Chapter 1 Introduction

Alzheimers Disease (AD) is one type of age-related neurodegenerative disorders which

slowly destroys memory and brain functions and eventually leads to loss of
independence and total disability(Batsch et al., 2012). In the USA, the prevalence of AD
was estimated to be 5.3 million in 2015 and, by 2050, is predicted to be over 14 million
(Batsch et al., 2012). Also, AD is one of the costliest chronic diseases to USA society. In
2015, the direct costs to USA society of caring for AD patients was 226 billion dollars,
with half of the costs borne by Medicare (alz.org, 2015).
However, currently, there isnt an effective treatment to slow or reverse the progression
of AD. Considering the problems this disease poses in terms of its prevalence and
financial cost, there is a need of an accurate statistical model to assess the progression
of AD since an AD progression model can help the development of potential treatments.
Besides, monitoring the pathological progression of AD with a proper statistical model
will help the pathological development of AD. More importantly, the possibility of
predicting the latent states changes will eventually help to improve prognosis for AD
patients in early stage which will provide a critical opportunity for therapeutic
intervention. This indicates that the AD progression model is supposed to be able to
detect more hidden states than the clinical diagnoses and thereby advance our
understanding of AD. In traditional clinical diagnosis, the progression of AD is in three
irreversible phases: Normal (pre-symptomatic phase) stage in which individuals are
clinically normal but have AD pathology, Mild Cognitive Impairment (prodromal phase)
stage in which cognition is impaired but the behavior functioning is minimally impaired,
and AD (symptomatic phase) stage when the clinical symptoms of AD such as loos of

function in activities appear (alz.org, 2015). These traditional three stages has been
used extensively in the clinical field. However, more evidence and analysis from both
genetic and clinical researches suggest the patho-physiological process of AD begins
before the diagnosis of AD even though the pathogenesis of AD remains unknown,
particularly in the early stages of the disease where therapies presumably would be
most effective. Many epidemiological studies suggest that the pathological process of
AD begins years before a clinical diagnosis is made (Li. C, 2015). In 2016, EJ. Mufson
and others have demonstrated that A plaques and NFTs are present in varying
degrees before the onset and throughout the progression of dementia. (EJ. Mufson,
2016) In the same year, another study suggests that BACE1 may serve as a sensitive
biomarker of early neuropathological changes before the diagnosis of AD (Lan Li, 2016).
All of this studies provide evidence that some of biomarkers have appeared
abnormalities in pre-symptomatic stage of AD progression which indicates that it is
critical to study the progression patterns in pre-symptomatic or prodromal phase
(Sperling et al., 2011). Note that the pathways of transitions from pre-symptomatic to
prodromal phase or from prodromal to clinical symptomatic phase are unclear, there is a
need to provide a more systematic and comprehensive assessment of the multiple
changes with the AD biomakers that occur during the transition from normal to MCI to
frank AD.

In recent years, numerous AD progression modelling studies aimed to assess the

progression of AD especially in the presymptomatic stage have been done. In 2010, a
time-to-event (TTE) simulation provided estimates of screening has been proposed for

Early-stage modeling of AD progression. In the same year, Rachelle S Doody used a

