Вы находитесь на странице: 1из 142

MSc Analytical Chemistry - N.

Blakiston 1

FACULTY OF SCIENCE

MSc DEGREE
IN
Analytical Chemistry

Neil Blakiston

K0429387

Development of a Near-Infrared
Spectroscopic Method for the
Determination of Water in Prozac

September 2007
Supervisor: Dr R. J. Singer
MSc Analytical Chemistry - N. Blakiston 2

Acknowledgements
First, I would like to thank my Supervisors, Dr Richard Singer and Dr Steve Barton

whose support, encouragement and feedback on both practical and written work was

invaluable to the completion of this thesis.

My gratitude also goes to Amanda Hibberd for her assistance and training on the NIR

spectrophotometer, to Jim Martindale, Ian Anderson, Yousuf Jeetoo, Ian Ross, Andy

Firth, Russell Freeman and Teresa Coyle for their review and input to the technical

reports and authorisation of the practical work.

I am also grateful to Eli Lilly & Co. Ltd. for their financial support and Jon Marks

whom supported my application, gave encouragement and technical guidance.

Finally, I would like to thank my family, especially my wife Jane, daughters

Hermione and Hannah for their continued encouragement and patience throughout all

of my studies.
MSc Analytical Chemistry - N. Blakiston 3

Abstract
The application of near-infrared (NIR) spectroscopy as an alternative analytical

procedure for analysis has been a controversial subject with both the analytical

community and regulatory authorities. This is due, in part, to the belief that the

International Conference on Harmonisation (ICH) Guidelines on the Validation of

Analytical Procedures could not be applied to NIR spectroscopy.

In this study Prozac was chosen to determine if NIR spectroscopy could replace the

more time consuming, costly and hazardous method of water determination by

volumetric Karl Fischer Titration (reference method).

This work demonstrated that a NIR method for determination of water was not only

feasible, but with only a few data points could produce results statistically comparable

to the existing method. During investigation the reference method required

development due to its unforeseen poor performance.

The guidelines of the International Conference on Harmonisation (ICH) for the

submission requirements of the pharmaceutical industry regulatory approval are also

discussed in relation to NIR methods.

The method proved suitable for the measurement of water concentrations from

approximately 2 to 9 % H2O with a maximum absolute error of 1.2 % H2O from the

average replicate value of the reference method. This is of a similar magnitude to the

error between replicate analysis by the reference method.


MSc Analytical Chemistry - N. Blakiston 4

Contents Page

FACULTY OF SCIENCE..............................................................................................1
Acknowledgements........................................................................................................2
Abstract..........................................................................................................................3
Contents of Figures........................................................................................................7
Contents of Tables and Attachments..............................................................................8
Abbreviations................................................................................................................9
General Introduction....................................................................................................10
Aims and Objectives................................................................................................10
Background..............................................................................................................10
Water Determination................................................................................................12
Definition:............................................................................................................12
Methods for the determination of water...............................................................13
Choosing a Suitable Method........................................................................................16
The Product of Choice..............................................................................................16
The Method of Choice..............................................................................................17
Introduction to the drug product PROZAC®...............................................................18
Sample Source..............................................................................................................20
Introduction to Karl Fischer Volumetric Titration.......................................................21
Karl Fischer Titration - History and Development..................................................21
Conditions for Karl Fischer Titration.......................................................................25
Hydranal® Composite 5 Titrant...............................................................................26
Solvent......................................................................................................................27
pH.............................................................................................................................28
Standard Sodium Tartrate Dihydrate........................................................................29
Automated Analysis.................................................................................................29
Orion Turbo2TM Volumetric Karl Fisher Titrator..................................................31
Limitations of KF.....................................................................................................34
An introduction to Near-infrared spectroscopy............................................................36
NIR Spectroscopy - History and Development........................................................36
Theory......................................................................................................................38
Instrumentation.........................................................................................................44
Mathematical Pre-Treatment and Processing...........................................................46
Bruker Opus NIR Spectroscopic Quant-Software...................................................46
Theoretical Background...........................................................................................47
Mulitvariate Calibration...........................................................................................48
Cross Validation.......................................................................................................49
Processing the Data..................................................................................................50
Experimental Analysis Karl Fischer............................................................................52
Chemical Safety (CoSHH, Risk and MSDS)...........................................................52
Sample Preparation..................................................................................................53
Loss on Drying Determination.................................................................................54
Good Manufacturing Practice (GMP) Requirements...............................................54
Execution of Technical Reports...............................................................................55
QCL-TR-014: Loss on Drying of Prozac Powder Blend.....................................55
QCL-TR-015: Water Determination on Dried Prozac Powder Blend by Karl
Fischer Titration and Near-Infrared.....................................................................56
Investigation of Failed Execution: QCL-TR-015.....................................................56
MSc Analytical Chemistry - N. Blakiston 5

Review of method AP1043-01.................................................................................59


Performance checks..............................................................................................60
Review of Site Certification Report B07766.......................................................61
Influence of solvent on the Karl Fischer reaction................................................62
Sample Transfer to the Karl Fischer Instrument..................................................62
QCL-TR-016: Water Determination on Dried Prozac Powder Blend by Karl
Fischer Titration and Near-Infrared II..................................................................64
Evaluation of Karl Fischer Data...............................................................................65
Data Acceptance Criteria......................................................................................65
Review of Variance between Replicates..............................................................69
Background Sample Analysis..............................................................................69
Summary..............................................................................................................70
Experimental Analysis NIR.........................................................................................70
Performance Calibration check................................................................................71
Sample Analysis.......................................................................................................72
Calibration Models...................................................................................................72
Mathematical Pre-treatment and Processing............................................................73
Calibration of Data Set R1.......................................................................................74
Calibration of Data Set R2.......................................................................................75
Calibration of Data Set R3.......................................................................................76
Comparison of R1, R2 and R3.................................................................................77
Repeatability Five Replicates - Between Methods...................................................82
Repeatability between methods............................................................................82
Review of water replicates by KF........................................................................83
Review of water replicates by NIR......................................................................83
Comparison of variance between methods...........................................................84
Repeatability 16 Replicates - Within Method..........................................................91
Repeatability NIR measurements.........................................................................91
ANOVA on 16 Replicates....................................................................................93
Outlier Check for Repeatability...........................................................................94
Review of Residuals for Repeatability.................................................................96
Discussion of Results.................................................................................................101
Further Discussion on Results................................................................................104
Cost.....................................................................................................................104
Speed..................................................................................................................105
Non-destructive..................................................................................................106
Containment and Safety.....................................................................................107
Conclusion..................................................................................................................109
A Review of Method Validation................................................................................111
Specificity...............................................................................................................112
Linearity.................................................................................................................112
Range......................................................................................................................112
Accuracy.................................................................................................................112
Precision.................................................................................................................112
Repeatability...........................................................................................................113
Intermediate Precision............................................................................................113
Reproducibility.......................................................................................................113
Detection Limit......................................................................................................113
Quantitation Limit ...............................................................................................114
Robustness/Ruggedness.......................................................................................114
MSc Analytical Chemistry - N. Blakiston 6

System Suitability Testing...................................................................................114


Reference Review......................................................................................................119
Applications of NIR Spectroscopy.........................................................................119
Qualitative Determinations................................................................................119
Quantitative Determinations..............................................................................122
Brief History of Water Analysis by NIRS.................................................................125
Future Work...............................................................................................................128
Reference to Raw Data...............................................................................................129
Bench Books..........................................................................................................129
Analytical Data Wallet...........................................................................................129
References..................................................................................................................130
MSc Analytical Chemistry - N. Blakiston 7

Contents of Figures
Figure 8.1 Three representations of the structure for fluoxetine;.................................19
Figure 9.1 MG2 Supermatic Capsule Filling Machine................................................20
Figure 10.5 Departure of the reaction rate constant K on the pH;...............................28
Figure 11.1a Diagram of the electromagnetic spectrum..............................................37
Figure 11.1b Full Electromagnetic Spectrum19..........................................................37
Figure 11.2a Vibrational energy levels for a diatomic molecule38.............................40
Figure 11.3a Diagram of diffuse reflectance and transmittance;.................................45
Figure 13.3b Primary water peak at 5180cm-1 (combination band)............................73
Figure 13.8a Calibration plot for Model R1.................................................................77
Figure 13.8b Calibration and Validation plot for Model R2........................................79
Figure 13.8c Calibration and Validation plot for Model R3........................................79
Figure 13.10.4 Frequency plots for calibration model R1, R2 and R3........................95
Figure 13.10.5 Matched pair plots for calibration model R1, R2 and R3....................98
MSc Analytical Chemistry - N. Blakiston 8

Contents of Tables and Attachments

Table 12.7.2 Site Certification of Method B07766.....................................................61


Table 13.8 Comparison of calibration models R2 and R3...........................................80
Table 13.9.1 Repeatability data for KF and NIR comparison......................................82
Table 13.9a F-test between KF and R1 for five replicate results.................................85
Table 13.9b F-test between KF and R2 for five replicate results.................................85
Table 13.9c F-test between KF and R3 for five replicate results.................................85
Table 13.9d t-test between KF and R1 for five replicate results..................................87
Table 13.9e t-test between KF and R2 for five replicate results..................................87
Table 13.9f ANOVA: Single Factor............................................................................88
Table 13.10a Predicted NIR values Vs KF - Repeatability..........................................92
Table 13.10b ANOVA: Single Factor - R1:R2:R3......................................................93
Table 13.10c ANOVA: Single Factor - R2:R3............................................................94
Table 13.10c Dixon’s Q-test applied to NIR repeatability study.................................94
Table 13.10d Residual analysis for predicted versus true for NIR..............................97
Attachment 1: Tabulated Overview moisture determination methods 30..................133
Attachment 2: Orion Turbo2TM Volumetric Karl Fisher Titrator............................134
Attachment 3: Bruker PQ Test Protocol....................................................................135
Attachment 4: Pay-Off Matrix for Water Determination A.......................................136
Attachment 5: Pay-Off Matrix for Water Determination B.......................................137
Attachment 6: Pay-Off Matrix for Water Determination C.......................................138
Attachment 7: Bruker MPA FT-NIR Default Parameter Settings.............................139
Attachment 8: Equipment...........................................................................................140
Attachment 9: Method AP-1043-01...........................................................................141
MSc Analytical Chemistry - N. Blakiston 9

Abbreviations
ANOVA Analysis of Variance
API Active Pharmaceutical Ingredient
BP British Pharmacopoeia
cGMP Current Cood Manufacturing Practice
CV Coefficient of Variation
FT Fourier-transform
HPLC High Performance Liquid Chromatography
ICH International Conference on Harmonisation
IR Infrared
KF Karl Fischer
MLR Multiple Linear Regression
NIR Near-infrared
PASG Pharmaceutical Analytical Sciences Group
PCA Principal Components Analysis
PhEur European Pharmacopoeia
PLS Partial Least Squares
PLSR Partial Least Squares Regression
QCL Quality Control Laboratory(s)
R2 Multiple Correlation Coefficient
%RSD Percentage Residual Standard Deviation
RSEP Relative Standard Error of Prediction
RSS Residual Sum of Squares
s Standard Deviation
se Standard Error
SEC Standard Error of Calibration
SEE Standard Error of Estimation
SEP Standard Error of Prediction
%SEP Percentage Standard Error of Prediction
SNV Standard Normal Variate
TR Technical Report
USP United States Pharmacopeia
UV Ultraviolet
µ Mean of the Population
σ Standard Deviation of the Population
MSc Analytical Chemistry - N. Blakiston 10

General Introduction
Aims and Objectives

The primary aim of this thesis was to conduct a feasibility study, investigating the use

of near-infrared (NIR) spectroscopy as a replacement analytical technique for Karl

Fischer volumetric analysis, for the determination of water in pharmaceutical product.

To achieve this end goal a suitable multicomponent material was required, which

would have suitable range of water contents, so as to produce a calibration set for the

NIR software to manipulate. The objective was to apply multivariate and

chemometeric tools to produce a good fit calibration from which the precision and

accuracy of the NIR method could be assessed.

Background

Water can have an adverse effect to the stability profile of an active pharmaceutical

product, through degradation of the sample over time. Both the physical and

chemical properties may be altered, which could influence the safety, strength,

quality, performance and shelf life of the manufactured product. An example of a

chemical change due to the quantity of water present in a sample is the increased rate

of degradation pathways which could lead to raised levels of related substances

(degradation products) above regulated limits. A physical example would be the

change to friability of a tablet or reduced performance of gelatine capsules.

The impact of these changes may result in a costly product recall, which may lead to

customer, share holder and inspectorate loss of faith.


MSc Analytical Chemistry - N. Blakiston 11

To support the regulatory commitment of product release and to demonstrate shelf life

stability various analytical techniques are conducted on the pharmaceutical product.

These include both in-process monitoring and finished product analysis.

Water determination on finished product within Eli Lilly and company limited is

routinely conducted by Karl Fischer volumetric determination. This analytical

technique is destructive and time consuming. It also requires the use of various

organic solvents. Due to the nature of these solvents the analysis requires fume

cupboard or bench extraction and appropriate containment.

A risk of skin contact with the solvents used presents a hazard to the analyst, hence

Personal Protective Equipment (PPE) must be worn for analysis at all times. In

implementing any risk reduction to a process containment and PPE should be the last

consideration according to industry hierarchy. Continued use of even the most hyper-

allergenic nitrile gloves poses its own health issues and can cause irritation and

sensitisation. Eliminating the risk or re-engineering the process should be the first

consideration.

NIR spectroscopy offers a safe alternative where, solvents and reagents are eliminated

and samples require no significant preparation. NIR offers a non-destructive analysis,

so the sample can be stored or used for further analysis and no waste is generated

other than the sample.


MSc Analytical Chemistry - N. Blakiston 12

Water Determination

Definition:

“The moisture contained in a material comprises all those substances which vaporize

on heating and lead to weight loss of the sample... According to this definition,

moisture content includes not only water, but also other mass losses such as

evaporating organic solvents…, aromatic components, as well as decomposition and

combustion products” 30.

Within the industry the term moisture and water are often interchangeable, even if not

correct. This is evident within Eli Lilly as many global analytical methods are titled

“Moisture determination by Karl Fischer…”

There are several different types of moisture

• Free Moisture; Readily available, dissolved, homo- or heterogeneous e.g.

perfume, beverages or solvents.

• Surface Moisture; Readily available, covers surface of solids, heterogeneous

e.g. sugar, polymers.

• Trapped Moisture; this must be liberated for analysis, heterogeneous e.g. cells,

food products, polymers.

• Capillary Moisture; Must be liberated for analysis, heterogeneous e.g. soil,

rocks, concrete.

• Water of Crystallisation; chemically bound, homogeneous e.g. soil, rocks,

concrete.
MSc Analytical Chemistry - N. Blakiston 13

Figure 4.3.1 The various forms of water

storage

Methods for the determination of water

Thermogravimetric methods

• Drying oven with balance

• IR drying and direct weighing

• Microwave drying and direct weighing

• Halogen drying and direct weighing

• Thermal Gravimetric Analysis

• Azeotropic distillation and weighing

• Phosphorous Pentoxide P2O5 method

Spectroscopic Methods

• Near Infrared spectroscopy

• Microwave spectroscopy

• NMR
MSc Analytical Chemistry - N. Blakiston 14

Chemical/Other

• Karl Fischer

• Calcium carbide method

• Density determination

• Refractory determination

• Conductivity

• Gas Chromatography

• Distillation

Each method has its own advantages and disadvantages and each will be discussed in

turn.

Gravimetric methods

The methods used for drying the sample to a predetermined or constant weight are

only suitable if the chemistry of the sample is known, as volatisation will remove not

only water, but other volatiles present. There is also the problem of thermal

decomposition of the sample. The samples required are generally large for most

drying techniques, which give a good composite result, but not suitable if the sample

size is small. Although analysis time can be long, for example oven drying times can

be several hours or even days, sample handling can be quite large depending on oven

size. Halogen, infrared, and microwave offer benefits as they do not suffer from

lengthy analysis times.


MSc Analytical Chemistry - N. Blakiston 15

Thermogravimetric analysis suffers from only being able to handle small sample

quantities, so is generally not suitable in the analysis of, for example, a batch of

pharmaceutical product. The analysis tends to be slow and has limited suitability for

liquids, but a wide application for solid samples. These techniques are generally used

in the pharmaceutical industry for in-process manufacturing controls, for example

during wet granulation or use of fluid bed dryers, where a certain end point of

moisture content is required. They are also used in the analysis of excipient material

following pharmacopoeia15 methods.

Spectroscopic methods

These have fast analysis times, can be easily automated and can handle large sample

volumes. Microwave and NIR analysis both need to be calibrated to a specific

substance and suffer interference from particle size and bulk density variation. The

instruments require high capital investment, especially NMR and a significant amount

of knowledge and time is required to develop methods and manipulate the data.

Chemical/Other

Karl Fischer has a wide application and is relatively cheap, but can not be automated

and analysis times vary depending on the substance analysed. The coulometric

method deals with trace analysis, while the volumetric deals with water contents of 1-

100%.

Refractometry and density measurement is a fast analysis technique, requires little

training and is highly mobile, but is suitable only for clearly defined samples.
MSc Analytical Chemistry - N. Blakiston 16

The calcium carbide method is favourably priced, but is prone to forming explosive

materials and requires specialist training and relevant safety precautions. Gas

Chromatography is suitable for multicomponent analysis, but requires specialist

knowledge in development and in daily operation. Automation proves useful for

large sample volumes, but set up time is lengthy if only one sample is required.

Attachment 1 gives a tabulated overview of the typical measurement range, accuracy

and water selectivity of a range of analytical techniques. Attachment 4, 5 and 6

indicates a pay-off matrix for the various techniques.

Choosing a Suitable Method

The Product of Choice

As this was my first introduction to NIR analysis it was suggested that a product with

few ingredients should be chosen. This would reduce the complexity of any spectra

obtained and should reduce any interferences with the water band(s) of interest.

The product chosen would need to present a range of water contents that could be

determined to construct a calibration set for the NIR software or it needed to be in a

form that could be processed in the laboratory to give the same. Any result generated

as part of this product could not jeopardise the release of product from the site or any

batch on stability.

The product that met these requirements was Prozac. Prozac has a relatively simple

manufacturing recipe containing just Fluoxetine Hydrochloride, Starch Flowable and

Silicone fluid (350 centistokes) 0.96%. The active ingredient represents


MSc Analytical Chemistry - N. Blakiston 17

approximately 10%, by weight, in the bulk powder. The powder blend is filled into

gelatine capsules with a running weight of approximately 230mg to give the

equivalent of 20mg active fluoxetine as the free base.

Prozac is manufactured by Basingstoke Manufacturing Operations, Eli Lilly and Co

Ltd under licence.

As Prozac is a granular/powder blend it was suitable for drying or doping to give

various concentrations of water content. The sample could also be collected in such a

way or adulterated by environmental conditions so that it would not represent any

released material.

The batch of Prozac powder blend used in this thesis was dispensed on the 30th May

2007 and the product was released by qualified person approval on the 13th June 2007.

The batch size was approximately 1.5million doses and review of the batch

documentation indicated no processing issues or any laboratory test result failures.

The Method of Choice

Due to the very stable nature of Prozac in the presence of water this product does not

have any regulatory release limits for water content. Interrogation of the Laboratory

Information Management System (LIMS) revealed that Global Method B07766

(Turbo2TM Karl Fischer Assay of Water in Fluoxetine Hydrochloride Capsules

Equivalent to 10 and 20 mg Fluoxetine) had been developed for the measurement of

water in Prozac formulations. This method had been used on the Basingstoke site for

the analysis of stability samples for a short period and is no longer performed. Global
MSc Analytical Chemistry - N. Blakiston 18

Method B07766 was imbedded into Basingstoke Analytical Procedure (AP) AP1043-

01 (see Attachment 9).

Introduction to the drug product PROZAC®


In a healthy individual serotonin is released, in the body, into the space between the

"sending" nerve cell and the "receiving" nerve cell. When serotonin is received on the

surface of the "receiving" cell, it stimulates or activates serotonin receptors.

Stimulation of these receptors generates an impulse and allows messages to move

forward.

When a person suffers from depression, there may be a problem with the balance of

the serotonin system. It is thought that this imbalance occurs when serotonin is

released from the "sending" nerve cell and is reabsorbed by an uptake pump. By

blocking the serotonin uptake pump, Prozac increases the amount of active serotonin

that can be delivered to the "receiving" nerve cell. This helps message transmission

return to normal.

Pharmacotherapy is currently the only proven method for treating major depressive

disorder and Prozac (active ingredient - Fluoxetine hydrochloride (HCl) Figure 8.1) is

one of the world's most widely prescribed antidepressants; it has been prescribed for

more than 54 million people worldwide.