mixed regression effects analysis on a set of clinical measurements to assess and
predict disease progression (Doody et al., 2010). Later, an event-based disease
progression model has been applied in AD (Fonteijn HM et al., 2011) in 2011 and it
described the progression of AD as a series of events with a significant change in
symptoms or in tissue. In the next year, Furiak NM suggested a simulation method for
the population-level impact of screening for preclinical AD (Furiak NM, 2012). However,
most of these event-based AD progression modellings were still in accordance with
traditional clinical progressive stages and the latent states of pathological process for
AD still havent been detected. Disease states is a general term that refers to any
quantifiable variable describing disease at a particular point in time (Holford, 2014),
which is valuable not only for potential drugs development but also for identifying criteria
to distinguish between types of treatment efficiency, with implications for long-term
patient care. Therefore, it is vital to detect the latent disease states with sensitive
presymptomatic biomarkers and the transition probability of AD progression, which
indicates an unsupervised learning method might be a proper way to describe the AD
progression with latent states from the changes of biomarkers. In 2012, Escudero J
suggested using clustering analysis to detect the pattern of AD progression (Escudero
J, 2012) and in the same year, Iordanescu G presented a novel automatic plaque
segmentation algorithm based on the intrinsic MR signal characteristics of plaques to
classify plaque or nonplaque by an unsupervised learning method (Lordanescu G,
2012). Later, Madsen SK used a novel unsupervised machine learning approach called
CorEx to evaluate cognitive decline and construct latent factors that explain the

correlation of biomarkers (Madsen SK, 2015). However, the transition probability and
the AD progression pattern profile with the latent states are still unclear. Understanding
the dynamic disease process is critical in early detection and diagnosis. Therefore, a
framework for describing AD progression pattern with latent states is needed.

Ideally, a framework for describing AD progression is supposed to detect unobserved

AD progression states with observed response items such as the change of sensitive
biomarkers to predict outcome events. Also, methods of time-to-events analysis are
needed especially in long-term progressive disease such as AD to predict the
pathological progression over time. Latent transition models, which assumes the
unobserved (latent) categorical variables (classes) can be explained with the observed
measurements, would be applied to detecting the progression of AD over time. In 2012
(Sukkar et al., 2012), Rafid Sukkar proposed a hidden markov model to monitor the
progression of AD with observed biomarkers and compare the hidden stage with stages
corresponding to clinical diagnoses. However, the approach to determine the hidden
stages and the profile of AD progression are still unclear. In addition, sufficient evidence
suggest that some patient factors, such as age, sex and apolipoproteinE4 (ApoE4)
expression, have been shown to account for some of the variability in progression
(Holford, 2013), and this may be useful for modeling AD progression.

In this study, the latent transition model for AD progression is employed in where the AD
progression is modeled with a set of bio-markers. Different bio-markers for AD may

potentially be complementary in diagnosis of AD and many bio-markers have proved to

be very sensitive to AD and MCI.(Gomar et al., 2011) The sensitive bio-markers were
selected and employed with the latent transition model. The ideal bio-markers should be
able to detect specific pathological features of AD neuropathology or have high
sensitivity for detecting AD progression. (Biagioni et al., 2011) Overall, AD bio-markers
can be categorized into following groups: pathological, structural or functional, and
clinical. The pathological group of bio-markers (CSF amyloid 42, tau) indicating
pathology appears in the presymptomatic phase of AD. The structural or functional
group reflects neuronal structure and functions which appears in the late
presymptomatic or early prodromal phase. Finally, the clinical group (behavior, cognition
and function) suggests the dementia with impairment and loss of function in activities
which usually appears in the prodromal phase of AD.(Biagioni et al., 2011) In AD
progression modeling, the convergence of multi-states latent transition models for ADNI
data might be adversely affected if too many biomarkers are employed. Considering this
concerns and the completeness of data, we selected volume of hippocampus, FDG and
MMSE for AD progression modeling. The latent transition model was proposed to
monitor the progression of AD with sensitive biomarkers which can help the
development of a given proposed treatment. Interestingly, the results suggested that the
AD progression is not simply from low to high disease states as traditional clinical
progression suggested, but a more complicated system which illustrated the pattern of
two profiles of the AD progression. One of them was primarily determined by FDGPET
and the other by the volume of hippocampus. And this latent transition model provided
more details about the pattern of intermediary transitions between states.