Prozac is approved by the FDA (Food and Drugs Administration) for the treatment of

the following disorders;

• Major Depressive Disorder, Obsessive-Compulsive Disorder, Bulimia

Nervosa and Panic Disorder in adults.


MSc Analytical Chemistry - N. Blakiston 19

• Major Depressive Disorder and Obsessive-Compulsive Disorder in paediatric

patients (children and adolescents).

Prozac was a first-in-class antidepressant, which lost U.S. market exclusivity in 2001

and experienced huge generic competition.

Unlike other treatments in this field Prozac is not associated with withdrawal

symptoms of a somatic or psychological nature and has a relatively benign side-effect

profile.

Figure 8.1 Three representations of the structure for fluoxetine;

The active ingredient in Prozac;

(N-Methyl-g-[4-trifluoromethyl)phenoxy]benzenepropanamine).
MSc Analytical Chemistry - N. Blakiston 20

Sample Source
The sample was taken from the powder bed of an Italian MG2 Supermatic Capsule

Filling Machine (Figure 9.1) capable of filling up to a 60,000 capsules per hour.

Capsule gravity feed

Capsule
delivery tubes

Powder Bed

Figure 9.1 MG2 Supermatic Capsule Filling Machine.

This is a high performance encapsulation machine, which consists of a capsule-

containing hopper that delivers capsules into feeding tubes, having vertical motion

and an orienting drum. The capsules are separated, filled and rejoined before passing

a metal check and de-duster area. Finally they are air fed into fibre board drums

ready for packaging. At the end of the operation the filling bed is separated and

several kilograms of powder blend are left behind, which is removed as waste.

This material was collected and placed inside a double lined plastic bag. The sample

contained both powder blend and waste gelatine capsules. To prepare the sample for

analysis the powder was poured into a shallow dish and the individual capsule shells

were removed with tweezers.


MSc Analytical Chemistry - N. Blakiston 21

This was not an ideal pure sample as it was not possible to remove all of the capsule

fragments. These fragments were present in a relatively small quantity as most

capsule shells were intact. It is not known if this has had any impact on the course of

the analysis carried out.

The sample was divided into several HDPE screw cap bottles, which provide a

relatively good barrier against water ingress.

Introduction to Karl Fischer Volumetric Titration

Karl Fischer Titration - History and Development

When required to determine the water content in an organic compound, considerations

of heat stability, volatility and the time factor involved make it difficult or impossible

to apply physical test methods.

In 1935 Karl Fischer initiated a new method for the determination of water. He

considered that the reaction involving iodine, sulphur dioxide and water, originally

investigated by Bunsen1 and later applied to the determination of sulphur dioxide in

flue gases, might form the basis for the estimation of small quantities of water.

The reaction may be represented by;

I2 + SO2 + 2H2O ↔ 2HI + H2SO4 (1)


MSc Analytical Chemistry - N. Blakiston 22

Karl Fisher dissolved the iodine (0.33 mol/l) and sulphur dioxide (0.50 mol/l) in

anhydrous methanol and added pyridine (1.67 mol/l) to displace the reaction to the

right by removing the acidic products of the reaction; he assumed, at the time, this had

no bearing on the course of the reaction, so that one mole of iodine would be

equivalent to two moles of water.

In the presence of a suitable base, the reaction consumes two molecules of water for

one molecule of iodine (Equation 1) and therefore there is a direct relationship

between water present and the consumption of iodine. Smith, et al.2 however, showed

that his assumption was unjustified, since both the pyridine and the methanol take

active part in the reaction process. In its simplest form they represented the reaction

as follows;

(2)

(3)

Therefore each molecule of water is equivalent to one molecule of iodine. This has

been universally accepted within the analytical community.

The early Karl Fischer reagents were not completely stable, due to the hygroscopic

nature of the constituents and other reactions, which take place even in the absence of

water. An established early procedure for reagent preparation was detailed in the

European Pharmacopoeia3 as follows;


MSc Analytical Chemistry - N. Blakiston 23

“In a 750 ml combustion flask mix 400 ml of anhydrous methanol and 80 g of

anhydrous pyridine. Immerse the flask in ice and bubble dried sulphur dioxide

through the mixture until the weight of the flask and contents has increased by 20 g.

Add 45 g iodine and shake the flask until solution is complete. Allow the solution to

stand for 24 hours before use”.

Several references suggest adding stabilising agents to the Karl Fisher titrant, such as

pyridinium iodide4 to absorb the hydrogen iodide and sulphuric acid formed; sodium

and zinc acetates5 can be added to eliminate side reactions, while addition of bromine6

allowed determination, iodometrically, of iodic acid from the oxidation of any

hydrogen iodide. The methanol present in the reagent can also promote side reactions

with the sample, which can increase its instability. Early reagents often substituted

the methanol with other more stable constituents’; 2-methoxy-ethanol7,

dimethylformamide8 and ethanediol9. Although methanol is routinely used as the

sample solvent care must be taken that unwanted side reactions do not occur with the

sample. For example; carbonyls, particularly aldehydes and ketones often need to be

converted into the cyanohydrins2 to avoid interfering reactions and the production of

water.

Two classes of compound therefore give anomalous results when titrated with Karl

Fischer reagent;

1. Those reacting with the iodine-sulphur dioxide portion of the reagent, which

include; ascorbic acid, quinols, per-acids, diacyl peroxides (and other


MSc Analytical Chemistry - N. Blakiston 24

oxidising agents), amines, mercaptans (thiols and other reducing agents) and

easily oxidised substances.

2. Those reacting with the components of the reagent, which result in water

formation. These include formic and acetic acids, which slowly form esters

with methanol in the reagent, basic oxides and salts of weak oxy-acids18. The

lower aldehydes and ketones similarly can form acetals and ketals7.

Modern instrumentation performs analysis at a rate, which reduces the extent of these

unwanted reactions. However if a potential problem exists, then alternative reagent

components and sample solvents may be required. For example British Standard

2511 :1970 recommends a modified reagent and pyridine-methanol solvent for the

determination of water in ketones.

The early Karl Fischer reagents also suffered from rapid decay, especially in sunlight,

which lead to erroneous results. The unpleasant smell and high toxicity of pyridine

opened the way to replace this undesirable component with a more suitable base, with

at least the same or better role in the reaction process.

Dr Eugen Scholz12/13 led development of the method in 1979 to eliminate pyridine as

the base. Investigation involved testing amines with a higher basicity and greater

affinity for methyl sulfite. Aliphatic amines and several other heterocyclic

compounds proved suitable replacements. However the base of choice proved to be

imidazole (C3H4N2), which shifts the equilibrium completely to the right, producing a

rapid reaction speed and stable end point (1980). Imidazole is the main base

component in Hydranal® Composite 5 KF reagent as supplied by Riedel-deHaën.


MSc Analytical Chemistry - N. Blakiston 25

With the development of new reagents, the stoichiometry of the reaction required

verification. It was found that earlier equations did not fully explain the course

previously described. Scholz and his team published the following accepted

representation;

CH3OH + SO2 + RN  [RNH]SO3CH3 (4)

Intermediate Methylsulfite

H2O + I2 + [RNH]SO3CH3 + 2RN  [RNH]SO4CH3 + 2[RNH]I (5)

Oxidation-Reduction step (rapid) RN=base

Sulphur dioxide reacts with methanol to form an ester, which is neutralised by the

base. The anion of the alkyl sulphorus acid is present in the KF reagent and is the

reactive component. The oxidation of the alkyl sulfite anion to alkyl sulphate by the

iodine consumes water in a 1:1 ratio with iodine.

Conditions for Karl Fischer Titration

Normally the water under analysis is not pure, but fixed onto a certain matrix and

certain conditions must be met for the analysis to proceed;

• The substance must dissolve in a suitable solvent or readily give up its water.

• It must not change the optimum working pH.

• Side reactions with the KF components must not take place.


MSc Analytical Chemistry - N. Blakiston 26

It is not always possible for these criteria to be met and therefore experimental

modification may be required. For the Analytical Procedure (AP) used in this thesis

certain changes to the basic Karl Fischer method have been employed to aid the speed

of titration, clarity of end point and release of water from the matrix. These are

discussed in turn below;

Hydranal® Composite 5 Titrant

Pyridine in the original Karl Fischer titrant is present to buffer the reaction process,

but is not present as a true reactant. The rate of the reaction is dependant on the base

chosen as it neutralises the intermediate CH3SO3H. The base activity rate of pyridine

is very slow and the end point is unstable due to its weak basicity, which can not

neutralise completely the methyl sulphurous acid. This can lead to poor repeatability;

hence pyridine can be replaced by bases which are superior for the application.

The Hydranal® Composite titrant is a one component reagent containing all of the

required elements for the reaction (imidazole, sulphur dioxide and iodine dissolved in

diethylene glycol monoethyl ether). It has been proven to reduce titration speed from

approximately 10 minutes, using conventional pyridine reagent, to about 4 minutes.

A feature of the Karl Fischer titration is the improved stability of the end point as the

time required for titration is shortened (Figure 10.3.).


MSc Analytical Chemistry - N. Blakiston 27

Figure 10.3 The course of three


titrations of 40 mg water; Using; A
conventional one component system (C),
Hydranal Composite 5 one component
reagent (B), and Hydranal two component
reagent (solvent and titrant) (A).

(Reidel-deHaën)11

The titrant has been adjusted to give a titre of about 5 mg H2O/ml and the minimum

quantity of water detectable is approximately 0.5 mg. The titrant has a shelf life of

two years and a low decay rate of approximately 0.2 mg H2O/ml per year.

Solvent

The solvent chosen for application must be able to dissolve the products of the

titration reaction, for this purpose methanol is the most frequently used solvent. The

solvent must give complete esterification of the sulphur dioxide (equations (4) and

(5)) to ensure a stoichiometric course for the KF reaction. When using methanol in

the working medium the concentration must not fall below 25%, otherwise the end

point can be shifted.

Formamide improves the solubility of polar substances and readily mixes with

methanol (the methanol content of the solvent should be greater than 50%) and is

good for the detection of water in proteins, carbohydrates and inorganic salts. It is

especially good in aiding the release of water bound in such matrix as lactose and

starch, and increases the course of the reaction. Formamide requires special

precaution when handling as it has shown to have teratogenic action, so women of

child bearing age must avoid inhalation and skin contact.


MSc Analytical Chemistry - N. Blakiston 28

pH

Karl Fischer has an optimum pH range of 4-7 (Figure 10.5), in which the reaction

runs quickly and stoichiometrically. At high pH side reactions are known to consume

iodine slowly and an end point reversal is observed. In strong acid conditions the rate

constant of the Karl Fischer reaction decreases, hence the course of the titration is

slower.

Figure 10.5 Departure of the

reaction rate constant K on

the pH;

according to Verhof and


Barendrecht 14

To balance any shift in the pH the addition of a weak acid or base is suitable to

neutralise the medium. An alternative approach is to add a buffering reagent to the

working medium. For method AP-1043-01 a 2% Hydranal buffer in methanol,

combined 50/50 v/v with Formamide is stated for the solvent medium. Review of the

method, chemistry and method validation indicates the reason for buffering the

solution, is due to the N-H group on the Fluoxetine API (Active Pharmaceutical

Ingredient) and the slight basicity of the starch excipient. The buffer is also available

as a base or acid reagent. However the method does not indicate which should be

used. As other methods in the laboratory use the Hydranal acid buffer, this was

chosen as the buffering reagent. The benefit of the buffer could be further

investigated through comparison of pH measurements during a titration using the

solvent, as prepared by the method in question, versus one run using 50/50 v/v

formamide and methanol.


MSc Analytical Chemistry - N. Blakiston 29

Standard Sodium Tartrate Dihydrate

Sodium tartrate forms a dehydrate, which is stable under normal conditions, does not

effloresce or absorb moisture. It therefore maintains a relatively constant water

content of 15.66%. It is used in KF determination for calibration or standardisation

purposes. However prior to use a loss on drying check may be periodically required

to ensure that no change of water content has occurred and also requires storage in a

desiccator. Due to any variability in its water content, a much cheaper reagent is now

routinely used, this being deionised water. This allows direct calibration at the level

of interest and can be administered quickly and directly into the titration vessel by

micro syringe.

Automated Analysis

Volumetric analysis is used to determine relatively high concentrations of water and

can be conducted using a single or two component titration system. The end point in

modern equipment is determined when either a defined potential (bipotentiometric

indication), or current (biamperometric indication) is not only reached, but remains

stable for a period of time. This is termed the “dead stop” method10 based on zero

current flow between two polarised electrodes enclosed in the titration vessel and is

employed with the Orion Turbo2TM Karl Fischer (as discussed below). An alternative

approach is the ‘drift stop’ method where the end-point is determined by a pre-

determined drift value measured in the titration cell. The measurement is detected by

changes across two platinum electrodes produced by the first stable appearance of free

iodine. As long as the added iodine reacts with the water present a high voltage is

required to maintain the specified polarization current across the electrodes. As free
MSc Analytical Chemistry - N. Blakiston 30

iodine becomes available it causes ionic conduction and the voltage is reduced to keep

the current constant. When the voltage drops below a defined value the titration is

terminated.

One of the highest sources of error in water determination is due to the ingress of

atmospheric moisture. For example one litre of air may contain about 20 mg of water

and to put this into perspective, one drop of Hydranal titrant (0.01 ml) corresponds to

0.05 mg of water, which is equivalent to 2.5 ml of extraneous air. For this reason the

stock reagents and vessel need adequate protection from the atmosphere. This is

provided by suitable seals or grease, which must be regularly checked and replaced as

needed. Generally, control of further water ingress is by a suitable hygroscopic

molecular sieve (drying tube) through which air is drawn to relieve any pressure

differential.

The sample addition of powder is usually conducted via back weighing addition from

a ‘long-spout’ glass weigh boat directly to the Karl Fischer cell. The sample transfer

process takes approximately 10 seconds, for the ‘Turbo2’, through a small stoppered

aperture; hence the exposure of the cell to the atmosphere is kept to a minimum.

The sample size is estimated according to the amount of water content in the sample.

Method AP-1043-01 gives guidance to the amount of sample required, which will be

discussed later. Literature values10 state that a sample size of approximately 500mg

should be taken if the water content is 10%.


MSc Analytical Chemistry - N. Blakiston 31

The results of the analysis are calculated from the consumption of reagent used in ml

(a), the water equivalency of the reagent in mg H2O/ml (WE) and the weight of the

sample in g (e), so that;

mg H2O = a . WE (6)

and

% H2O = a . WE (7)
10 . e

Orion Turbo2TM Volumetric Karl Fisher Titrator

The Turbo2 titrator combines microprocessor technology with an efficient in-built

multi-speed blender, capable of running at 7500rpm. This effectively homogenises

the sample, increasing the penetration of titrant and the release of bound water is

greatly improved. This function allows the handling of a greater range of sample

matrix eliminating the need to pre-grind samples, which can introduce moisture or

thermally degrade the sample. The titration vessel is large and can accommodate

about 750ml of solvent and sample, reducing the requirement to empty and refill. The

titrant is introduced with a high precision peristaltic pump system employing a stepper

drive motor with detachable head and tubing platen16. The reaction vessel has an

automatic vacuum waste extraction with a 3cm sample port, which is sealed with a

solid plastic stopper.

The Titrator is calibrated daily following local procedure; BPD-071003-007 (Orion

Turbo 2 Karl Fischer-operation of) and during analysis guidance in Analytical

Procedure; AP-1043-01 (Turbo2TM Karl Fischer Assay of Water in Fluoxetine

Hydrochloride Capsules Equivalent to 10 and 20 mg Fluoxetine) is followed.


MSc Analytical Chemistry - N. Blakiston 32

Calibration requires a consecutive calibration test using 25μl water injections to give

three results that fall within a 1.0% correlation variation (C.V.) limit. No more than

five injections are permitted to conduct an acceptable calibration. The results of three

acceptable calibration samples are used to calculate a calibration factor, which is

stored in the memory for determination of analytical samples. The instrument has a

literature precision of 0.5% at 25mg H2O.

The instrument has several programmable variables;

1. End-point Time: At the end of a titration, the addition of excess titrant will

temporarily drive the system into an over-titration situation from which it will

need time to recover. This feature provides a variable wait time before the

‘dead stop’ reading is taken. The end-point time can be set between 1-30

seconds with 10 seconds being the default.

2. End-point Level: This variable value (5-23 arbitrary units) changes the

conductance measured at the electrode. By varying the value the end-point

will be moved along the titration curve. By lowering the value the end-point

reaction will move to an over-titration condition.

3. Step Level: This has a direct relationship to the end-point level and has

arbitrary units of 1-18. The value determines the point along the titration

curve where the peristaltic dispenser slows down and enters a pulse mode until

the final end-point is reached. The relationship between end-point level and
MSc Analytical Chemistry - N. Blakiston 33

step level should be maintained and too low a level gives rapid titrations, but

precision may be lost.

4. Turbo Speed: The action of the blender head has three speed settings;

• Speed 1: (approx 1000 rpm) is for liquid samples and fine powders.

• Speed 2: (approx 3000 rpm) is for granules and viscous materials.

• Speed 3: (approx 7500 rpm) is for breaking down solid materials.

5. Turbo time: The time period for homogenising prior to the start of titration.

6. Titration mode: There are four modes available for analysis; ‘Sample Only’,

where the sample is titrated directly after addition to the vessel. This is

generally used for fully miscible samples. ‘Background Sample’, used for

insoluble solids using high speed blending and a pre-titration to determine the

background water level. This is automatically subtracted from the sample

analysis. ‘Background - Sample - Background’, determines the moisture level

before and after analysis and the average is subtracted from the sample.

Finally the ‘Search’ mode allows analysis of samples with an unknown water

release profile. It will alternately blend and titrate until a set criterion is

reached. For example, a point at which the water content is lower than the

background, two consecutive results fail to increase the cumulative result by

1% H2O, a total of 30 minutes Turbo time elapses etc.

A schematic of the instrument can be seen in Attachment 2.


MSc Analytical Chemistry - N. Blakiston 34

Limitations of KF

1. As previously discussed the hydroscopicity of the reagents means that routine

standardisation/calibration is required. The period of which may be weekly or

daily dependant on operating procedures and instrument validation.

2. Changes in room temperature of the laboratory will cause fluctuations in the

titre, as the reagents contain organic solvents; they are more prone to

significant thermal expansion compared to aqueous solutions. Generally, a

temperature increase of 1˚C will result in a titre decrease of about 0.1%.

3. When conducting method development a sound knowledge of the chemistry,

nature of the matrix, available reagents, environmental parameters and

equipment operating conditions is an essential requirement.

4. The new reagents used in the Karl Fischer, although less harmful than the

older generation, still pose particular safety issues.

5. The main safety issue is that of containing the inhalation of volatile substances

and for this reason the technique should be undertaken in a fume cupboard or

under an extraction unit.

6. The biggest draw back is the lack of automation and the time it takes for each

analysis. Analysis times are generally between 5 to 20 minutes (some take

longer). This not only limits the analyst to the number of samples that can be

run in a day, but also involves a lot of non-value added time waiting for the
MSc Analytical Chemistry - N. Blakiston 35

titration to finish. The short wait period means it is very difficult to conduct

other activities between sampling.

7. The KF instrument requires connection to a balance or the balance weights

need to be manually entered. If the balance is connected, stability of the

balance may be a problem due to the flow of air from extraction. Manual

transfer of weights increase the time per analysis and could lead to

transcription errors.

8. The Karl Fischer technique is well contained, but a number of reagent

transfers do occur; the stopper from the sample port is removed each time a

sample is added and the vessel needs to be emptied, cleaned and refilled

routinely. All these actions lead to possible spillage and staining of equipment

with KF reagent. The drying tubes, seals, and delivery tubes need routine

maintenance and replacement, so ownership of the system is essential.

Near-infrared spectroscopy offers an attractive solutions to these limitations if a

viable method can be developed to give equivalent or better measurements compared

to the Karl Fischer technique of water determination.


MSc Analytical Chemistry - N. Blakiston 36

An introduction to Near-infrared spectroscopy

NIR Spectroscopy - History and Development

NIR radiation was first discovered by the English astronomer Sir William F Herschel

in 1800. His area of interest lay in identifying which colour of the visible spectrum

was responsible for the heat in sunlight (“Experiments on the Refrangibility of the

invisible rays of the Sun”). Using a prism he separated sunlight into its constituent

visible light spectrum and by means of a thermometer measured the temperature

associated with each colour. No discernible increase in temperature was observed

until the thermometer was placed just outside the red end of the spectrum. At this

location there was no light visible to the naked eye so Herschel named this region

infrared, the prefix being Latin for below.