Chapter2 Latent transition models

Modeling the progression of Alzheimer's Disease plays an important role in the
development of potential treatments. Latent transition models, as one of the analysis of
stage-sequential processes, can be used to assess the progression of Alzheimer's
Disease. Since the assumption of latent transition models is that the latent states at
each time point are unobserved, but can be measured with a set of manifest items, this
statistical method might be ideal to monitor the progression of AD only with observed
bio-markers. In addition, the latent transition models are the extension of latent class
models, which assess that the latent states through their responses to manifest items.
Specifically, the manifest items in this study represent the sensitive AD biomarkers. In a
latent transition model, the measurement in each time point is a latent class model and
the transition probability of latent class membership over each time point can be
estimated. (Chung et al., 2005)
\\
In general, the following sets of parameters are estimated from latent transition models:
i) : latent states membership probabilities at Time 1; ii) : probabilities of transitions
between latent statuses over time; iii) : item-response probabilities conditional on latent
status membership and time. In addition, the Rho parameters interpret the
correspondence between the observed items and the latent and when covariates are
includes, two additional sets of

parameters can be estimated: i) a set of parameters

which are logistic regression coefficients for covariates predicting latent status

membership at baseline time point states; ii) a set of parameters which are logistic
regression coefficients for covariates predicting transitions over time. Here, assuming a
latent transition model with n estimated latent states can be estimated from C
categorical response measured at

times, the individual is responses for all the times

t=1, 2, T and response items r=1,2, , R is:

Yi i i i M i i i M iT iT iTM = ( , , , , , , , , , , , , )
Set

as individual is latent state membership at baseline time point,

as individual

is latent state membership at time 1, and so on; and I() as the indicator function which
equals 1 if y equals k and 0 otherwise; and as individual Is group membership, the
latent transition model can be expressed as:

Here, Xi is the value of the covariate X for individual i.

In this model, the probability of belonging to latent status s at baseline time point is
given by the parameter | 1 ( ) = ( = | = , = ) s g i i i x P S s X x G g. The
correspondence between the parameters and covariates can be interpreted as
category multinomial logistic model (see, e.g., Agresti, 2002). So, with one covariate X,
the parameters can be expressed in this way:

For

Similarly, | , 2 2 1 1 2 1 ( ) = ( = | = , = , = ) s s g i i i i x P S s S s X x G g is also expressed by

a multinomial logistic model which estimates the probability of individual i's from latent state s1
to latent state s2. For example, the probability of individual i transitioning from latent status s1 at
Time 1 to latent status s2 at Time 2 given membership in group g and covariate value x is
0 | , 1 | , 2 1 2 1 | , 1 2 1 0 | , 1 | , 2 1 2 1 =1 exp ( ) = 1 exp s s g s s g s s g n s s s j s s j j x x x
(7) for s2 = 1, , ns.
(Here latent status ns is serving as the reference status.)

In this study, since we have only two time points (baseline time point and after two
years), the general latent transition model can be specified in this way:

Here, the parameters

Chapter 3 ADNI data

The proposed latent transition model was performed with real data from Alzheimers
Disease Neuroimaging Initiative (ADNI) to seek an adequate explanation of the
progression of MCI and early AD by using various clinical and neurocognitive
measurements. The ADNI database was launched in 2003 as a public-private
partnership, which is led by the Principal Investigator Michael W. Weiner, MD.(Liu et
al., 2016) At first, the main goal of ADNI is to investigate whether serial magnetic
resonance imaging (MRI), positron emission tomography (PET), other biological
markers, and clinical and neuro-psychological assessments can be used to measure