Although some investigative work continued after this discovery (a full history of

works are reference by Workman37) there was no significant volume of work

published until the 1960’s, when a prolific series of papers were released. It was only

in 1973 when P. Williams reported the use of a commercial NIR grain analyser for

analyses of cereal products that the technique established a strong interest.

NIR is comprised of two sub-regions, 780 to 1100 nm and 1100 to 2500 nm, the

former being referred to as the Herschel region. This lies between the red end of the

visible spectrum and what is now called the mid infrared region (Fig 11.1a and 11.1b)

of the electromagnetic spectrum. This region was not widely investigated as a useful

analytical application due to scepticism arising from the apparent lack of sharp peaks,
MSc Analytical Chemistry - N. Blakiston 37

loss of baseline resolution and lack of sensitivity that is typically two or three orders

of magnitude less relative to the infrared region.

X-Ray UV VIS NIR Mid IR Radio Waves

200 380 780 2500 25000 nm

Figure 11.1a Diagram of the electromagnetic spectrum

Figure 11.1b Full Electromagnetic

Spectrum19

Band assignments are also difficult due to the

numerous overtone and combinations of

vibrations observed in the mid infrared region.

The consequence of this is that NIR absorptions

are less intense, broader and more overlapping

than the parent absorptions.

Recently this initial reluctance to investigate applications of NIR spectroscopy has

been compensated, due to instrumental breakthroughs that include developments in

detectors; fast-scan and Fourier-transform (FT) instruments, fibre-optics and

increased mathematical processing capabilities of modern computers. In addition,

unique combination bands provide information not available in the infrared region,

and these reduced intensities allow for direct measurements on undiluted samples.
MSc Analytical Chemistry - N. Blakiston 38

Although NIR spectroscopy is not useful for trace analysis, the benefits of NIR

spectroscopy are that it is rapid, non-destructive and in general, there is little or no

sample preparation required. NIR spectra can therefore be used to both identify and

quantify samples. NIR spectra also contain information relating to the physical

properties of the sample, such as particle size and compaction density which can be

measured with this technique. As an analytical technique, its penetration potential is

far greater than that of many other spectral techniques. This allows determination

deep into a bulk material and can even be conducted through outer packaging

components and manufacturing inspection windows.

Theory

Vibrational spectroscopy utilises the concept that atom-to-atom bonds within

molecules vibrate with frequencies according to the laws of physics. When these

molecular vibrators absorb light of a particular frequency, they are excited to a higher

energy level. In order to absorp infrared radiation, a molecule must undergo a net

change in dipole moment as a consequence to its vibrational or rotational motion.

Generally at room temperature most molecules are at the zero energy level so are

vibrating at the least energetic state allowed by quantum mechanics.

Infrared spectroscopy is used to investigate the molecular vibrational properties of a

sample by interpretation of the resultant absorption bands. The infrared radiation

absorbed by a molecule causes individual bonds to vibrate in a similar fashion to a

diatomic oscillator. Therefore, in the simplest vibrating model of a diatomic molecule

the molecule behaves as a harmonic oscillator and Hooke’s law is obeyed

(Equation (8)).
MSc Analytical Chemistry - N. Blakiston 39

F = −ky (8)

Where F is the restoring force and is proportional to the displacement, k is the force

constant of a spring (equivalent to the measure of a diatomic molecule bond strength)

and y is the distance from the equilibrium position38. In this model, as the atoms are

displaced from the equilibrium position, the potential energy of the vibrating system

increases.

This results in a symmetrical parabolic curve about the equilibrium position or bond

length. This works well for the fundamental vibrations of simple diatomic molecules

and is not too far from the average value of a two-atom stretch within a polyatomic

molecule36.

In real molecules the electron withdrawing/donating properties of neighbouring atoms

and groups influence the bond strength and length and thus the frequency of the

diatomic bonds. The harmonic oscillator model therefore has limits similar to those

of a spring attached to two masses. As one mass approaches the other compression

forces are fighting against the bulk of the spring. As the spring stretches, it eventually

loses its shape and fails to return.

In molecules the electron clouds around two bound atoms, together with the nuclear

charges, limit the approach of the nuclei during the compression step. In addition, at

the extension of the stretch the bond eventually reaches breaking point when the

vibrational energy level reaches the dissociation energy. Therefore, in practise, a

better model for the potential energy in a diatomic model is the anharmonic oscillator

illustrated in Figure 11.2a.


MSc Analytical Chemistry - N. Blakiston 40

Figure 11.2a Vibrational energy levels for a diatomic molecule38.

As can be seen from this graphic representation, the energy levels in the anharmonic

oscillator are not equal, the levels being slightly closer as the energy increases. An

expression for the deviation away from the simple harmonic model is the ‘Morse

function’, which is an empirical equation where the allowed energy levels, E, for this

anharmonic oscillator are given by;

E = (v + 0.5) hf − (v + 0.5)2 hfx (9)

Where h is Planck’s constant, v is the vibrational quantum number, f is the

equilibrium frequency of oscillation and x is the anharmonicity constant.

The vibrational quantum number is an integral value so the anharmonicity constant is

a measure of how the potential energy curve deviates from Hooke’s law. The

selection rules for the anharmonic oscillator are Δv = ± 1, ± 2, ± 3 etc. and therefore

the following transitions are allowed v1 ← 0, v2 ← 0, v3 ← 0, v2 ← 1 etc.


MSc Analytical Chemistry - N. Blakiston 41

In practice, only transitions starting at v = 0 are observed and these are referred to as

the fundamental transitions, first overtone, second overtone etc. respectively. These

transitions occur at frequencies, which approximate to f, 2f, 3f etc. due to the

anharmonicity constant being, for practical purposes, typically less than 5%36.

The fundamental transitions typically occur in the mid infrared region, while the

overtones bands which are 10 to 1000 times weaker than the fundamental vibrations

generally occur in the NIR region and this therefore is the basis of NIR spectroscopy.

In polyatomic molecules, there are different vibrational modes each of which has an

associated frequency and energy. These include symmetric stretching, asymmetric

stretching, scissoring, rocking, wagging and twisting as illustrated in Figure 11.2b.

Figure 10.2b Representation of different modes of bond vibration; the + sign

indicating the plane perpendicular to the page towards the reader and the ─ sign the

opposite. The principle fundamental vibration bands of water are at wavelengths

5150 cm-1 (Combination Band = Stretch + Scissoring ≈ 5352 cm-1), 3756 cm-1
MSc Analytical Chemistry - N. Blakiston 42

(Asymmetric Stretch) and 1596 cm-1 (Scissoring) 63. These bands are not only strong,

but are typically well resolved from absorption bands of other compounds. The

diffuse reflected intensity is measured, which is proportional (in a non-linear relation)

to the water concentration.

There is the potential for simultaneous changes in the energies of two or more

vibrational modes that results in a frequency that is the sum of the individual

frequencies, referred to as combination bands. These combination bands are typically

weak with a low probability of occurrence.

Typically the near infrared bands are broad and therefore spectra can be very difficult

to interpret with regards to specific chemical components. Multivariate (multiple

wavelength) calibration techniques are generally performed to clean the spectra and to

extract the chemical information required. This requires significant understanding in

the application of mathematical manipulations, such as principle component analysis

(PCA) and partial least squares (PLS) when building calibration sets.

To conduct analysis the instrument must be taught how to interpret spectra to build up

a calibration set. A calibration set is a set of numbers which convert the signal from

the detector into a predicted value of constants. The constants in the case of a NIR

instrument are assigned to each of the filters or grating slits, which are used to

conduct the scan. These applied constants are then multiplied by the absorbance

reading attained. The sum of these calculations is then plotted to give a representative

spectra. Initially a multiple linear regression is conducted (Equation (10)) which uses

various terms to describe a straight line;


MSc Analytical Chemistry - N. Blakiston 43

y = C0 + S(C1L1 + …+ CnLn) (10)

The term C0 is the intercept value and represents the S slope, the values for C are the

constants applied and have to be determined for each parameter that requires

measurement. The terms in the equation are log values and represent the output

information regarding the sample, where A is the apparent absorbance and is generally

plotted against the log function. The log values include the reflectance value R (the

amount of light reflected) and is described by;

A = log (1/R) (11)

These calculations are automatically applied by the instrument controller.

The most prominent absorption bands occurring in the NIR region are related to the

overtone and combination bands of the fundamental molecular vibrations of

-CH, -NH and -OH functional groups observed in the infrared spectral region.

Measurement is undertaken on the principle that the number of photons absorbed is

proportional to the concentration, but a strict adherence to Beer’s law is not the

general case when dealing with NIR spectra. This is not only due to matrix effects

such as particle size and compaction, but also to the complex nature of overlapping

bands in the spectra. A paper written by Miller (1993) 64 reviews a number of sources

that can cause non linearity in NIR methods including detectors, stray light and

chemical interactions.
MSc Analytical Chemistry - N. Blakiston 44

A good overview of NIR spectroscopy can be read in the paper by Morisseau et al,

(1995) 65.

Instrumentation

NIR radiation can be measured by either reflectance (R) or transmittance (T).

Reflectance is the ratio of the intensity of the radiation reflected by the sample (Ir) to

the intensity of the incident radiation (I0), equation (12). The instrument was used in

Diffuse Reflectance Mode for sample analysis of the powder blend samples.

R = Ir / I0 (12)

In diffuse reflectance, the radiation penetrates the surface layers, to an estimated depth

of 0 to 5 mm51 into the sample, and undergoes multiple reflections before re-emerging

after undergoing various characteristic absorptions.

The penetration depth depends on particle size, particle shape, degree of compaction

and the chemical nature of the material. In order to concentrate on the chemical

information, mathematical transformations such as spectral derivitisation can be used

to effectively remove the physical effects.

The radiation undergoes many internal reflections before emerging in all directions. A

schematic illustration of diffuse reflectance and transmittance is in Figure 11.3a.


MSc Analytical Chemistry - N. Blakiston 45

Figure 11.3a Diagram of diffuse reflectance and transmittance;

Showing a few of the many paths taken by the radiation.

Spectrophotometers are based on filter, grating, FT or acousto-optical tuneable filter

systems (AOTF), and diode arrays. Tungsten-halogen lamps are used as the energy

source whilst silicon, lead sulphide, indium gallium arsenide and deuterated triglycine

sulphate are common materials used in detectors.

The samples are placed on a turntable, which places each sample in turn over the

radiation source for measurement (Figure 11.3b). The NIR radiation in this

instrument is separated into discrete wavelengths by dispersing the light with a

holographic diffraction grating. The radiation is split equally into two beams by an

optical interference beam splitter and on to two mirrors mounted perpendicular to

each other, one of which is moveable parallel to the NIR radiation beam. The mirrors

reflect and recombine the two beams which can interfere constructively or

destructively depending on the position of the moveable mirror resulting in an


MSc Analytical Chemistry - N. Blakiston 46

interferogram. Spectral information can be obtained by combining the interferograms

and applying the mathematical function Inverse Fourier transform.

Figure 11.3b Sphere detector; The detector beneath

the sample is a sphere detector and will receive a high

amount of reflectance, so increasing the signal to noise

ratio.

Mathematical Pre-Treatment and Processing

Near-Infrared spectra can be very complicated with multiple components present in

the sample, overlapping bands, matrix effects and other physical properties of the

sample. For this reason complex mathematical manipulation can be conducted to

simplify the spectra by reducing the ‘Noise’ caused by light scattering.

Bruker Opus NIR Spectroscopic Quant-Software

The Quant software supplied by Bruker was used to acquire and analyse spectroscopic

data for this thesis. It was specifically designed for NIR spectral acquisition,

chemometric model development and routine analysis.

It comprised a graphical, user-friendly, interface and provided tools for instrument

performance assessment to the recommendations of the European Pharmacopoeia, the

United States Pharmacopeia Chapter 1119 on NIR analysis and fully 21 CFR Part 11
MSc Analytical Chemistry - N. Blakiston 47

(21 Code of Federal Regulations Electronic Records; Electronic Signatures) 41

compliant.

A full set of diagnostic tools were provided to conduct instrument performance tests

and the software was designed to ease method development in a regulatory

environment.

A tutorial manual and CD demonstrated hands on methods with sample spectra being

provided to facilitate understanding of the key features available.

The software included six spectral pre-treatments, multiple sample selection methods,

library identification and qualifications methods, regression methods and routine

analysis operations.

Theoretical Background

In general the aim of a NIR quantitative analytical method is to apply a suitable

mathematical solution to produce values as near to the ‘True’ value as possible. The

following gives a brief overview of the various steps undertaken to apply these

solutions.

The most demanding task in developing a NIR spectroscopic method is the

preparation of samples for calibration and the vast number of samples required to

develop a good calibration set. In developing a chromatographic or U.V.

spectroscopic method it is easy to produce standards covering the range of interest.

Interference analysis of excipients is routine and scaling dilutions to suit the detection

method is straight forward. This may be one of the draw backs that have stopped NIR
MSc Analytical Chemistry - N. Blakiston 48

analysis from exploding in popularity. For most pharmaceutical companies there is a

race to get the product from discovery to the shelf. The patents start ticking; hence

the revenues are greater the longer the market share is held.

In the early stages of development batch sizes are small and infrequent so a NIR

method is probably not appropriate and so HPLC, GC or U.V. methods are developed

and routinely used. However there is a great benefit in replacing these methods once

full production commences. By running the methods in parallel a NIR spectroscopic

technique can be developed with little effort. Once enough data has been collected

the replacement method can be fully developed reducing analysis time, consumables

and reagent costs.

In an ideal scenario the manufacturing samples could be read directly during

processing, reducing lead times and the risk of forward processing a poor quality

batch. The instrumentation and methods, once validated are simple to operate for

routine analysis reducing the requirement for highly trained operatives.

Mulitvariate Calibration

Most analytical test methods are conducted using a univariate calibration set. There

are several draw backs to this method of analysis;

• Outliers caused by additional unknown components are not recognized.

• Statistical fluctuations caused by detector noise are directly reflected by the

concentration values.
MSc Analytical Chemistry - N. Blakiston 49

• Peaks used for the analysis of multicomponent systems must be well

separated.

• The analysis of multicomponent systems assumes the validity of the Beer

Lambert Law.

Multivariate calibrations will take into account spectral features over a wide range,

thus these deficiencies can be eliminated. This method assumes that systematic

variations observed in the spectra are a consequence of the concentration change of

the components and the change in the infrared signal does not need to be linear.

The down side of multivariate analysis is the large number of samples that are

required to produce the required validation (hundreds or thousands of data points may

be required).

To develop a chemometric method the NIR instrument is taught the relationship to

analyte concentration using these calibration samples. To validate the test set two

methods can be used; “Cross Validation” and “Test Set Validation”. The former is

generally used for small sample sets and was chosen for the purpose of this project.

Cross Validation

With any NIR method the sample used for validating the system must not be part of

the calibration set. However in any calibration set each data point can be taken out in

turn and used to validate the calibration. The system will conduct this automatically

comparing the expected result to the actual result. To produce the validation set the

samples must be independently analysed quantitatively on a reliable test method.


MSc Analytical Chemistry - N. Blakiston 50

This will then generate a set of ‘True’ values, which can be assigned to the data points

in the NIR instrument. The calibration samples should cover at least the expected

range of the samples intended to be tested routinely. Ideally this should be beyond

any regulatory or control limits and the sample values should be homogenously

spaced across the concentration range.

Once the calibration samples have been acquired the known concentration values can

be applied ready for data manipulation and the application of statistical tools.

Processing the Data

There are various data pre-treatments that reduce light scattering effects due to surface

effects, sample compaction and changes in particle size. These scattering effects can

cause problems in analysis as they lead to base line shifts between spectra. It is

therefore advantageous to remove these effects prior to performing any data analysis.

When pre-processing the data a good correlation between the spectral data and the

concentration values is required. The first step is usually the application of Partial

Least Squares (PLS) regression 42. This is a method for constructing predictive

models when the factors are many and collinear. PLS was developed in the 1960’s by

Howard Wold as an econometric technique. It is used to extract multiple overlapping

components from samples containing various concentrations once true values have

been applied. It is generally used to predict a set of dependant variables from a (very)

large set of independent variables.


MSc Analytical Chemistry - N. Blakiston 51

A combination of the following methods can also be applied to improve the

correlation from the initial linear regression67, 68:

• Linear Offset Subtraction: shifts the spectra in order to set the y-minimum to

zero.

• Straight Line Subtraction: fits a straight line to the spectrum and sub tracts it.

• Vector Normalization: normalizes a spectrum by first calculating the average

intensity value and subsequent subtraction of this value from the spectrum.

Then the sum of the squared intensities is calculated and the spectrum is

divided by the square root of this sum.

• Min-max Normalization: first subtracts a linear offset and then sets the y-

maximum to a value of 2 by multiplication with a constant.

• Multiplicative Scatter Correction: performs a linear transformation of each

spectrum for it to best match the average spectrum of the whole set.

• First Derivative: calculates the first derivative of the spectrum. This method

emphasizes steep edges of a peak. It is used to emphasize pronounced, but

small features over a broad background. Spectral noise is also enhanced.

• Second Derivative: similar to the first derivative, but with a more drastic

result.

These mathematical formulations can be chosen and applied manually from a drop

down list within the system or the software can automatically run through all

connotations to derive successful rules to aid a good linear fit. Review of the

statistical correlation is then required by the operator before deciding on the required

mathematical application.
MSc Analytical Chemistry - N. Blakiston 52

Experimental Analysis Karl Fischer

Chemical Safety (CoSHH, Risk and MSDS)

Before any analytical work commenced a thorough review of the associated

health/hazard information was conducted.

This included;

• Material Safety Data Sheets (MSDS) for each reagent, API and Excipients

used (Effective date: 30-June-2005).

• Control of Substances Hazardous to Health (CoSHH) assessments for

Moisture determination by Karl Fisher - Raw Materials (Ref: 000123).

• CoSHH assessments for Moisture determination by Turbo Karl Fisher (Ref:

000285).

For all chemical reagents used, the Health Hazard Evaluation reviews state;

“Based on an assessment, as required by the corporate document Chemical Safety in

Laboratories the above mentioned chemicals / quantities do not exceed the quantity of

each chemical handled inside a ventilated enclosure that would exceed the

Occupational Exposure Limit (OEL).

This statement is based on the minimum requirements that suitable Personal

Protective Equipment (PPE) is worn at all times when conducting analysis (Lab coat,

safety glasses and nitrile gloves), and that all work is conducted in a fume cupboard”.
MSc Analytical Chemistry - N. Blakiston 53

Sample Preparation

Several methods of sample preparation were discussed in order to construct a

calibration set. It was known from previous stability analysis that Prozac routinely

contains 8% H2O rising to approximately 11% H2O after 36 months on stability. This

additional uptake of water is from the humid storage condition or dehydrating of the

gelatine capsule into the matrix.

The initial experiments were going to be focused on systematic drying of samples to

produce a range of moisture contents down to the limit of detection for the KF

apparatus (approximately 1%) followed with incremental ‘wetting’ of samples to

produce water contents above the nominal content.

An alternative approach to sample preparation, which is more broadly recognised, is

to use live manufactured batches and collect NIR data at the same time as normal KF

analysis slowly building a spectral library with batch to batch variation. The limits of

using this method are in constructing a good linear plot which accurately brackets the

expected theoretical values. Unfortunately Prozac is of relatively low manufacturing

volume and is run in periodic campaigns of approximately 10 batches over several

weeks. This would not have been enough to produce the multiple samples required

and would have taken to long in the time window required for this project.

For this reason a single batch was chosen for the analysis as this would give a good

indication of the feasibility of using NIR spectroscopy on production batches.


MSc Analytical Chemistry - N. Blakiston 54

Further discussion on the process of wetting included mixing a known quantity of

water with the blend or using a high humidity environment and allow the matrix to

absorb water. Neither method, it was thought, would deliver a suitable homogenous

blend, although the latter method appeared to be the better option. The process would

have included the use of a salt solution to deliver a known and constant % Relative

Humidity in a closed vessel (Desiccator) or stability chamber. Due to the

unpredictable nature of the absorption this was not pursued as part of this project.

Loss on Drying Determination

Stage one of the project was to establish the water content of the powder blend and to

identify the drying times that would be required to produce a linear calibration

between approximately 2% H2O and the un-dried value. Once this was identified

portions of the powder blend were dried for various periods before conducting Karl

Fischer analysis.

Good Manufacturing Practice (GMP) Requirements

Eli Lilly and Company Limited is regulated by the Medicines and Healthcare products

Regulatory Agency (MHRA) and all activities are under taken to current Good

Manufacturing Practice (cGMP). To this end any non-routine activity needs to be

documented and approved before any work is conducted. As stated previously the

material was chosen as it did not represent the finished pharmaceutical item and the

data generated would not impact any regulatory test requirement.


MSc Analytical Chemistry - N. Blakiston 55

All analytical work was documented prior to execution in the following technical

reports. This ensured that all work undertaken had prior approval and a set path of

analysis would be conducted.

Execution of Technical Reports

QCL-TR-014: Loss on Drying of Prozac Powder Blend

Purpose: The purpose of this protocol was to determine a baseline water content for

Prozac powder blend and to construct a time versus weight loss drying curve.

Results: The material was dried to constant weight in a period of 1 hour. The

percentage loss on drying was determined to be 7.74%.

Discussion and Conclusion: The percentage loss is in general agreement with the

predicted water content. It is known that there are no residual solvents in the raw

materials and the manufacturing process is a dry manufacture process, therefore this

value represents the great majority of water that can be liberated by gentle heating.

Due to the unexpected rapid drying of the sample a representative drying curve could

not be established, but this would add as a guide to further drying experiments.
MSc Analytical Chemistry - N. Blakiston 56

QCL-TR-015: Water Determination on Dried Prozac Powder Blend


by Karl Fischer Titration and Near-Infrared

Purpose: The purpose of this protocol is to generate a number of data points to

construct a quantitative model for the determination of water by NIR spectroscopy.

Results: The results generated were not reportable due to failed method acceptance

criteria.

Discussion and Conclusion: The execution of this protocol failed due to the poor

performance of the Karl Fischer method. This is discussed further below.

Investigation of Failed Execution: QCL-TR-015

When the KF titration result for the water standard check appears to be greater than

110% H2O, the test answer is printed as 999.000% H2O and the result is automatically

excluded from the statistics (see later discussion 12.7.1).

Observation of the course of the titration indicated that the peristaltic pump ran at

constant speed, delivering a steady stream of titrant, and then stopped. After a pause

of 5 seconds, as per method the system terminated the titration as it perceived a stable

end point had been achieved.

Once a sample has been administered, the titration should be started immediately and

rapidly. As the end point is approached the titration rate should be reduced, allowing

time between small aliquot additions, for the reactants to equilibrate. If the rate up to
MSc Analytical Chemistry - N. Blakiston 57

the end point is too rapid an excess of un-reacted titrant will be present and if the wait

period is not long enough (End-Point Time) the electrode will detect the excess titrant

and assume the end point is reached. If the wait time is too long then reproducibility

can become a problem. This is due to variations in the release of tightly bound

residual water or atmospheric ingress, resulting in sporadic over titration, or very long

titration times due to ingress of background water.

A balance therefore, needs to be struck between an accurate end point

‘approximation’ and the time taken to reach a stable reading. A stable end-point is a

significant indication of the course of a titration with no complications. A vanishing

end point would indicate that there is a slow release of water from the matrix or that a

side reaction may be interfering.

By setting a longer ‘End-Point Time’ in the Turbo KF method the requirement above

was met. The titration started rapidly until a point near the ‘End-Point’ was

approached. The instrument then went into the ‘End-Point Time’ mode for

approximately 6-8 seconds before adding small increments of titrant, pausing between

additions. This continued for approximately 20-40 seconds before the titration was

terminated. The experimental optimisation results are not documented or discussed

any further here, but a value of 10 seconds for the ‘End-Point Time’ was found to

give ideal results.

The method allows two modes of instrument operation ‘Sample Only’ and

‘Background Sample’. It was decided as an added precaution to use ‘Background

Sample’ mode to account for residual water and carryover. However this added
MSc Analytical Chemistry - N. Blakiston 58

considerable amount of time to each analysis with an overall testing time of

approximately 40 minutes per sample replicate.

Further investigation found two statements in a laboratory report “When using the

Mettler DL18 for moisture analysis in Atomoxetine® capsules the end point takes a

long time to be reached (typically 30-40 minutes per titration). Results can be

variable as residual water may remain within the sample matrix and therefore may not

be detected by the Karl Fischer. This is believed to be due to the presence of Silicone

fluid in the product formulation, restricting the release of water from the sample”.

Atomoxetine has a similar formulation to Prozac.

“It has been suggested that by adding a proportion of chloroform to the methanol in

the titration vessel that the solubility of the sample may be improved. This has been

proven to work well with compounds containing silicone oil17”. This alternative was

considered and a full review of available literature was conducted to understand the

impact of the solvents used.

Methanol does not dissolve long-chained hydrocarbons, such as starch satisfactorily.

The addition of 1-propanol to the working medium can also alleviate this problem.

Another benefit of 1-propanol addition is that it will remove oil deposits, from coating

the indication electrode. A literature reference32 states “Starch products will not

dissolve;…, formamide extracts the water from them extremely effectively. You can

optimise the extraction capacity by increasing the temperature (50 ˚C). The amount

of formamide …should not exceed 30%, or the stoichiometry…will change and the
MSc Analytical Chemistry - N. Blakiston 59

results falsified”. It was found that during operation, due to the long ‘Turbo’ times

and constant stirring that the temperature of the solution rapidly rose to between 45-

52 ˚C and remained relatively constant.

The changes made to the operation parameters of the instrument proved successful in

generating good reproducibility between samples and water standard checks gave

good response. Therefore no changes were made to the solvent composition.

This was an unexpected result from what was assumed a suitable method for the

analysis of water in Prozac. Modification of Global Method B07766 was not an

intended objective of this project. However the additional time taken to fully

understand and experiment with the various parameters of the Turbo2 KF proved

beneficial from a technical knowledge perspective.

As part of the investigation into the failure of Technical Protocol QCL-TR-015 the

site certification report for Global Method B07766 was reviewed. The summary of

this review and a general review of the method are discussed further.

Review of method AP1043-01

This method was developed for the determination of water in fluoxetine

hydrochloride capsules equivalent to 10 and 20mg fluoxetine. It was used to support

primary stability studies, but early indications proved water did not cause degradation

to the product on storage. There have been no regulatory limits applied to this test

and the method was deleted from use in 2001. The method and product was chosen

for experimentation, as there would be no regulatory impact to producing results from

live powder blend material.


MSc Analytical Chemistry - N. Blakiston 60

The Prozac powder blend under study has a water content of about 8% H2O. This

equates to approximately 30-60 mg of H2O per sample on the basis of a sample size of

375-750 mg respectively, as per method AP-1043-01 (equations (13) and (14)).

Literature values11 state a systematic error of +0.06% and a scatter of 0.24%, which

infers a total error of approximately 0.3% at this level of water determination.

Method acceptance criteria states that if the results calculated from equation (13) do

not lie between 30-60 mg H2O, then the weight of sample taken for analysis should be

adjusted using equation (14).

Weight (mg) H2O Sample = Sample Weight (mg) x % H2O Sample (13)
100%

Original Sample Weight (mg) x 45 (14)


Weigh (mg) H2O in Original Sample

Due to the nature of the analysis (the result generated for samples was an unknown)

with regards to the sample transfer method (discussed later) and the expected

variation in water contents, it was not possible to meet this method requirement. The

method stated to take approximately 690 mg (contents of three capsules). However

for the purpose of this investigation a weight between approximately 500 - 1000 mg

was deemed acceptable (see calculated values in Table 11.8a).

Performance checks

In order to ascertain the integrity of the instrumentation throughout the analysis, a

water check was performed at the end of a set of sample analysis. A water check is an
MSc Analytical Chemistry - N. Blakiston 61

addition of a defined (weighed) volume of water introduced into the reaction vessel

and titrated as per the samples. The result has confidence limits of 90.0% - 110.0%

associated with it that the water check result must fall between. This is often termed

‘method data acceptance criteria’.

Review of Site Certification Report B07766

Global Method B07766 was transferred to the Basingstoke Quality Control

Laboratory (Receiving Laboratory) on the 14th December 2000. The requirement for

a successful site transfer (certification) was via a collaborative study with the

Indianapolis Development Laboratory (Originating Laboratory). The batch used for

the transfer was the same as that used for the original method validation.

Previous testing of the batch obtained a mean value of 8.9%. It is not known on how

many samples this was conducted or the range/%RSD of the results. Acceptance

criteria were that the receiving laboratory will generate an average result from four

replicates, which falls between 6.9-10.9% H2O. Two replicates were analysed on two

separate days to give the following results;

Table 12.7.2 Site Certification of Method B07766.

Replicate 1 Replicate 2 Average


Set Up
(% water) (% water) (% water)
1 8.2 8.4 8.3
2 7.5 8.3 7.9
Mean 8.1

The data met the criterion. However the % RSD between replicates for ‘Set Up 2’

was 7.2 % and the % RSD between all four replicates was 5.0 %. The four highest %

RSD’s for replicate results generated during the experimental stage of this thesis was

5.0, 6.4, 7.0 and 9.3 % (the latter result was investigated at the time of analysis – see
MSc Analytical Chemistry - N. Blakiston 62

Dixon’s Q test discussion). This would suggest that the adjustments made to the

method improved the reproducibility significantly from the original method with an

average %RSD of 2.2% and a median of 1.2% for all results.

It was also noted that a sodium tartrate dihydrate control was run at the start of the

analysis each time to check the operation of the system, but no post control was run.

This may indicate why the issue of carry over was not noted during method transfer or

during routine analysis. The method states that analysis of a control is optional.

Influence of solvent on the Karl Fischer reaction

The stoichiometry (molar ratio of H2O:I2) depends on the type of solvent used. In an

alcoholic solvent such as methanol the stoichiometry is 1:1. In a non-alcoholic

solvent, such as Formamide the stoichiometry is 2:1. Studies by Eberius31

demonstrated that iodine and water react in the ratio of 1:1 if the percentage of

methanol in the solvent is greater than 20%. The method meets this requirement by

having a solvent composition of 2% Hydranal buffer in methanol, combined 50/50 v/v

with Formamide.

Sample Transfer to the Karl Fischer Instrument

During the familiarisation and method adjustment phase it was noted that the balance

being used was extremely unstable when trying to back weigh material between the

balance and the KF instrument. This was first thought to be due to the static

behaviour of the Prozac powder blend.

To eliminate the problem of static; a fresh glass weigh boat and antistatic ion source

was used in combination with anti static gloves and an earth connecting wrist strap.
MSc Analytical Chemistry - N. Blakiston 63

This modification made no significant difference to the stability of the balance. As

the fluctuation of the balance appeared to be in an upward direction (continually

increasing in weight), an alternative hypothesis was that the powder once weighed out

from the sealed weighing dish, was absorbing atmospheric water. This problem had

not been encountered when conducting familiarisation testing on un-dried bulk

powder, but presented an issue with dried samples. To test this, weighing was

conducted using a sealed vial. This required the following procedure;

• Weigh and tare container and sample.

• Transfer this weight to the KF instrument.

• Remove container, open KF sample port, remove lid of container and transfer

a quantity of powder into the vessel.

• Seal sample, close sample port and weigh sample container.

• Transfer weight to the KF instrument.

This method gave a stable balance reading and was deemed suitable for the

experiment. This method did increase the sample transfer time to between 20-30

seconds. The slight increase in time the sample port was open and the time the

sample container remained open was not thought to have increased any water ingress

significantly.

The NIR instrument uses 12 ml sample vials, which are sealed with a rubber stopper.

These vials were suitable for conducting analysis as they reduced sample handling

and transfer. Samples were prepared by weighing approximately 4 g of material into


MSc Analytical Chemistry - N. Blakiston 64

each vial. The vials were sealed with the rubber stopper and parafilm was used for

added protection against water ingress.

QCL-TR-016: Water Determination on Dried Prozac Powder Blend


by Karl Fischer Titration and Near-Infrared II

Purpose: The purpose of this protocol is to generate a number of data points to

construct a quantitative model for the determination of water by NIR spectroscopy.

Results: The results generated are presented in Table 11.8b. All results generated

met data acceptance criteria.

Discussion and Conclusion: When conducting the analysis for the 50 minute dried

sample 2 (sample 15 Rep 1 and Rep 2 in the protocol) the first result generated was

significantly different to the replicate (5.187% and 4.343% H2O) with a %RSD of

12.5. A third replicate was conducted which gave a result of 4.568% H2O. To decide

whether all data should be accepted and the result averaged or if the first replicate

should be removed a Dixon’s Q test was conducted. The results for all 50 minute

samples were used as the population and the result of 5.187% was reviewed as a

potential outlier.
MSc Analytical Chemistry - N. Blakiston 65

Dixon’s Q-test

The experimental Q-value (Qexp) is the ratio defined as the difference of the suspect

value from its nearest one divided by the range of the values (Q: rejection quotient).

Qexp = Xn – Xn-1 (15)


Xn – X1

The obtained Qexp value is compared to a critical Q-value (Qcrit) found in tables. This

critical value should correspond to the confidence level (CL) which is usually 95%.

If Qexp > Qcrit, then the suspect value can be characterized as an outlier and it can be

rejected.

The critical value of Q for eight data points at a 95% CL is 0.526.

Q = 5.187 – 4.947 = 0.268


5.187 – 4.293

As this value is smaller than Qcrit the three results will be averaged.

Evaluation of Karl Fischer Data

Data Acceptance Criteria

All Karl Fischer determinations were reported after the initial method development

stage. As indicated in Protocol QCL-TR-016 one sample had three replicates

conducted as there was an apparent high variance between results, as discussed above.

All standard water checks (Control Samples) passed the limit of 90-110% H2O.
MSc Analytical Chemistry - N. Blakiston 66

One replicate was repeated due to an assignable cause, this was agreed acceptable by

the responsible scientist reviewing the work. A higher than expected result was

generated. The instrument continued to titrate, for longer than normal during the end-

point time stage. Investigation indicated that the stopper in the sample port had not

been replaced correctly and air was able to enter the vessel. The result was deleted

and a fresh replicate was analysed (Un-dried Sample 1 result of 9.311% deleted).

A review of method acceptance criteria for AP1043-01 using equation (13) indicates

that 21 samples are outside of the required 30-60 mg H2O (Table 12.8a). This was to

be expected and does not appear to have any implication on the results. Additionally

the standard water checks conducted gave results between 26.63 - 30.54 mg H2O,

with the majority being below 30 mg H2O. The method state to use 25μl of water for

the instrument calibration and this was assumed to be appropriate for the performance

checks. To meet the method requirements it would therefore be more appropriate to

conduct the instrument calibration on 30 μl aliquots of water.


MSc Analytical Chemistry - N. Blakiston 67

Table 12.8a Karl Fischer Results: Method acceptance criteria

%H2O in Sample Sample Weight (g) Weight mg H2O


Rep. Rep. Rep. Rep. Rep. Rep.
Rep. 1 Rep. 2 Rep. 3
Sample 1 2 3 1 2 3
Dried Sample 1 3.248 3.313 1.06664 0.71044 35 24
QCL-TR-015 Samples run with QCL-

Dried Sample 2 2.720 2.882 0.64332 0.57232 17 16


Un-Dried Sample 1 8.184 8.279 0.60203 0.58962 49 49
15min-1 7.341 7.402 0.69742 0.53078 51 39
15min-2 7.896 7.425 0.52818 0.92134 42 68
TR-016

15min-3 7.428 7.378 0.52299 0.71828 39 53


15min-4 7.239 7.245 0.61071 0.55829 44 40
20min-1 7.781 7.349 0.53075 0.56326 41 41
20min-2 7.472 7.460 0.67764 0.64445 51 48
20min-3 7.529 7.347 0.57997 0.52877 44 39
20min-4 7.381 6.981 0.60058 0.61636 44 43

20min-1 6.334 6.802 0.68954 0.83398 44 57


6.719 6.647 0.54493 0.71329 37 47
QCL-TR-016 Samples with additional un-dried and dried

20min-2
20min-3 6.561 6.570 0.64731 0.52973 42 35
20min-4 6.641 6.928 0.56422 0.68015 37 47
30min_1 6.024 5.881 0.60333 0.65708 36 39
30min_2 5.869 5.928 0.68133 0.83157 40 49
30min_3 5.937 6.121 0.70193 0.74624 42 46
30min_4 5.542 5.583 0.70978 0.84698 39 47
samples

40min_1 6.134 6.029 0.68525 0.97489 42 59


40min_2 6.149 6.192 0.71237 0.68269 44 42
40min_3 6.377 6.281 0.64525 0.81971 41 51
40min_4 5.666 5.702 0.86135 0.60078 49 34
Un-Dried Sample 1 8.586 8.528 0.73939 0.75297 63 64
Dried Sample 3 4.450 4.441 0.74192 1.01197 33 45
50min_1 4.790 4.905 0.76289 0.76250 37 37
50min_2 * 5.187 4.343 4.568 0.90798 0.70370 0.80131 47 31 37
50min_3 4.756 4.947 0.80860 0.81679 38 40
50min_4 4.701 4.293 1.01671 0.79395 48 34

A051452 Undried 8.635 8.800 0.64054 0.54095 55 48


A051452 Undried 8.683 8.644 0.77375 0.60887 67 53
BAS 1607 page 34

A347575 Undried 7.870 7.656 0.76582 0.67774 60 52


A347575 Undried 7.493 7.585 0.64183 0.58869 48 45
A347575 70min 2.827 3.029 0.85102 0.80616 24 24
A347575 90min 2.267 2.158 0.84886 0.83842 19 18
A347575 110 2.519 2.567 0.87229 0.70916 22 18
A347575 130 1.896 1.874 0.78739 1.05112 15 20

Sample A Un-dried 8.290 8.084 1.23037 0.90941 102 74


BAS 1607 page 36

Sample B 20min 6.777 6.466 0.57779 0.68788 39 44


Sample C 50min 4.346 4.799 0.51722 0.56804 22 27
Sample D 120min 2.414 2.398 0.57691 0.73258 14 18

Sample A Un-
8.055 7.981 7.957 0.55285 0.53843 0.72517 45 43 58
dried**
Sample D 120min** 2.451 2.462 2.652 0.85870 0.63378 0.57292 21 16 15

* Three results generated for this sample


* Five replicate results generated for these sample.
MSc Analytical Chemistry - N. Blakiston 68

Table 12.8b Karl Fischer Results: Background Titre

Background Titre mg
%H2O in Sample
H2O
Sample Rep 1 Rep 2 Rep 3 Mean Diff. %RSD Rep 1 Rep 2 Rep 3
016QCL-TR-015 Samples run with QCL-TR-

Dried Sample 1 3.248 3.313 3.281 0.065 1.4 1.38 4.10


Dried Sample 2 2.720 2.882 2.801 0.162 4.1 1.25 3.10
Un-Dried Sample 1 8.184 8.279 8.232 0.095 0.8 3.10 2.01
15min-1 7.341 7.402 7.372 0.061 0.6 1.89 0.83
15min-2 7.896 7.425 7.661 0.471 4.3 1.05 2.34
15min-3 7.428 7.378 7.403 0.050 0.5 1.32 4.48
15min-4 7.239 7.245 7.242 0.006 0.1 2.65 4.66
20min-1 7.781 7.349 7.565 0.432 4.0 0.64 1.32
20min-2 7.472 7.460 7.466 0.012 0.1 1.17 1.27
20min-3 7.529 7.347 7.438 0.182 1.7 1.28 1.44

20min-4 7.381 6.981 7.181 0.400 3.9 1.74 1.90


samplesQCL-TR-016 Samples with additional un-dried and dried

20min-1 6.334 6.802 6.568 0.468 5.0 1.64 3.17


20min-2 6.719 6.647 6.683 0.072 0.8 2.78 2.84
20min-3 6.561 6.570 6.566 0.009 0.1 3.72 2.12
20min-4 6.641 6.928 6.785 0.287 3.0 3.60 4.05
30min_1 6.024 5.881 5.953 0.143 1.7 2.11 4.23
30min_2 5.869 5.928 5.899 0.059 0.7 4.70 4.32
30min_3 5.937 6.121 6.029 0.184 2.2 5.41 4.44
30min_4 5.542 5.583 5.563 0.041 0.5 4.10 4.48
40min_1 6.134 6.029 6.082 0.105 1.2 0.59 2.22
40min_2 6.149 6.192 6.171 0.043 0.5 2.03 1.66
40min_3 6.377 6.281 6.329 0.096 1.1 1.93 1.47
40min_4 5.666 5.702 5.684 0.036 0.4 1.42 3.23
Un-Dried Sample 1 8.586 8.528 8.557 0.058 0.5 1.52 0.80
Dried Sample 3 4.450 4.441 4.446 0.009 0.1 2.11 3.58
50min_1 4.790 4.905 4.848 0.115 1.7 4.32 3.66
50min_2 * 5.187 4.343 4.568 4.699 0.844 9.3 2.86 3.64 1.51
50min_3 4.756 4.947 4.852 0.191 2.8 4.89 3.26

50min_4 4.701 4.293 4.497 0.408 6.4 4.55 6.25

A051452 Undried 8.635 8.800 8.718 0.165 1.3


Printer Jammed - results
BAS 1607 page 34

A051452 Undried 8.683 8.644 8.664 0.039 0.3


not available.

A347575 Undried 7.870 7.656 7.763 0.214 1.9


A347575 Undried 7.493 7.585 7.539 0.092 0.9
A347575 70min 2.827 3.029 2.928 0.202 4.9
A347575 90min 2.267 2.158 2.213 0.109 3.5
A347575 110 2.519 2.567 2.543 0.048 1.3
A347575 130 1.896 1.874 1.885 0.022 0.8

Sample A Un-dried** 8.290 8.084 8.187 0.206 1.8 1.05 4.04


BAS 1607 page 36

Sample B 20min 6.777 6.466 6.622 0.311 3.3 0.35 2.76


Sample C 50min 4.346 4.799 4.573 0.453 7.0 3.13 4.12
Sample D 120min** 2.414 2.398 2.406 0.016 0.5 1.80 3.25

Sample A Un-dried** 8.055 7.981 7.957 8.073 0.074 1.6 3.73 2.77 2.41
Sample D 120min** 2.451 2.462 2.652 2.475 0.011 4.1 4.32 4.95 4.28

* Three results generated for this sample


** Five replicate results generated for these sample (%RSD Calculated on all 5).
MSc Analytical Chemistry - N. Blakiston 69

Review of Variance between Replicates

To understand if there was any correlation between the water content analysed and the

reproducibility between replicates a plot of average % water content versus % RSD

was conducted (Figure 12.8.2).

Plot of Average %H2O Vs %RSD


10.0 Figure 12.8.2. Replicate
9.0
8.0 y = -0.2945x + 3.9037
7.0
2
R = 0.0774 Variance; Correlation of
6.0
%RSD

5.0
values to identify if
4.0
3.0
2.0 measured water content is
1.0
0.0
1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5 independent of the
%H2O

reproducibility of replicates.

A relationship between the two parameters would be evident if the trend line had an

R2 value approaching 1.0 with data points falling on or close to it. The data points are

relatively random and no positive correlation appears to exist. The ‘Best Fit Line’

constructed suggests that reproducibility is slightly improved with increased water

content, but this is tenuous.

Background Sample Analysis

The result for the background water content varied between 0.35 – 6.25 mg. The

average value was 2.78 mg and the Standard Deviation on all results was 1.37 and the

% RSD was 0.01.


MSc Analytical Chemistry - N. Blakiston 70

Further investigation could be conducted to identify any relationship between the

background titre and the sample results. However this is beyond the scope of this

project. It can be stated that the background content of the vessel measured after 5

minutes of Turbo time is equivalent to between 1 and 15% of the water value in the

sample. This variance might be expected to have a consequence on the sample

replicates, but the reproducibility of sample replicates for this type of water release

profile appears little affected.

Without further analysis it is assumed that the background titre is giving a relatively

accurate result for the water content present between analyses.

Summary

Forty one samples in all were run on the Karl Fischer and the NIR instrument. Two

samples were of an un-dried stability batch A051452, the rest were un-dried and dried

powder blend samples from batch A347575.

Experimental Analysis NIR


Before any of the project samples were run initial training on the instrument was

conducted using various ‘doped’ organic solvents. These samples were ‘doped’ with

a range of water contents by micro syringe and produced excellent calibration sets

with R2 values greater than 99.5%. A sample of the dried bulk Prozac powder and un-

dried was also run and a disenable difference could be seen between the two spectra.

This was a critical step in the initiation and feasibility of this project.
MSc Analytical Chemistry - N. Blakiston 71

The instrument used was a BrukerTM MPA FT-NIR, using Opus Quant version 5.5

software. Procedure: BPD-071708-003-024 (BrukerTM MPA FT-NIR Spectrometer -

Routine Operation and Maintenance) was followed for all operations.

The instrument was used in integrating sphere mode as per Attachment: Default

Parameter Settings.

Performance Calibration check

Prior to performing any analysis with the instrument a daily system suitability test

was performed, for which certain criteria had to be met. An example of a successful

test report can be seen in Attachment 3 - Bruker PQ Test Protocol.

Typical automated, internal checks conducted by modern NIR instruments include;

• Wavelength Linearisation

• Photometric Instrument Noise

• Peak to Peak Noise

• Electronic Communication

• Bias

• Wavelength Precision, Accuracy and Bandwidth

These checks and adjustments are conducted against built in reference materials and

electronic diagnostic checks. The self calibration takes about 20 minutes to run and is

ready for operation immediately the test certificate is printed and verified. As I did

not have an account for the instrument all performance checks and sample analysis
MSc Analytical Chemistry - N. Blakiston 72

were observed by a trainer. For more information refer to the manufacturer’s

instrument manual67.

Sample Analysis

The sample vials were rotated by hand to evenly distribute the powder. Each sample

was placed on the sampling turntable and analysed once only for the determination of

the calibration set.

For the repeatability determination the sample was analysed then the vial was gently

shaken to redistribute the sample and read again. The samples were also randomly

tapped on the bench to increase compaction and bulk density.

Calibration Models

The 41 samples analysed were assigned their respective ‘True’ value as obtained from

the average result of the replicate Karl Fischer analysis. The overlain spectra for each

sample are depicted in Figure 13.3a.

Figure 13.3a An overlay of all 41 sample spectra; depicting the greatest change in

peak intensity at approximately 5180cm-1.


MSc Analytical Chemistry - N. Blakiston 73

It can be seen that the area around 5180cm-1 has a spectral peak which shows variation

with water content - Peak height increases with increasing water content. A similar,

but less apparent correlation can be seen in other areas of the spectra. There is clear

benefit to overlaying the spectra as you can immediately see if there are any ‘rogue’

spectra. A closer inspection of the primary spectral peak can be seen in Figure 13.3b.

Figure 13.3b Primary water peak at 5180cm-1 (combination band).

Mathematical Pre-treatment and Processing

The first step after the initial look at the sample spectra is to distinguish any spectral

peaks that appear to vary with concentration of the analyte. From this stage on it can
MSc Analytical Chemistry - N. Blakiston 74

be a case of trial and error (“No general recommendations can be given as to which

data processing methods should be used. The best method has to be found

empirically by trial and error” 67) with applying various pre-treatments and

chemometrics, or an individual may have a preferred selection of mathematical

applications they like to use or the system can analyse the data for you and give

various options to the operator.

Without any prior experience in NIR spectral analysis options 1 and 3 were attempted.

After much button pushing three acceptable models were chosen to compare and

contrast. Each gave a relatively good linear plot and spread of data points.

The three calibration models; Linear Regression, Vector Normalised + PCA and

System Optimised were saved in the data base as R1, R2 and R3 respectively.

Calibration of Data Set R1

The first model applied to the data was a simple linear regression;

In the simplest model Beer’s law applies and the constituent of interest has a single

absorption band and is contained in a non-absorbing matrix (Burns et al, 200136). If

the spectrophotometer also has no noise, non-linearity or any other electrical or

mechanical fault, then in this ideal case the absorbance peak is exactly proportional to

the concentration of the constituent. The slope and the intercept can be easily

determined and the usual formula for Beer’s law applies.

A = εcl (16)
MSc Analytical Chemistry - N. Blakiston 75

Where A is the absorbance, є is the molar absorption coefficient, c is the concentration

of the active and l is the pathlength. Where the pathlength is unity then ε is equal to

the slope of the line.

This relationship proved to exist for the primary water band as depicted above and so

the peak between 5454.4 cm-1 and 4902.3 cm-1 was integrated using the local maxima

and minima, representing the start and the end of the peak. The peak was integrated

using a forced baseline model, which dropped the local maxima down as a

perpendicular to define the area of interest. A linear regression was then applied with

no pre-treatment or other mathematical manipulation.

Calibration of Data Set R2

For the second calibration the same data was used but ‘Vector Normalisation’

mathematical pre-treatment was applied. Normalizing a vector produces a unit

vector. This unit vector is produced by measurement of the vector length, then

dividing the x, y and z values by this value. This then describes a new vector. The

sum of the squared intensities is calculated and the spectrum is divided by the square

root of this sum. A full theoretical view of various chemometric manipulations is

described by Cowe, et al. (1995) 69.

A Principal Component Analysis (PCA) was then applied to ‘clean’ the spectra. The

band of interest was picked at 5473.4 cm-1 to 3795.5 cm-1.

“PCA is mathematically defined as an orthogonal linear transformation that

transforms the data to a new coordinate system such that the greatest variance by any
MSc Analytical Chemistry - N. Blakiston 76

projection of the data comes to lie on the first coordinate (called the first principal

component), the second greatest variance on the second coordinate, and so on”

(Wikipidea).

PCA reduces the spread of the data set by retaining those characteristics of the data

set that contribute most to its variance (eigenvalue: matrices are broken down into

their eigenvectors, which are called factors or principle components). This reduces

the large multivariate data of the original spectra, and is undertaken by keeping lower-

order principal components and ignoring higher-order ones. Such low-order

components often contain the "most important" aspects of the data. With increasing

numbers of factors less and less variance is modelled until additional factors are only

modelling variations due to noise.

A PCA factor of 5 was applied with a rank of 4. This is the number of Partial Least

Squared regression algorithms that are applied to the data, which selects the best

correlation function between spectral and concentration matrix. It also determines the

number of principle components used; too few leads to poor reproduction of the

spectral data, hence spectral changes may not be observed (“underfitting”), too many

leads to the measurement of spectral noise in the regression (“overfitting”)

Calibration of Data Set R3

For this calibration the Quant software was used to perform an optimisation routine.

This performs multiple chemometric and pre-treatment operations to produce the best

possible solution for the calibration model. This process takes several minutes of

computation and provides several possible solutions, which the operator must then
MSc Analytical Chemistry - N. Blakiston 77

decide on. The integration of the best model was extremely complex and appeared to

have used multiple and varied manipulations across the whole of the spectral region

chosen. This included 1st and 2nd derivation, straight line subtraction, constant offset

elimination, various normalization techniques (vector, min, and max) and other

corrections.

The region of analysis was 7502.3 cm-1 to 5446.4 cm-1. This area does not look at the

primary combination water band, so it is assumed that other spectral information is

available, which relates to water concentration. This may also be due to a flatter

baseline across this region.

Comparison of R1, R2 and R3

All three calibration models demonstrated a visibly good linear correlation (Figures

13.8a, 13.8b and 13.8c).

Figure 13.8a Calibration plot for Model R1.


MSc Analytical Chemistry - N. Blakiston 78
MSc Analytical Chemistry - N. Blakiston 79

Figure 13.8b Calibration and Validation plot for Model R2.

Figure 13.8c Calibration and Validation plot for Model R3.


MSc Analytical Chemistry - N. Blakiston 80

Table 13.8 provides a comparison of the calibration models R2 and R3.

Table 13.8 Comparison of calibration models R2 and R3.

R2 R2 Cross R3 R3 Cross
Calibration Validation Calibration Validation
R2 96.31 94.42 99.82 95.74
RMSECV 0.451 0.394
RMSEE 0.386 0.0952

The determination coefficient (R2) gives the percentage of variance present in the true

component values, which is reproduced in the regression. R2 approaches 100% as the

fitted concentration values approach the true values. This is different to the

conventional use of R2, which approaches 1.0 with the true value (17):

(17)

Where; the residual (Res) is the difference between the true and the fitted value. Thus

the sum of squared errors (SSE) is the quadratic summation of these values (18).

(18)

The root mean square error of estimation RMSEE is calculated from this sum, with M

being the number of standards and R the rank. The true value is approached as the

value approaches zero:


MSc Analytical Chemistry - N. Blakiston 81

(19)

A comparison of the resulting analysis values, from cross validation, with the original

raw data allows the calculation of the predictive error of the complete data system, the

RMSECV (Root Mean Square Error of Cross Validation). This is a quantitative

measure for the mean accuracy of the predictive capability of the chemometric model.

The smaller the RMSECV, the better the quality of the model.

It can be concluded that model R3 gives a better linear fit than R2 for both the

calibration and cross calibration models. The RMSEE value for the calibration model

is significantly improved for the R3 model compared to R2. However the R3 model

is not significantly better than R2 for the cross validation model. This is confirmed by

visual comparison of the plots, which show an extremely tight population, close to the

line for the calibration plot of R3 compared to R2, but greater variability in the cross

validation plot. This is also seen by comparison of the R3 calibration and validation

R2 values of 99.82 and 95.74. This would indicate that all of the chemometrics

applied produce a good calibration fit, but is less effective at predicting the true value.

For R2 both calibration and validation plots demonstrate similar variation.

It is not so easy to compare model R1 to the other two models as the statistics applied

by the software are different. R1 produced a linear calibration model with a

correlation coefficient of 0.9736, which translates to an R2 value of 94.79. This is

lower than the other two models, which is apparent by visual comparison.
MSc Analytical Chemistry - N. Blakiston 82

When constructing the data models initially the data population resembled a non-

linear fit. This improved as more data points were modelled. It was first thought that

the powder blend had a maximum water saturation limit and a minimum quantity due

to bound water. This produced a correlation similar to a Boltzman sigmoid

distribution. This is still slightly evident in the linear regression model with a

flattening of the data points at the top and bottom of the chart.

The ‘Best-Fit’ line for each model can be seen to intersect very closely with the origin

and would suggest that measurement outside of the calibrated range is possible.

Repeatability Five Replicates - Between Methods

Repeatability between methods

Two samples of powder blend were analysed five times by KF analysis. The 120

minute dried sample was also analysed five times by NIR analysis to compare the

repeatability between the two methods. The results are shown in Table 13.9.1.

Table 13.9.1 Repeatability data for KF and NIR comparison.

A347575 A347575
Replicate R1 R2 R3
Undried 120 min.
1 8.290 2.414 2.626 2.666 2.428
2 8.084 2.398 2.513 2.651 2.377
3 8.055 2.451 2.576 2.676 2.66
4 7.981 2.462 2.536 2.755 1.839
5 7.957 2.652 2.522 2.785 2.707
Mean 8.073 2.475 2.555 2.707 2.402
sd 0.132 0.102 0.047 0.060 0.346
%RSD 1.632 4.126 1.825 2.199 14.388
Min 7.957 2.398 2.513 2.651 1.839
Max 8.290 2.652 2.626 2.785 2.707
MSc Analytical Chemistry - N. Blakiston 83

Review of water replicates by KF

For the Karl Fischer analysis there appears to be less variability on results at the

higher water concentration. This may be due to poor homogeneity within the dried

sample (binding, clumping etc.) or due to instrument precision at this lower level.

Further work would be needed to substantiate this. To check variability of the

sampling process the un-dried KF analysis was performed on variable sample

weights; 1.23037, 1.23037, 0.72517, 0.55285 and 0.53843 g respectively. The results

indicated that a lower reading was attained for the lower weights taken.

The alternative conclusion might be that a high result is attained by using large

weights. In this instance the first two samples had values outside the method

acceptance criteria of 30-60 mg of water (equation 6 and 7). The values were 102 and

74 mg of water respectively.

The range for both batches was approximately 0.3% absolute water content.

Review of water replicates by NIR

The five readings were taken by the same method as stated above, then compared

using the three models; R1, R2 and R3.

The mean value for models R1 and R3 fell between the minimum and maximum

value obtained by KF analysis. This would suggest that at this level they offer a

better predicted value than model R2.


MSc Analytical Chemistry - N. Blakiston 84

The repeatability for models R1 and R2 were better than those obtained by KF

analysis 1.8 and 2.2 respectively compared to 4.1 %RSD. Model R3 demonstrated a

high variance between replicates of 14.4 %RSD. On this basis, with this limited data,

the linear regression model offers good reproducibility with a mean predicted value

close to the true value.

Comparison of variance between methods

F-test

An F-test was conducted on each calibration model comparing the five replicate

readings on the NIR to the five replicate titrations on the reference method. This

compares the standard deviations of two data sets to establish whether they represent

normal populations and that the population variances are equal.

The calculation follows;

F = S12/S22 (20)

F is always calculated to have a value greater than 1. This value is then compare to

table values (Fcrit), which take into account both sample populations (in this case 5)

and the significance level (in this case 95%). The test used is a two-tailed test; as

results can be higher or lower than the theoretical values (for tests on the mean) or in

this case if the two standard deviation values differ significantly. It is useful to note

that Microsoft Excel has built in functionality to conduct this analysis;


MSc Analytical Chemistry - N. Blakiston 85

Table 13.9a F-test between KF and R1 for five replicate results

F-Test Two-Sample for Variances R1

Variabl Variabl
e1 e2
Mean 2.4754 2.5546
Variance 0.01043 0.00217
Observations 5 5
df 4 4
F 4.798
P(F<=f) 0.079
F Critical 9.604

Table 13.9b F-test between KF and R2 for five replicate results

F-Test Two-Sample for Variances R2


Variabl Variabl
e1 e2
Mean 2.4754 2.7066
Variance 0.01043 0.00354
Observations 5 5
df 4 4
F 2.945
P(F<=f) 0.160
F Critical 9.604

Table 13.9c F-test between KF and R3 for five replicate results

F-Test Two-Sample for Variances R3


Variabl Variabl
e1 e2
Mean 2.4022 2.4754
Variance 0.11946 0.01043
Observations 5 5
df 4 4
F 11.453
P(F<=f) 0.018
F Critical 9.604

It can be seen that F calculated, for R1 and R2 is less than F critical (table value) and

therefore any difference between the NIR data and the reference method is not

significant. The value of P (the probability) is also greater than the 95% confidence

level entered for alpha (0.025) in excel, for a two tailed test, which also confirms this
MSc Analytical Chemistry - N. Blakiston 86

result. For model R3 the opposite is true and therefore a significant difference is

present between the reference method and the NIR data.

t-test

This confirmation allows us now to conduct further analysis to compare the average

results from each data set to understand if there is any significance between them. If

no significant difference is found then this will suggest that the two sets of data, in

each case, could come from the same population.

The comparison of two means is undertaken using one of two types of t-test depicted

below;

There are two methods to compare the mean (average) values depending on the

results generated from the F-test these are;

t = (mean1 - mean2)/√( S12/n1 + S22 /n2) (21)

and

t = (mean1 - mean2)/S√( 1/n1 + 1/n2) (22)

Where S is the pooled standard deviation and the value t has two degrees of freedom

(2 × n-1). The latter equation will be used for comparison of the reference method to

R1 and R2 as this is applied for a successful F-test i.e. a‘t-test assuming equal

variance’. The value for alpha (α) was set at 0.05 (equivalent to 95% significance).
MSc Analytical Chemistry - N. Blakiston 87

Table 13.9d t-test between KF and R1 for five replicate results

t-Test: Two-Sample Assuming Equal Variances


R1 Vs KF

Variable R1 Variable KF
Mean 2.555 2.475
Variance 0.0022 0.0104
Observations 5 5
Pooled Variance 0.0063
Hypothesized Mean
Difference 0
df 8
t Stat 1.5774
P(T<=t) two-tail 0.1534
t Critical two-tail 2.3060

Table 13.9e t-test between KF and R2 for five replicate results

t-Test: Two-Sample Assuming Equal Variances


R2 Vs KF
Variable R2 Variable KF
Mean 2.707 2.475
Variance 0.0035 0.0104
Observations 5 5
Pooled Variance 0.0070
Hypothesized Mean
Difference 0
df 8
t Stat 4.3736
P(T<=t) two-tail 0.0024
t Critical two-tail 2.3060

For the R1 model; As t Stat is less than tcrit (table value) and the value for P is greater

than the value for α, then any difference between the means is not significant. This

indicates that the two sets of data have come from the same population.

For the R2 model; As t Stat is greater than tcrit and the value for P is less than the

value for α, then any difference between the means is significant.


MSc Analytical Chemistry - N. Blakiston 88

ANOVA on 5 replicates

The results of the F-test and t-test appear to prove that model R1 (linear regression

with no mathematical manipulation) has produced a sample mean and standard

deviation within the same population as the reference method for five replicate

analysis.

To confirm this, an Analysis of Variance (ANOVA) will be conducted to compare all

four data sets;

Table 13.9f ANOVA: Single Factor.

SUMMARY
Averag Varianc
Groups Count Sum e e
Reference
Method 5 12.377 2.475 0.0104
R1 5 12.773 2.555 0.0022
R2 5 13.533 2.707 0.0035
R3 5 12.011 2.402 0.1195

ANOVA
Source of
Variation SS df MS F P-value F crit
Between Groups 0.25509 3 0.08503 2.50808 0.09582 3.23887
Within Groups 0.54244 16 0.03390
Total 0.797534 19

If Fcalc < Fcrit and Pvalue is > than α (0.05), then there is no significant difference between

the data sets.

This proves that although the ANOVA indicates that all four data series are from the

same population the F-test and t-test prove that only model R1 is actually

representative to the reference method.


MSc Analytical Chemistry - N. Blakiston 89

Confidence Limits

To assess how representative the mean data and standard deviation of each data set

are, with respect to the reference method we can look at the confidence limit for each

mean value. This will indicate any systematic error associated with the data.

The calculation used for this analysis is;

μ = mean ± t(s/√n) (23)

where;

μ (greek letter mu) = the theoretical value 2.475 % H2O

t = critical table value for n-1 degree of freedom at the 95%

confidence interval.

n = number of samples.

Model R1

μ = 2.555 ± 2.57(0.047/√5)

μ = 2.555 ± 0.054 % H2O

Model R2

μ = 2.707 ± 2.57(0.06/√5)

μ = 2.707 ± 0.068 % H2O

Model R3

μ = 2.402 ± 2.57(0.346/√5)

μ = 2.402 ± 0.398 % H2O

As μ = 2.475 % H2O it can be seen that the theoretical value μ does not fall within the

confidence limit of Model R1 or R2, but does for R3. This demonstrates that an
MSc Analytical Chemistry - N. Blakiston 90

inherent, systematic error is involved in the models for R1 and R2 using five replicate

readings. However this is not unexpected as model R3 has a mean close to the true

value and a high variance, so this does not necessarily indicate this is the better

method.

Random Error

Further determination was conducted to identify if any difference between mean

model values and the reference method were solely as a result of random error. The

lower the probability of such a difference occurring by chance, the less likely it is that

the null hypothesis is true. The null hypothesis is generally rejected if the probability

of such a difference occurring by chance is less that 1 in 20 (i.e. 0.05 or 5%) and in

such a case the difference is said to be significant at the 0.05 level. The following

equation is used for this analysis;

t = (mean - μ)√n/s) (24)

Where t is the critical table value for n-1 (tcrit = 2.78) samples and μ is the expected or

theoretical quantity, with s being the standard deviation.

For Model R1

t = (2.555 - 2.475) √5/0.047


t = 3.81

For Model R2

t = (2.707 - 2.475) √5/0.06


t = 8.65

For Model R3

t = (2.402 - 2.475) √5/0.346


t = 0.472
MSc Analytical Chemistry - N. Blakiston 91

This would indicate that systematic error is involved and model R1 and R2 fails and

that there is significant bias above the calculated mean at the lower end of the

determination (circa 2.5 % H2O). This also indicates that model R3 is a more

acceptable solution at low water content levels.

Repeatability 16 Replicates - Within Method

Repeatability NIR measurements

Sixteen readings were taken for Batch A051452 using the above strategy to present

the sample in different orientations. The sample vials were sealed throughout the

analysis. The results of the 16 readings indicate any variance (absolute) between the

instrument performance or the bulk powder composition/orientation. The readings

taken were then calculated as predicted values from each of the three calibration

models and compared to the ‘True’ value, as produced for the reference KF method.

The results are depicted in Table 13.10a.


MSc Analytical Chemistry - N. Blakiston 92

Table 13.10a Predicted NIR values Vs KF - Repeatability

Predicted Values Taken From the Calibration Plots %H2O


Linear Regression Vector Normalization + PCA System Self Optimisation
R1 R2 R3
Result Residual Result Residual Result Residual
8.229 0.435 8.625 0.039 8.613 0.051
8.263 0.401 8.766 0.102 8.556 0.108
8.269 0.395 8.876 0.212 9.056 0.392
8.400 0.264 8.884 0.220 8.911 0.247
8.287 0.377 8.834 0.170 8.995 0.331
8.365 0.299 8.834 0.170 8.772 0.108
8.356 0.308 8.869 0.205 8.687 0.023
8.138 0.526 8.913 0.249 9.382 0.718
8.241 0.423 8.974 0.310 9.213 0.549
8.275 0.389 9.026 0.362 9.563 0.899
8.235 0.429 8.959 0.295 9.095 0.431
8.192 0.472 8.863 0.199 8.444 0.220
8.057 0.607 8.994 0.330 8.722 0.058
8.163 0.501 8.938 0.274 8.387 0.277
8.088 0.576 9.046 0.382 9.859 1.195
8.057 0.607 9.155 0.491 8.902 0.238

Max 8.400 0.607 9.155 0.491 9.859 1.195


Min 8.057 0.264 8.625 0.039 8.387 0.023
Range 0.343 0.343 0.530 0.452 1.472 1.172
Residual 0.438 8.226 0.246 8.413 0.283 8.299
Mean 8.226 0.438 8.910 0.251 8.947 0.365
Median 8.238 0.426 8.899 0.235 8.907 0.262
SD 0.105 0.105 0.123 0.112 0.407 0.331
%RSD 1.276 23.956 1.376 44.531 4.554 90.499

True Value from KF = 8.664 %H2O Average of two replicates; 8.683 and 8.644 %H2O

No mean result for the calibration models falls within the upper and lower replicate

values for the reference method. This was not unexpected as the %RSD for the KF

replicates was 0.3. The %RSD and range for model R1 was better than R2 and both

were significantly better than R3. The mean result for R1 showed a low bias while

R2 and R3 showed a similar mean (%RSD between them of 0.3) with high bias. This

data was too limited to reach any further conclusions in comparison to the true mean

due to the tight nature of the values around the mean. Repeat analysis of this

experiment would elucidate the best performing model. Further statistical analysis

would established how many replicate NIR readings are required to give the required

confidence level that the average falls within a suitable value of the true result.
MSc Analytical Chemistry - N. Blakiston 93

To accomplish this, a frequency distribution can be conducted and the confidence

limits can be established using;

mean ± tn-1 (s/√n) (24)

The value for t can be taken from confidence tables for a certain number of degrees of

freedom at a defined confidence interval.

ANOVA on 16 Replicates

This was conducted as previously on the five replicates;

Table 13.10b ANOVA: Single Factor - R1:R2:R3

SUMMARY
Averag Varianc
Groups Count Sum e e
R1 16 131.615 8.226 0.0110
R2 16 142.556 8.910 0.0150
R3 16 143.157 8.947 0.1660

ANOVA
Source of Variation SS df MS F P-value F crit
Between Groups 5.277 2 2.638 41.207 6.76E-11 3.20432
Within Groups 2.881 45 0.06403

Total 8.158 47

If Fcalc < Fcrit and Pvalue is > than α (0.05), then there is no significant difference between

the data sets. There is significant difference between the three data sets. To test if this

failure is due to model R1, the data set is removed and ANOVA conducted again.

There is no significant difference between model R2 and R3 once R1 is removed;


MSc Analytical Chemistry - N. Blakiston 94

Table 13.10c ANOVA: Single Factor - R2:R3

SUMMARY
Averag Varianc
Groups Count Sum e e
R2 16 142.556 8.910 0.0150
R3 16 143.157 8.947 0.1660

ANOVA
Source of Variation SS df MS F P-value F crit
Between Groups 0.011 1 0.0113 0.1247 0.7265 4.1709
Within Groups 2.716 30 0.0905

Total 2.727 31

Outlier Check for Repeatability

A Dixon’s Q-test was applied to each of the 16 samples results for each of the

calibration models. No outliers existed in the data set predictions for either models

(Table 13.10c).

Table 13.10c Dixon’s Q-test applied to NIR repeatability study.

Dixon's Q test
R1 R2 R3
8.057 8.625 8.387
8.057 8.766 8.444
8.088 8.834 8.556
8.138 8.834 8.613
8.163 8.863 8.687
8.192 8.869 8.722
8.229 8.876 8.772
8.235 8.884 8.902
8.241 8.913 8.911
8.263 8.938 8.995
8.269 8.959 9.056
8.275 8.974 9.095
8.287 8.994 9.213
8.356 9.026 9.382
8.365 9.046 9.563
8.400 9.155 9.859

Qexp
0.000 0.266 0.039
Upper
Qexp
0.102 0.206 0.201
Lower

Qcrit for 15 data points = 0.384


Qcrit for 20 data points = 0.342
MSc Analytical Chemistry - N. Blakiston 95

Residual Analysis - Frequency

The residuals have been plotted below for each of the models to identify if they

follow an expected distribution with the highest frequency closest to the true value;

R1 follows an expected distribution with the higher percentage of residuals for the 16

replicates falling near the true value. R2 appears on this data set to have a bias around

0.2 to 0.3 % away from the true value and R3 appears random. This would indicate

that for replicate analysis R1 performs better.

Figure 13.10.4 Frequency plots for calibration model R1, R2 and R3.

1.25 Quantiles R1
100.0% maximum 1.1950
99.5% 1.1950
1
97.5% 1.1950
90.0% 0.9878
75.0% quartile 0.5195
0.75
50.0% median 0.2620
25.0% quartile 0.1080
10.0% 0.0426
2.5%0.5 0.0230
0.5% 0.0230
0.0% minimum 0.0230
0.25
Moments
Mean 0 0.3653125
Std Dev 0.3306044
Std Err Mean 0.0826511
upper 95% Mean 0.5414792
lower 95% Mean 0.1891458
N 16

0.5 Quantiles
100.0% maximum 0.49100
99.5% 0.49100
0.4
97.5% 0.49100
90.0% 0.41470
75.0% quartile 0.32500
0.3
50.0% median 0.23450
25.0% quartile 0.17725
10.0% 0.08310
2.5%0.2 0.03900
0.5% 0.03900
0.0% minimum 0.03900
0.1
Moments
Mean 0 0.250625
Std Dev 0.1116064
Std Err Mean 0.0279016
upper 95% Mean 0.3100958
lower 95% Mean 0.1911542
N 16
MSc Analytical Chemistry - N. Blakiston 96

0.65 Quantiles
100.0%
0.6 maximum 0.60700
99.5% 0.60700
0.55
97.5% 0.60700
90.0% 0.60700
75.0%
0.5 quartile 0.51975
50.0% median 0.42600
25.0%
0.45 quartile 0.38000
10.0% 0.28850
2.5%0.4 0.26400
0.5% 0.26400
0.0%
0.35 minimum 0.26400

0.3
Moments
0.25
Mean 0.4380625
Std Dev 0.1049403
Std Err Mean 0.0262351
upper 95% Mean 0.4939812
lower 95% Mean 0.3821438
N 16

Review of Residuals for Repeatability

The residuals for each model were calculated by comparing the predicted values to

the true, reference value. The results are shown below.


MSc Analytical Chemistry - N. Blakiston 97

Table 13.10d Residual analysis for predicted versus true for NIR.
R1 R2 R3
Predicte Predicte Predicte
Sample TRUE
d
Residual
d
Residual
d
Residual

Prozac2blend_1.0 6.568 6.788 -0.220 6.611 -0.043 7.043 -0.475


Prozac2blend_2.0 6.683 6.684 -0.001 6.474 0.209 6.362 0.321
Prozac2blend_3.0 6.566 7.142 -0.577 7.141 -0.576 6.457 0.109
Prozac2blend_4.0 6.785 6.881 -0.097 7.059 -0.275 6.601 0.184
Prozac2blend_5.0 5.953 6.189 -0.237 5.982 -0.030 6.337 -0.385
Prozac2blend_6.0 5.899 6.395 -0.496 6.261 -0.363 6.261 -0.363
Prozac2blend_7.0 6.029 6.274 -0.245 6.145 -0.116 6.094 -0.065
Prozac2blend_8.0 5.563 6.136 -0.574 5.971 -0.409 6.166 -0.604
Prozac2blend_9.0 6.082 5.533 0.549 5.261 0.821 5.444 0.638
Prozac2blend_10.0 6.171 5.649 0.522 5.516 0.655 5.578 0.593
Prozac2blend_11.0 6.329 5.701 0.628 5.548 0.781 6.258 0.071
Prozac2blend_12.0 5.684 5.706 -0.022 5.450 0.234 5.346 0.338
Prozac2blend_13.0 4.848 7.787 -2.940 4.549 0.299 4.586 0.262
Prozac2blend_14.0 4.699 5.087 -0.388 4.854 -0.155 5.077 -0.378
Prozac2blend_15.0 4.852 5.162 -0.311 5.011 -0.160 4.933 -0.082
Prozac2blend_16.0 4.497 4.801 -0.304 4.762 -0.265 4.570 -0.073
Prozac2blend_Dried.0 4.446 4.234 0.212 4.147 0.299 3.711 0.735
Prozac2blend_UnDried.0 8.557 7.868 0.689 8.021 0.536 8.001 0.556
Prozacblend_10.0 7.661 7.514 0.147 7.595 0.066 7.765 -0.105
Prozacblend_11.0 7.403 7.537 -0.134 7.706 -0.303 7.526 -0.123
Prozacblend_12.0 7.242 7.556 -0.314 7.651 -0.409 7.854 -0.612
Prozacblend_13.0 7.565 7.126 0.439 7.183 0.382 7.338 0.227
Prozacblend_14.0 7.466 6.964 0.502 6.811 0.655 6.988 0.478
Prozacblend_16.0 7.181 7.019 0.162 7.038 0.143 7.197 -0.016
Prozacblend_A051452_1.0 8.718 8.236 0.481 8.663 0.054 8.987 -0.270
Prozacblend_A051452_2.0 8.664 8.229 0.435 8.595 0.068 8.333 0.330
Prozacblend_A347575_1.0 7.763 8.031 -0.268 8.095 -0.332 7.694 0.069
Prozacblend_A347575_110minutes.0 2.213 2.419 -0.207 2.534 -0.322 2.393 -0.181
Prozacblend_A347575_130minutes.0 1.885 1.885 0.000 2.054 -0.169 1.376 0.509
Prozacblend_A347575_2.0 7.539 7.870 -0.331 7.973 -0.434 7.703 -0.164
Prozacblend_A347575_90minutes.0 2.543 2.628 -0.085 2.840 -0.297 2.746 -0.203
Prozacblend_Dry_1.1 3.281 2.421 0.860 2.703 0.578 3.104 0.177
Prozacblend_Dry_2.0 2.801 2.056 0.745 2.293 0.508 2.382 0.419
Prozacblend_Original.0 8.292 7.849 0.443 7.761 0.531 7.983 0.309
Prozacblend_9.0 7.372 7.431 -0.059 7.682 -0.311 7.871 -0.500
Prozacblend_15.0 7.438 7.047 0.391 7.080 0.358 7.281 0.157
Prozacblend_A347575_120.1 2.406 2.626 -0.220 2.741 -0.335 2.826 -0.420
Prozacblend_A347575_20min.0 6.622 7.301 -0.680 7.367 -0.746 7.070 -0.449
Prozacblend_A347575_50min.0 4.573 5.505 -0.933 5.809 -1.237 3.945 0.628
Prozacblend_A347575_70min.0 2.928 3.671 -0.743 3.667 -0.739 3.837 -0.909
Prozacblend_A347575_Undried_A.0 8.187 8.014 0.173 8.230 -0.043 8.170 0.017

Mean 0.000 -0.022 0.018


Min 0.000 0.030 0.016
Max 0.933 1.237 0.909

It can be seen from the mean residual values that each model on average produces a
result within 0.02% of the true water content. In particular the mean result for model
R1 reports a value of 0.000%. The greatest error from the true value for each model
was 0.933, 1.237 and 0.909 % water respectively. For model R1 and R2 this is the
same data point.
MSc Analytical Chemistry - N. Blakiston 98

Matched Pair Analysis


The matched pair analysis is a comparison of the two sets of results; predicted and

actual (true – column 1). The analysis shows that there is reasonable correlation

between the results. However there is also significant spread, most likely due to the

error associated with the reference method and that of the NIR method.

Figure 13.10.5 Matched pair plots for calibration model R1, R2 and R3.

Difference: Column 2-Column 1 - R1

1.0

0.5
Difference: Column 2-Column 1

0.0

-0.5

-1.0
1 2 3 4 5 6 7 8 9
Mean: (Column 2+Column 1)/2

Column 2 5.95005 t-Ratio -0.00071


Column 1 5.9501 DF 40
Mean Difference -4.9e-5 Prob > |t| 0.9994
Std Error 0.06887 Prob > t 0.5003
Upper95% 0.13915 Prob < t 0.4997
Lower95% -0.1392
N 41
Correlation 0.97363
MSc Analytical Chemistry - N. Blakiston 99

Difference: Column 4-Column 1 - R2

1.0

0.5
Difference: Column 4-Column 1

0.0

-0.5

-1.0

1 2 3 4 5 6 7 8 9
Mean: (Column 4+Column 1)/2

Column 4 5.97156 t-Ratio 0.301272


Column 1 5.9501 DF 40
Mean Difference 0.02146 Prob > |t| 0.7648
Std Error 0.07124 Prob > t 0.3824
Upper95% 0.16545 Prob < t 0.6176
Lower95% -0.1225
N 41
Correlation 0.97179
MSc Analytical Chemistry - N. Blakiston 100

Difference: Column 6-Column 1 - R3


1.0

0.5
Difference: Column 6-Column 1

0.0

-0.5

-1.0
1 2 3 4 5 6 7 8 9
Mean: (Column 6+Column 1)/2

Column 6 5.93156 t-Ratio -0.2977


Column 1 5.9501 DF 40
Mean Difference -0.0185 Prob > |t| 0.7675
Std Error 0.06227 Prob > t 0.6163
Upper95% 0.10731 Prob < t 0.3837
Lower95% -0.1444
N 41
Correlation 0.97891
MSc Analytical Chemistry - N. Blakiston 101

Discussion of Results
The reference method (Karl Fischer Titration) has been discussed at length to have

not been entirely reliable. With this in mind the NIR models can only be as good as

the data used to produce them. However the more data that can be used in the

calibration model the greater the accuracy of the method. This is especially true if the

data is adjusted by removal of outliers.

The problems associated with the reference method are nicely surmised by

MacDonald et. al;

“Karl Fischer methods are commonly used to determine water contents, but can be

prone to errors in all but experienced hands, and risk compromising the sample when

the vial is opened to atmosphere in order to dissolve and titrate the contents” 51.

He further stated that for quantitative analysis;

“The calibration set must comprehensively cover all variations in analyte and matrix

components likely to be encountered in real samples, in order that the calibration

equation will be robust and reliable” 51.

In this instance the calibration models were mainly determined on dried samples of

one bulk batch and it is not known what impact this has had on the matrix. The one

sample run that was of a different batch (A051452) gave results from the true value of

0.07% to 0.48% H2O for the three models used. This is within acceptable limits of the

reference method, which had variability between replicates of up to 0.884% H2O, with
MSc Analytical Chemistry - N. Blakiston 102

an average of 0.164% H2O between replicates. To put this into perspective the site to

site transfer validation of the KF method had a variance of 0.8 % H2O between

replicates.

A disadvantage with multivariate analysis is that the prediction model often requires

an extensive calibration set to account for all of the possible sample variances in order

to achieve optimal performance. The calibration samples also need to resemble those

of real analytical samples and this is not always possible.

Brutsche (1996) 75 states; “For simple one component preparations, 40-60 spectra are

enough for calibration”, and “A basic requirement is that different batches of

excipients and active ingredients, as many as possible…”

Therefore this initial feasibility will need further development to incorporate a greater

range of batches to include the required variation in matrix. This was also borne out

by Plugge et al. whom used 474 results obtained by the spectroscopic method that

were compared to 4952 batch samples run by KF analyses72. This gives some idea of

the large amount of data required to produce a good calibration model. For

pharmaceutical use there may a limited number of batches made per year. For Eli

Lilly at Basingstoke there are several products that are only made a few times per year

and less than one hundred for even the highest volume product.

The analysis of 16 repeat readings by NIR demonstrated that there was a significant

variability. As the calibration models were produced by only one reading each time it

is assumed that improvement would be made with the averaging of multiple readings.
MSc Analytical Chemistry - N. Blakiston 103

For instance Webster et al. states “Each spectrum from the FT-NIR was an average of

20 readings”70. Windham et al. measured the sample 10 times and produced a 0.39 %

RSD for forage samples76 and achieved R2 values between 0.84 - 0.97 for the

calibration model, which gave values of 7.43-9.11% water for 20 forage samples.

The use of complicated chemometric techniques and mathematical manipulations lend

the technique of NIR analysis to statisticians rather than the humble analytical chemist

and this may be another reason for the slow uptake of the technique in an industry

dominated by chromatography. The use of NIR spectroscopy is a fantastic

opportunity for the future development of quality control in the pharmaceutical

industry and acceptance by the regulating bodies (discussed later) is an encouraging

sign.

The forty one samples used for the calibration model demonstrated that good linear

response can be achieved. This is true no matter whether a simple linear regression is

used, or a complicated multivariate manipulation is conducted. The R2 values for the

limited data were more than acceptable and demonstrated a good visual fit. The

predicted values showed a variation of up to 1.237%, which could be improved I am

sure with further work. However even at this level of error the method may be

suitable for certain applications. For example in a drying process for bulk material

this level of accuracy may be acceptable to validate a total blend time. For finished

product this may be acceptable if the expected results are significantly below

regulatory limits. It may be possible to use this as a check method and only

implement the reference method if a result above a certain control limit is obtained.
MSc Analytical Chemistry - N. Blakiston 104

The statistical analysis used to review the three NIR models did not highlight any one

model being superior to another. For pure simplicity and relatively good performance

in the analysis model R1 (linear regression) appears to have performed slightly better

than the other two models. As more data is collected this may be confirmed or

negated.

Further Discussion on Results

The NIR procedure had several advantages compared to the reference technique;

• Cost

• Speed

• Safety

• Ease of Use

• Non-Destructive

• Investigation of Suspect Results

Cost

The Bruker NIR Instrument was purchased at a cost of £45K this is compared to the

combined cost of £20K for the Turbo2 titrator and associated balance.

Servicing costs for the two instruments are comparable. The Karl Fischer requires

additional routine maintenance to replace titrant tubing and molecular sieve. This

reduces the resources available for testing (see speed).


MSc Analytical Chemistry - N. Blakiston 105

Consumables for the NIR include sample vials and stoppers. This is a relatively

cheap commodity. The Karl Fischer reagents and consumables cost, on average £200

per month, assuming that approximately 60 batches per month are analysed.

Within the Eli Lilly laboratories at Basingstoke there are four Karl Fischer

instruments, which include three different models. These are required to give flexible

use of resources, but also because different methods require different instrumentation

and set-up conditions. The KF instrumentation has only one purpose and that is to

test the concentration of water in a sample. The ability of NIR spectroscopy to

replace many different pieces of test equipment makes this technique an attractive

proposition. It is already being used for excipients and API identification and could,

potentially make the Karl Fischer equipment redundant.

Speed

The NIR calibration process required lengthy analysis by the KF reference technique

to produce the required number of data points for the NIR model. Once the

calibration had been established and is stored in the NIR instrument it is available for

use. A daily NIR performance check is conducted, which requires approximately 2-5

minutes of analyst time to run. Quantitative determinations can then be performed in

less than five minutes per sample, depending on the number of readings required. The

run time to analyse 12 samples by NIR spectroscopy for the determination of water

took approximately 30 minutes. The same analysis conducted by the KF method took

approximately 12 hours.

As Analyst time is costed to approximately £40 p/hr this can be directly related into a

saving of nearly £500.


MSc Analytical Chemistry - N. Blakiston 106

The simplicity of the NIR analysis passes on another advantage in that there is a huge

saving in training time and possibly validation time in the transfer of methods from

one laboratory to another.

This would need to be balanced against the volume of batches that require to be tested

and the time involved in validating a suitable method.

Non-destructive

NIR is a non destructive technique and any suspect analytical result can be fully

investigated. This may be by repeat analysis or confirmation by a reference method.

Karl Fischer is a destructive technique, which makes any investigation potentially

very difficult. Karl Fischer analysis has a high potential to produce failures within the

laboratory and is one of the techniques that produce the most investigations at Eli

Lilly. There are several reasons for this;

• Most methods have acceptance criteria, which may be a range or %RSD

between replicates. Due to the nature of the analysis this will occasionally

fail.

• There is a number of sample transfer steps required and any mistake by the

analyst or link between the balance and KF instrument can cause failure.

• Some methods require a final method suitability check to be conducted. If this

is not conducted the test needs to be repeated.

• Routine performance checks between samples may fail. The reasons for

failure are usually difficult to investigate.


MSc Analytical Chemistry - N. Blakiston 107

• The sample may be spilt when transferred.

• The balance may not be stable when a reading is taken or environmental

conditions may effect the balance operation.

• The wrong reagents may be used.

• Reagents may be exhausted during analysis.

• Poor judgement in sample preparation by the analyst.

• Tablets have been known to stick to the vessel and do not disintegrate causing

failure.

All of these issues have occurred in the last three years within the Basingstoke

laboratory.

Brandenberger et al. stated; “This comparison shows considerable deviation of the

results from the different methods… The KF method has been used by untrained

personnel, which could explain at least partially the difference between the results of

the KF and the new method” (instant coffee powders 0.5% absolute difference on a

water value of approximately 1.5%)73.

Containment and Safety

Karl Fischer analysis involves unavoidable exposure of the sample to the atmosphere

and possible change in moisture content during weighing and measurement. This was

a problem experienced with the dried samples. This means that the analyst is also

exposed to the product.


MSc Analytical Chemistry - N. Blakiston 108

The risk assessment for KF operation requires the use of a ventilated area; usually

work is conducted within a fume cupboard. The analyst must wear personal

protective equipment including nitrile gloves during any material or reagent

operation. The reagents used are particularly toxic and are flammable.

NIR analysis offers an extremely safe alternative, which requires no reagents and the

sample can remain sealed throughout.


MSc Analytical Chemistry - N. Blakiston 109

Conclusion
Quantitative measurement by Fourier Transform Near-Infrared (FT-NIR)

spectroscopy is reliant on data from a reference method. The Karl Fischer (KF)

volumetric titration method proved initially to give poor performance and results that

could net be relied upon.

During the feasibility study this method was further developed to produce results,

which met current method acceptance criteria in line with other KF methods. This

was not the original aim of the objectives stated and required significant investment of

investigative time.

Forty one samples were analysed in duplicate by the reference method and by FT-NIR

spectroscopy. The results from the reference method were used to develop three

calibration models using the NIR software package (Opus-Quant). The three models

included a linear regression model (R1), a Vector Normalized + Principle Component

Analysis (R2) model and a fully system optimised model (R3).

The calibration models produced values of 0.9736 (correlation coefficient), 96.31 (R2)

and 99.82 (R2) respectively. Models R2 and R3 had Root Mean Square Error of

Estimation values of 0.386 and 0.0952 respectively.

The validation models for R2 and R3 produced R2 of 94.42 and 95.74 with Root Mean

Square Error of Cross Validation values of 0.451 and 0.394.


MSc Analytical Chemistry - N. Blakiston 110

The predicted cross validation values gave a mean residual value of 0.000, -0.022 and

0.018 for each model with a maximum variance of 0.933, 1.237 and 0.909 % H2O

respectively. Models R2 and R3 identified one outlier each, this was not unexpected

and full use of the software package clearly show these data points lie significantly

outside of the normal population. However this was not fully investigated or

discussed as part of this feasibility study.

The NIR models gave acceptable prediction based on only 41 data points and were

comparable to the reference method with respect to both precision and accuracy.

There was no clear distinction in superiority between the three models as statistical

analysis demonstrated, but the linear regression model appeared to perform slightly

better than the other two models. Many NIR analysis methods reviewed were based

on complex multivariate chemometric manipulation. However this investigation has

proven that simple is often better and should certainly be the first approach to

establishing a linear relationship.

The NIR method gave rapid and non-destructive measurement with no sample pre-

treatment and eliminated all hazards associated with the reference method. The

savings of materials, equipment utilisation and resource are considerable over the

reference method once a fully validated NIR method is available.


MSc Analytical Chemistry - N. Blakiston 111

A Review of Method Validation


Analytical method validation has been defined as a procedure used to prove,

unequivocally that a test method should consistently yield what it is expected to do,

with adequate accuracy and precision. This is borne out by the statement in the

United States Pharmacopeia20: “The accuracy of an analytical method is the closeness

of test results obtained by that method to the true value”.

Regulatory authorities require that data provided for applications for marketing have

been acquired using validated methods. The Food and Drugs Administration

Guideline21 requires that data be provided”… over the range of interest (ca. 80-120%

of label claim).

In the highly regulated pharmaceutical industry there has been a history of increasing

method validation requirements. Prior to 1979 there was no mention of this

requirement in the guidance documents for Good Manufacturing Practice (GMP).

There are now numerous guidance notes21, 23 and regulated requirements to ensure a

method is ‘Fit for Purpose’.

The International Conference on Harmonisation of Technical Requiremements for

Registration of Pharmaceuticals for Human Use (ICH) has helped to rationalise some

of the key requirements across the industry. The guidelines introduced have helped

define the general validation characteristics required for analytical test methods;
MSc Analytical Chemistry - N. Blakiston 112

Specificity

Specificity is the ability to assess unequivocally the analyte in the presence of

components that may be expected to be present. Typically, these might include

impurities, degradants, matrix, etc. Lack of specificity of an individual analytical

procedure may be compensated by other supporting analytical procedure(s).

Linearity

The linearity of an analytical procedure is its ability (within a given range) to obtain

test results that are directly proportional to the concentration (amount) of analyte in

the sample.

Range

The range of an analytical procedure is the interval between the upper and lower

concentration (amounts) of analyte in the sample (including these concentrations) for

which it has been demonstrated that the analytical procedure has a suitable level of

precision, accuracy, and linearity.

Accuracy

The accuracy of an analytical procedure expresses the closeness of agreement

between the value which is accepted either as a conventional true value or an

accepted reference value and the value found. This is sometimes termed trueness.

Precision

The precision of an analytical procedure expresses the closeness of agreement

(degree of scatter) between a series of measurements obtained from multiple

sampling of the same homogeneous sample under the prescribed conditions.


MSc Analytical Chemistry - N. Blakiston 113

Precision may be considered at three levels: Repeatability, intermediate precision

and reproducibility. Precision should be investigated using homogeneous, authentic

samples. However, if it is not possible to obtain a homogeneous sample it may be

investigated using artificially prepared samples or a sample solution. The precision

of an analytical procedure is usually expressed as the variance, standard deviation,

or coefficient of variation of a series of measurements.

Repeatability

Repeatability expresses the precision under the same operating conditions over a

short interval of time. Repeatability is also termed intra-assay precision.

Intermediate Precision

Intermediate precision expresses within laboratories’ variations: Different days,

different analysts, different equipment, etc.

Reproducibility

Reproducibility expresses the precision between laboratories (collaborative studies,

usually applied to standardisation of methodology).

Detection Limit

ICH Definition: The detection limit of an individual analytical procedure is the lowest

amount of analyte in a sample which can be detected but not necessarily quantitated

as an exact value. This is only required for limit testing for impurities.
MSc Analytical Chemistry - N. Blakiston 114

Quantitation Limit

ICH Definition: The quantitation limit of an individual analytical procedure is the

lowest amount of analyte in a sample which can be quantitatively determined with

suitable precision and accuracy. It is used particularly for the determination of

impurities and/or degradation products.

Robustness/Ruggedness

The robustness of an analytical procedure is a measure of its capacity to remain

unaffected by small, but deliberate, variations in method parameters and provides an

indication of its reliability during normal usage.

System Suitability Testing

System suitability testing is an integral part of many analytical procedures. The tests

are based on the concept that the equipment, electronics, analytical operations and

samples to be analysed constitute an integral system that can be evaluated as such.

System suitability test parameters to be established for a particular procedure depend

on the type of procedure being validated.

The ICH guidelines were developed mainly for the validation of analytical procedures

primarily based on the analyte being in solution and using univariate mathematical

treatment of resultant data. This proved specific and suitable for the validation of

many routinely used chromatographic methods, but is more difficult in application to

NIR spectroscopic methods. To this end by 1990 there were a series of requirements

and guidelines published24 by various authorities that included:


MSc Analytical Chemistry - N. Blakiston 115

• Guideline for Submitting Samples and Analytical Data for Methods Validation,

1987

• Guidance Notes on Applications of Product Licences, 1987

• Committee for Proprietary Medicinal Products Guidance Note on Analytical

Validation, 1989

• Guidelines on the Quality, Safety and Efficacy of Medicinal Products for Human

Use, 1990

These guidelines attempted to define terms, but were vague as to what was exactly

required for a validation exercise. To help clarify the position several papers were

published on the practical approaches to method validation in pharmaceutical

analysis25 and this work was presented at the “Second International Symposium26 on

Pharmaceutical and Biomedical Analysis”. The Pharmaceutical Analytical Sciences

Group (PASG), which is an association of analytical chemists within the research

based UK pharmaceutical industry, agreed to determine current practice employed by

its membership. The results of a conducted survey, published in 199427 proved useful

for assessing the approaches to method validation and recommended a harmonised

guideline for validation within the pharmaceutical industry. This finally led to

revision of Pharmacopoeial text to allow for the use of new techniques of analysis;

The BP 1993 contains the following statement in the General Notices:

“this does not imply that performance of all tests in a monograph is necessarily a

prerequisite for a manufacturer in assessing compliance with the Pharmacopoeia

before release of a product. The manufacturer may assure himself that a product is

of Pharmacopoeial quality, from in-process controls or from a combination of both”.


MSc Analytical Chemistry - N. Blakiston 116

Further, the USP 23 stated in its General Notices that:

“ data derived from manufacturing process validation and from in process controls

may provide greater assurance that a batch meets a particular monograph

requirement than analytical data......On the basis of such assurances, the analytical

procedures in the monograph may be omitted.......Compliance may be determined

also by the use of alternative methods, chosen for advantages in accuracy, sensitivity,

precision, selectivity or adaptability to automation or computerised data reduction or

other special circumstances. Such alternative methods shall be validated”.

Official recognition of NIR spectroscopy was not achieved until 1990. The edition of

AOAC Official Methods of Analysis discussed a method titled “Piperazine in drugs”

(Official Methods of Analysis, 1990) was the first official NIR method in

pharmaceutical analysis. More recently a reflectance NIR assay analysis of intact

bulk tablets was conducted (Sterwin 500 mg paracetamol tablets), which

demonstrated successful application in meeting the regulatory submission

requirements for the pharmaceutical industry39, 40.

In the latter paper each validation characteristic was discussed in turn and proven to

be met through successful validation of the NIR method. It also discussed additional

routes to the analysis for validating more problematic spectral analysis.

Specificity can be achieved through selectively labelling all the components of

particular spectra, showing zero-order interference, or mathematical pre-treatment and

chemometrics could be applied to deconvolute the spectra. Demonstrating that spiked

interfering compounds or increases in others components do not interfere would be


MSc Analytical Chemistry - N. Blakiston 117

another suitable method. Proving a qualitative match to a reference library would also

be acceptable.

Linearity is demonstrated by the construction of calibration and sample sets. NIR

calibration plots use the predicted assay value Vs reference values and applies

statistical evaluation to demonstrate that the method is linear.

The Range of the linear calibration is dependant on the ease of using a suitable

calibration set. This is limited to the variation of manufacturing performance (tablet

weight, thickness or quantity of active present) unless representative samples can be

made. An ideal situation is one where the drug product is manufacture in the same

physical parameters with only a change in the API content (e.g. 25, 50, 100 mg).

Accuracy is conducted as part of the calibration and can be easily determined with

comparison to other known analysis methods. The Precision (degree of scatter) is the

variation between different readings of the same homogenous sample. Most NIR

methods suggest the averaging of multiple readings to increase the precision of the

true value as this only takes a few minutes per sample. The Intermediate Precision or

Repeatability can be achieved easily by measurements with different analysts on

different equipment (if available) on different days.

Inter-laboratory Precision or Reproducibility is not a general requirement unless the

method is for broad use, such as a Pharmacopoeial reference method. The final

requirement for validation is normally applied to solution stability, tolerance in

weighing practice, volumetric process and extraction times. This is the Robustness
MSc Analytical Chemistry - N. Blakiston 118

(the ability to remain unaffected by deliberate method variation) of the method and

due to the nature of sample preparation with NIR analysis, has little implication.

However other variances need to be reviewed such as; environmental conditions,

sample preparation (especially for ground or powdered samples, due to particle size

and bulk density variance), sample orientation and variation in the method of analysis

(probe parameters, vial diameters, cell thickness).

The daily operation of any analytical instrument is checked using a set calibration

process. This may be conducted as part on the analytical run using internal standards,

or using external standards such as in balance calibration. The System Suitability

check for a modern NIR spectrometer is an automated integral part of the system.

The instrument will conduct internal diagnostics and performances checks of the

filters using set parameters. This will take approximately twenty minutes at the start

of each day and will generate a report detailing the tests performed and the criteria

met.

As more companies invest and develop the use of the versatility and analysis speed of

this technique, greater guidance will become available from the regulators as to the

requirements for method validation. This may generate a snowball effect, but as yet

movement forward in this area has been slow.


MSc Analytical Chemistry - N. Blakiston 119

Reference Review
Applications of NIR Spectroscopy

The first examples of the application of NIR spectroscopy was in the examination of

cereal and forage products, specifically looking at moisture content and protein

analysis. This is still a significant area of application for the technique, but has now

spread into many other material handling sectors including industries such as;

petrochemicals, fine chemicals, polymers, food, packaging, medical and

pharmaceuticals. A good review of the development of NIR analysis has been

documented by Drennen et al, (1993) 74.

Some of these applications are discussed in more detail below, with the emphasis

mainly on applications in the pharmaceutical industry;

Qualitative Determinations

Examples of NIR spectroscopy used for qualification purposes in the pharmaceutical

industry include; identification of raw materials (API and Excipients), finished

product, assessing homogeneity and in-process monitoring of key stages of

pharmaceutical production such as blending and coating.

Identification
NIR reflectance spectroscopy is used in Basingstoke QCL to identify raw materials

and there have been well documented examples including; the identification and

discrimination of penicillin-type drugs (1982) 53 and more recently rapid, non-

destructive methods of analysis for the identification of pure powdered drugs and

actives in tablets33.
MSc Analytical Chemistry - N. Blakiston 120

A published application by Dempster (1993)34, 35, described the development of non-

invasive methods to confirm the identity of blister-packed tablets for clinical trial

supplies. The approach described the examination of the tablets through unopened

opaque blister packs and presentation of single exposed tablets. Four strengths of the

experimental drug were examined (2, 5, 10 and 20% w/w of the active), a placebo and

a marketed comparator containing 80% w/w of the active. The method was also able

to distinguish film-coated and non-film coated tablets.

These papers illustrate the diverse applications of NIR spectroscopy for identification

purposes.

In-Process Monitoring

This is a great strength for NIR spectroscopy analysis due to its non-destructive nature

and rapid analysis time. Investigation by Rantanen (1998)50 to monitor the moisture

content in a fluidized bed granulator with a multi-channel NIR moisture sensor proved

useful for validation purposes. The Standard Error of Performance (SEP) was found

to be 0.2%. There is a movement within the industry, supported by the FDA, to

incorporate Process Analytical Technologies (PAT). This on-line sampling and

testing builds additional quality into the validation of a process, which ensures greater

confidence in the product and can lead to reduced end item monitoring66.
MSc Analytical Chemistry - N. Blakiston 121

Homogeneity

The following are examples of the use of NIR spectroscopy to determine

homogeneity. The homogeneity of a typical direct compression pharmaceutical

powder blend was monitored by reflectance NIR spectroscopy by Wargo (1996) 54.

They concluded that NIR spectroscopy had great potential as an analytical tool in

powder blend analysis.

Another paper published on the on-line monitoring of powder blend homogeneity by

Sekulic (1996) 57 described the use of a diffuse reflectance fibre-optic probe during the

blending process. Spectral changes occurred which eventually converged to a point

of constant variance when a blend of sodium benzoate, microcrystalline cellulose,

lactose and magnesium stearate were suitably blended. Their results demonstrated

that a blend was suitably homogeneous before a typical blending period was

complete.

These papers illustrate how, by the use of NIR spectroscopy, the blending process

need not be a time limiting factor in the pharmaceutical industry due to, for example,

waiting for laboratory analysis to confirm batch homogeneity. Furthermore, NIR

spectroscopy can more accurately predict, through greater sampling, the optimum

blending time for homogeneity.


MSc Analytical Chemistry - N. Blakiston 122

Quantitative Determinations

Particle Size

The particle size distribution of powdered pharmaceutics is important as it affects

powder flow, dissolution rate and compressibility. Typical particle size analysis is by

forward angle light scattering or sieve analysis. However this is a time consuming

and sample destructive technique.

A study was published on the use of reflectance NIR spectroscopy to the

measurement of granule particle size growth of microcrystalline cellulose during

granulation by Rantanen (1998) 50. The results demonstrated the use of NIR

spectroscopy for on-line determination of the particle size.

This was followed by a publication on the determination of cumulative particle size

distribution of microcrystalline cellulose using MLR and PLSR by O’Neil (1999) 55.

This study demonstrated a rapid determination of the cumulative frequency which

represented a development over previously published studies by Ciurczac (1986) 56

which proved a linear relationship between the ratio of two wavelengths versus mean

particle size.

NIR spectroscopy therefore has the potential to be used in the pharmaceutical industry

for determining both the mean particle size and cumulative particle size of powders

and granules.
MSc Analytical Chemistry - N. Blakiston 123

Powders and Tablets

There are numerous examples of NIR spectroscopy being used to non-destructively

quantify the active in powders and tablets (crushed or whole).

As early as 1987, reflectance NIR spectroscopy was used to determine the

nicotinamide content in pre-mixes (Osborne 198758). This enabled 36 samples to be

quantified in around 30 minutes.

Later Corti (199059) published a paper on reflectance NIR spectroscopy being used to

quantify the following solid binary mixtures:

• diazepam (1.5% w/w,.SEP:.0.11%.w/w)

• otillonium bromide (14.8%.w/w,.SEP:.0.37%.w/w)

• dicloxacillin (34.0%.w/w,.SEP:.0.52%.w/w)

• sodium ampicillin (66.0%.w/w,.SEP:.0.67%.w/w)

The authors concluded that the results were satisfactory if the percentage active

content was not below 1% w/w.

Corti et al, continued work in this field and described how streptomycin sulphate

(70.2% w/w) and cloxacillin benzathine (96.0% w/w) powders were quantified using

reflectance NIR spectroscopy. The SEP for these quantitative determinations was 1.03

and 0.95% w/w respectively. Several antibiotic compounds were also reviewed,

sodium ampicillin (85.0% w/w, SEP:.1.08%.w/w) and gentamicin sulphate

(66.0%.w/w,.SEP:.0.78%.w/w).
MSc Analytical Chemistry - N. Blakiston 124

Zappala (1977) 60, measured meprobamate in ground tablets by measuring the

transmittance of NIR radiation. The results compared favourably with the reference

United States Pharmacopeia method with a coefficient of variation (CV) of 0.7%.

Broad (2000) 61, published a paper on the application of transmittance NIR

spectroscopy to the uniformity of content of intact steroid tablets. Tablets containing

5, 10, 15, 20 and 30 mg (2.94, 5.88, 8.82, 11.76 and 17.64% w/w steroid) were

quantified with a measurement of error for a single tablet having a relative standard

error of less than 2.5% for a given active level.

A guideline has also been published, Smith et al., (2002)61, for the transfer of a

reflectance NIR reflectance assay for paracetamol in intact tablets between two

instruments of the same type. The authors successfully demonstrated the transfer of

PLS regression and MLR model assays along with the criteria deemed necessary to

conclusively prove transfer and justify any correction method utilised.

These papers illustrate that NIR spectroscopy can be used to quantify powders and

tablets, whether pharmaceutical premixes, antibiotics or binary mixtures and therefore

can be used as a replacement for the reference techniques such as HPLC.


MSc Analytical Chemistry - N. Blakiston 125

Brief History of Water Analysis by NIRS


Typical pharmaceutical methods for water determination are based on weight loss by

drying or by Karl Fischer titration techniques16.

NIR absorption bands in the region of 1900 to 2000 nm are indicative of the presence

of water as these are due to the combination of fundamental bending and stretching

vibrations of the OH bond. The intensity of this absorption band and the shape

depends on the degree of hydrogen bonding in environment in which the water is

located. The stronger the hydrogen bonding, the longer the wavelength of the NIR

absorption observed.

The determination of water is one of the more extensively published applications of

NIR spectroscopy and quantitative water determinations in the order published and

the major findings are detailed below:

Keyworth52 (1961), detailed the use of transmittance NIR spectroscopy to determine

the water content of methanol and ethylene glycol by single wavelength linear

regression. The content of water assayed was 0.03 to 4.00% v/v for methanol and

2.2 to 5.0% v/v for ethylene glycol. He concluded that the NIR method compared

favourably with the reference method and had the advantage of being non-destructive.

Hollenberg45 (1982), utilised transmittance NIR spectroscopy to determine hydration

numbers in 16 soluble anhydrous amino acids by single wavelength linear regression.


MSc Analytical Chemistry - N. Blakiston 126

The results demonstrated that the method was precise and comparable to available

literature values for both monomer and homopolypeptide hydration numbers.

Williams46 (1983), used reflectance NIR spectroscopy to determine the moisture

content of wheat by Multiple Linear Regression (MLR) using up to six wavelengths.

The NIR instruments were found to be superior to the reference oven method for the

determination of moisture in most cases.

Last47 (1993), used reflectance NIR spectroscopy for the determination of the

moisture in an experimental freeze-dried injection product by single wavelength, dual

wavelength linear regression and Partial Least Squares Regression (PLSR). This

study emphasised the need to include all possible manufacturing process variables in

order to obtain robust calibrations.

Blanco48 (1997), used reflectance NIR spectroscopy and a fibre-optic probe to

determine the water content in ferrous lactate. The results departed from the reference

Karl Fischer titration method by less than 1.5%.

Zhou49 (1998), determined the moisture levels in hygroscopic drug substances by

diffuse reflectance NIR spectroscopy. Samples were prepared with moisture levels

from 0.5 to 11.4% w/w and calibration models were built using PLSR giving a

Standard Error of Prediction (SEP) of 0.11% w/w.

Webster (2003) 70, developed and qualified a quantitative method for the

determination of moisture in topical formulations. This was developed for in process


MSc Analytical Chemistry - N. Blakiston 127

monitoring and as a check on finished goods. This method was suitable for routine

release of moisture, but further development was required to support active purity and

stability indication.

Brűlls (2007) 70, looked at OH combination bands and overtones (OH-stretch) using

spectral peak area analysis accurate predictions were made for moisture content of

lyophilized PVP in the range of 0-22%.

The above publications illustrate the wide spread application of NIR spectroscopy to

the determination of water content in both solid and liquid samples with examples

dating as far back as 1961.


MSc Analytical Chemistry - N. Blakiston 128

Future Work
Further development of an NIR method for Prozac determination would not be a

viable project, as this investigation was only a feasibility study and has no real benefit

to Eli Lilly. However as a student project this work could be continued to prove that

full validation of a NIR method for water determination is possible.

NIR analysis as a replacement technique for moisture determination has demonstrated

that a great opportunity exists to reduce lead times, costs, increase analyst safety, be

non-destructive and reduce potential errors.

With this in mind a suitable product candidate for development would be one that is

of high volume, has water content well below regulatory limits and has a high toxicity

level. NIR spectroscopy has been proven successful in the analysis of whole coated

tablets and thus testing could easily be conducted on-line in the manufacturing area.

One product manufactured at Basingstoke has a small variability in water content

between manufactured sections. If tested as a composite sample, the replicate KF

determinations would fail the method reproducibility limit. As batches are produced

as a four or even eight section size the time savings that NIR analysis offer are vast.
MSc Analytical Chemistry - N. Blakiston 129

Reference to Raw Data

Bench Books

BAS 1318 page 27, 28


BAS 1420 page 116, 117, 118
BAS 1441 page 41, 43, 44
BAS 1447 page 64, 65
BAS 1448 page 132
BAS 1607 page 16 – 37
BAS 1609 page 17
BAS 1613 page 13, 15, 16, 17

Analytical Data Wallet

ADW: 32352
ADW: 32365
MSc Analytical Chemistry - N. Blakiston 130

References
1. Bunsen, R., Annalen, 1835, 86, 265
2. Smith, D. M., and Bryant, W. M. D., J. Amer. Chem. Soc., 1935, 57, 841
3. European Pharmacopoeia, 1969, Vol. 1, 165
4. British Patent No. 722,873; U.S. Patent No. 2,780,601
5. British Patent No. 728,974
6. Johannson, A., Acta Chem. Scand., 1949, 3, 1058
7. Peters, E. D., and Jungnickel, J. L., Analyt. Chem., 1955, 27, 450
8. Klinova, V. A., Sherman, F. B., and L’vov, A. M., Analyt. Abstr., 1969, 16,
1893
9. Gonzalez Garcia-Gutierrez, A., Analyt. Abstr., 1970, 18, 2572
10. BDH, Moisture determination by Karl Fischer, 2nd ed., 1967, 13
11. Riedel-de-Haën., Hydranal® - Water Reagent According to Eugen Scholz for
Karl Fischer Titration., 3d ed., 1982
12. Scholz E., “Karl Fischer Reagents without Pyridine”, Fresenius Z. Anal.
Chem., 303, 203-207, (1980)
13. Scholz E., “Karl Fischer Titration of Aldehydes and Ketones”, Anal. Chem.,
57, 2965-2971, (1985)
14. Riedel-de-Haën., Hydranal® - Manual, Eugen Scholz reagents for Karl
Fischer Titration., Sigma Aldrich, 2001
15. Loss on Drying - European Pharmacopoeia, Fifth ed., Volume 1, 1. General
Notices, page 6, 5.0.
16. Turbo2TM Instruction Manual, “Blending Karl Fischer Titrator”, Orion, 1991,
215662-001.
17. Anon (2001), Hydranal®-Manual – “Eugen Scholz Reagents for Karl Fischer
Titration”, Germany.
18. “Vogels - Quantitative Inorganic Analysis”, XVII, 18, page 688, (edition not
known - photocopied page for reference).
19. Wikipedia Reference; http://en.wikipedia.org/wiki/Electromagnetic_spectrum
20. United States Pharmacopeia XXII, section 1225, 1710-1712.
21. Guidelines for Submitting Samples and Analytical Data for Methods
Validation, US Department of Health and Human Services, Fod and Drugs
Administration, Maryland, USA, Feb. (1987).
22. Code of Federal Regulations, US. 21CFR211.165, 28 March 1979.
23. Code of FederalRegulations, US. 21CFR211.194, 28 March 1979.
24. World Health Organisation. Validation of Analytical Procedures used in the
examination of pharmaceutical materials. WHO Technical Report Series 823,
1992.
25. Mark, H., Ritchiw, G.E., Roller, R.W., Ciurczak, E.W., Tso, C., MacDonald,
S.A., “Validation of NIR transmission spectroscopic procedure, part A:
validation protocols”, Journal of Pharmaceutical and Biomedical Analysis,
2002, 28, 251-260
26. Carr, G.P., Wahlich, J.C., “A practical approach to method validation in
pharmaceutical analysis”, Journal of Pharmaceutical & Biomedical
Analysis”, 1990, 8, 613-618.
27. Clarke, G.S., “The validation of analytical methods for drug substances and
drug products in UK pharmaceutical laboratories”, Journal of Pharmaceutical
and Biomedical Analysis, 1994, 12(5), 643-652
MSc Analytical Chemistry - N. Blakiston 131

28. “Molecule of the Month for February 1996”, Update by Karl Harrison,
http://www.3dchem.com/moremolecules.asp?ID=30&othername=Fluoxetine
29. “Image:Fluoxetine structure.svg”, Wikipedia, the free encyclopedia, Self-made
in BKChem, 16th March 2007,
http://en.wikipedia.org/wiki/Image:Fluoxetine_structure.svg
30. Methods of Moisture Content Determination, Halogen Moisture Analyser
from Mettler Toledo, Application Brochure, 02/2002
31. Mettler Toledo, “Fundamentals of the Karl Fischer Titration”,
32. Mettler Toledo DL31/DL38 Titrators, “Karl Fischer Application Brochure
26”, 10/98, 03/99.
33. Khan, P.R., Jee, R.D., Watt, R.A., Moffat, A.C., The identification of active
drugs in tablets using NIRS, Pharmaceutical Sciences, 1997, 3(9), 447-453
34. Dempster, M.A., Jones, J.A., Last, I.R., MacDonald, B., Prebble, K.A., “NIR
methods for the identification of tablets in clinical trial supplies”, Journal of
Pharmaceutical & Biomedical Analysis, 1993, 11(11/12), 1087-1092.
35. Dempster, M.A., MacDonald, B.F., Gemperline,P.J., Boyer, N.R., “A NIR
reflectance analysis method for the non-invasive identification of film-coated
and non-film-coated, blister-packed tablets”, Analytica Chimica Acta, 1995,
310, 43-51.
36. Burns D., A., Cuirczak, E., W., “Handbook of Near-Infrared Analysis”,
Marcel Dekker, New York, 2001.
37. Workman J., “An Introduction to Near Infrared Spectroscopy”, Spectroscopy
Now, 12 September 2005, Wiley.
38. D.A.Skoog, T.A. Nieman, F.J. Holler, “Principles of Instrumental Analysis”,
5th ed., W Brookes/Cole, ISBN 0-03-002078-6.
39. Trafford, A.D., Jee, R.D., Moffat, A.C., Graham, P., “A rapid quantitative
assay of intact paracetnol tablets by reflectance near-infrared spectroscopy”,
Analyst, 1999, 124, 163.
40. Moffat, A.C., Trafford, A.D., Jee R.D., and Graham, P., “Meeting the ICH
guidelines on validation of analytical procedures”, Analyst, 2000, 125, 1341.
41. U. S. Food and Drug Administration., “21 Code of Federal Regulations
Electronic Records; Electronic Signatures”, http://www.fda.gov/.
42. Tobias R.T., “An introduction to PLSR”, SAS Inst. Inc., Cary., NC.
43. Morisseau, K.M., Rhodes, C.T., “Drug Development and Industrial
Pharmacy”, 1995, 21, 1071.
44. Keyworth, D. A., Talanta, 1961, 8, 461.
45. Hollenberg, J.L., Ifft, J.B., J. Phys. Chem, 1982, 86, 1938.
46. Williams, P.C., Norris, K.H., Gehrke, C.W., Berstein, K., Cereal Foods
World, 1983, 28, 149.
47. Last, I.R., Prebble, K.A., J. Pharm. Biomed. Anal., 1993, 11, 1071.
48. Blanco, M., Coello, J., Iturriaga, H., Maspoch, S., de la Pezuela, C., Appl.
Spectrosc., 1997, 51, 240.
49. Zhou, X., Hines, P., and Borer, M. W., J. Pharm. Biomed. Anal., 1998, 17,
219.
50. Rantanen, J., Lehtola, S., Ramet, P., Mannermaa, J., Yliruusi, J., “On-line
monitoring of moisture content in an instrument fluidized bed granulator with
a multi-channel NIR moisture sensor”, Powder Technology, 1998, 99, 163-
170.
MSc Analytical Chemistry - N. Blakiston 132

51. MacDonald, B.F., Prebble, K.A., “Some applications of NIR reflectance


analysis in the pharmaceutical industry”, Journal of Pharmaceutical &
Biomedical Analysis, 1993, 11(11/12), 1077-1085.
52. Keyworth, D. A., “Talanta”, 1961, 8, 461.
53. Rose, J. R., “Qualitative and Quantitative Aanalysis of Pharmaceuticals with
NIRA”, Proceedings of the 2nd Annual Symposium on NIRA, Technicaon,
Tarrytown, NY, 1982.
54. Wargo, D. J., Drennen, J. K., J. Pharm. Biomed. Anal., 1996, 14, 1415.
55. O’Neil, A. J., Jee, R. D., Moffat, A. C., Analyst, 1999, 124, 33.
56. Ciurczak, E., Torlini, P., Demkowicz, P., Spectroscopy, 1986, 1, 36.
57. Sekulic, S. S., Ward, H. W., Brannegan, D. R., Stanley, E. D., Evans, C. L.,
Sciavolino, S. T., Hailey, P. A., Aldridge, P. K., Anal. Chem., 1996, 68, 509.
58. Osborne, B. G., Analyst, 1987, 112, 313.
59. Corti, P., Dreassi, E., Corbini, G., Ballerini, R., Gravina, S., Pharm. Acta.
Helv., 1990, 65, 189.
60. Zappala, A. F., Post, A., J. Pharm. Sci., 1977, 66, 292.
61. Broad, N. W., Jee, R. D., Moffat, A. C., Smith, M. R., Analyst, 2000, 125,
2054.
62. Smith, M. R., Jee, R. D., Moffat, A. C., Analyst, 2002, 127, 1682.
63. Franklin, E. Barton II, “Theory and Principles of NIRS”, Spectroscopy
Europe, 14/1, (2002), p12.
64. Miller, C. E., “Sources of non-linearity in NIR methods”, NIR News, 1993,
4(6), 3-5.
65. Morisseau, K.M., Rhodes, C.T., “Pharmaceutical uses of NIRS”, Drug
Development and Industrial Pharmacy, 1995, 21(9), 1071-1090.
66. Forcinio, H., “Pharmaceutical industry embraces NIR technology”, 2003,
18(9), 17-19.
67. Bruker Opus Version 5.5 Quant Spectroscopic Software, “Handbook”, 2004.
68. Fearn, T., “A look at some standard pre-treatments for spectra”, NIR news,
1999, 10(3), 10-11.
69. Cowe, I. A., McNicol, J. W., Appl. Spectrosc., 1985, 39, 257.
70. Webster, G. K., Farrand, D. A., Johnson, E., Litchman, M. A., Broad, N.,
Maris, S., “Use of NIRS for quantitative determination of elamectin and
moisture in topical formulations”, J. Pharmaceuticals and Biomedical Anal.,
33(1), 2003, 21-32.
71. Brűlls, M., Folestad, S., Sparén, A., Rasmuson, A., Salomonsson, J.,
“Applying spectral peak area analysis in NIR spectroscopy moisture assays”,
Journal of Pharmaceutical and Biomedical Anal., 44(1), 2007, 127-136.
72. Plugge, W., Van Der Vlies, C., “The use of NIR spectroscopy in the quality
control laboratory of the pharmaceutical industry”, Journal of Pharmaceutical
& Biomedical Analysis, 10(10-12), 797-803, 1992.
73. Brandenberger, H., Bader, H., “Determination of. Powder Moisture by
Azeotropic Distillation and Near-Infrared Spectrophotometry”, anal. Chem.,
33(13) 1961, 1947-1949.
74. Drennen, J. K., Lodder, R. A., “Pharmaceutical Applications of NIRS”,
Advanves in NIR measurement, 1, 93-112, (1993).
75. Brutsche, A., “Advances in NIR Spectroscopy”, European Pharmaceutical
Review, 1996, Sep, 45-51.
76. Windham, W. R., Barton, F.E., “Moisture in Forage NIR Reflectance
Spectroscopy”, J. ASSOC. OFF. ANAL. CHEM., 74(2), 1991, 325-331.
MSc Analytical Chemistry - N. Blakiston 133

Attachment 1: Tabulated Overview moisture determination methods 30.


MSc Analytical Chemistry - N. Blakiston 134

Attachment 2: Orion Turbo2TM Volumetric Karl Fisher Titrator.

1. Blender drive cover


2. Pump observation window
3. Tube platen
4. Drain pump switch
5. Platen release/engage lever
6. Pump rotor
7. Glass sealing plate
8. Pivot release catch
9. Vessel clamp screws
10. Protective guard
11. Spillage tray
12. Vessel
13. Electrode
14. Sample port stopper
15. Cam arm
16. Protective guard retaining screw
17. Blender motor inter-lock
position (magnet)
18. ‘P’ clip jet tube retainer
19. Blender drive belt
20. Shaft seal retainer

Orion Turbo2TM Volumetric Karl Fisher Titrator - Rear View

8. Waste bottle vacuum tube


10. Desiccant guard tubes, vessel
11. Desiccant guard tube, reagent
12. Interconnecting power cable sockets
13. Fuse, power cable
14. Reagent tube connector
15. Socket for console interconnecting
signal cable

(Turbo2TM Instruction Manual)


MSc Analytical Chemistry - N. Blakiston 135

Attachment 3: Bruker PQ Test Protocol


MSc Analytical Chemistry - N. Blakiston 136

Attachment 4: Pay-Off Matrix for Water Determination A

Speed of Analysis versus Ease of Use


Speed of analysis

NIR
DE/RE

Halogen

Ease
of
KF
use

Oven

TGA
MSc Analytical Chemistry - N. Blakiston 137

Attachment 5: Pay-Off Matrix for Water Determination B

Wide Sample Range versus Measuring


Range
Wide Sample Range

NIR
KF

DE/RE
TGA
Oven
Measuring
Halogen Range
MSc Analytical Chemistry - N. Blakiston 138

Attachment 6: Pay-Off Matrix for Water Determination C

Accuracy versus Selectivity


Accuracy

NIR
KF
TGA DE/RE

Halogen Oven Selectivity


MSc Analytical Chemistry - N. Blakiston 139

Attachment 7: Bruker MPA FT-NIR Default Parameter Settings.


MSc Analytical Chemistry - N. Blakiston 140

Attachment 8: Equipment

Prozac Powder blend Sample containing


capsule fragments.

Turbo 2 Karl Fisher: Peristaltic pump, delivery tubes,


motor drive, sample aperture and Turbo blade.

Turbo Karl Fischer set up with;


Balance, Reagents, Printers, Calibration
weights and syringe.

Bruker MPA FT-NIR

This can measure in a multiple of


orientations including fibre optic
probe.
MSc Analytical Chemistry - N. Blakiston 141

Attachment 9: Method AP-1043-01

Turbo2 Karl Fischer Assay for Water


In Fluoxetine Hydrochloride Capsules
Equivalent to 10 and 20 mg Fluoxetine.
MSc Analytical Chemistry - N. Blakiston 142

Attachment 10: Technical Protocols

QCL-TR-014

QCL-TR-015

QCL-TR-016

Вам также может понравиться