the progression of mild cognitive impairment (MCI) and early Alzheimers disease
(AD) (Liu et al., 2016). Recently, the ADNI database has been used in various
studies about Alzheimers disease. Data of participants in three traditional clinical
groups, which are NC (normal control), MCI (Mild cognitive Impairment) and AD
(Alzheimers disease), are available in ADNI database. The longitudinal data
performed in the latent transition model contain 1,879 participants from 2005 to
2014 who were examined at baseline time and with repeat visits every 6 months for
a period of up to 108 months.
In addition, there are three stages of ADNI including ADNI 1, ADNI GO and ADNI 2
[18]. As the earliest stage, ADNI1 was aimed to find out more accurate diagnose
method for detecting AD in the earliest or pre-diagnosis stage and define the
pathology using biological markers. The various biological biomarkers in ADNI
include existed brain-imaging techniques, such as PET (Positron Emission
Tomography) and MRI (Magnetic Resonance Imaging) for detecting early stage of AD
as well as a series tests including PET scans, lumbar punctures for CSF, blood tests,
neuropsychological tests and tests of psychiatric function for mood [17]. Then, the
data of these biomarkers are collected and analyzed for the progression of cognitive
decline as well as assisting researchers to develop the potential treatments. Later,
the second stage of ADNI GO was launched from June 2009 to 2011. Based on ADNI
1, ADNI GO aimed at developing standardized methods for clinical tests according to the previous
ADNI studies in order to have a better understanding of cognitive decline in the early stage of AD.
Then, ADNI2 as the third stage of ADNI started in 2011 and it mainly focus on comparing
cerebrospinal fluid and genetics biomarkers of those aged people who are not suffering from AD with
those who have mild symptoms of AD. The data employed in this study are from ADNI 1.

In this study, the demographics of participants must be considered for modeling the progression of
AD. The main risk factors for developing Alzheimers disease are age and gender [19]. Several
studies suggested that the incidence of the disease is higher in women than man which cannot be
simply attribute to the longer longevity of women but men and women have different pathogenic
mechanism for the progression of AD. [19,Cacciottolo M et al., 2016; Galen Buckwalter et

al., 1993) From Table 1, we observed significant differences between women and
men, which indicate the relevance of gender to studies of AD. Additionally, many
studies revealed that genes play an important role of Alzheimers progression.
Among all of the genetic risk factors, APOE4 alleles are the greatest genetic risk factor for
late-onset Alzheimers [20]. However, this doesnt mean that an individual will absolutely get
Alzheimers disease if he or she has one or two copies of APOE4, this just indicates increasing the
risk for developing the disease as well as lower the age of potential disease onset. Since some
scientific evidence exists to suggest that certain proposed Alzheimers prevention therapies may
work differently in people who carry at least one copy of the APOE4 gene, which indicates the
importance of the APOE4 gene, this gene should be considered as one of covariates for AD
progression modeling. Besides considering the demographics and genetic risk factors of the
participants, there are lots of other risk factors that are related to the AD, however, it should also

raise a concern that we might not have enough data from the ADNI dataset (table
**) to accurately specify the covariates effects in this complex modeling application
and that we are potentially over-fitted. In this study, the gender and the gene
APOE4 are the covariates for this modeling.

Biomarkers selection:
The disease progression model might be most effective with the pathologically
sensitive biomarkers that provide important avenues of research to enhance the

diagnosis of those individuals at risk for developing AD. Ideally, bio-markers should
be able to detect specific pathological features of AD neuropathology or have high
sensitivity for describing AD pathophysiological processes (Biagioni et al., 2011). As
we introduced before, the AD biomarkers for the pathophysiological process of AD in
the diagnostic criteria can be categorized into three groups: pathological, structural
or functional, and clinical. This ADNI dataset include CSF _ amyloid 42, tau as
pathological biomarkers indicating
pathology appears in the presymptomatic phase of AD. The FDG-PET and volume of
hippocampus as structural and functional biomarkers are included in our dataset.
Additionally, MMSE as one of biomarkers in clinical group (behavior, cognition and
function) in our dataset suggests the dementia with impairment and loss of function
in activities which usually appears in the prodeomal phase of AD.
In this study, the convergence of multi-states complex model for ADNI data was
adversely affected when too many biomarkers involved in modeling application.
Additionally, if data are inadequate for AD progression modeling, it might also raise
a concern for over-fitting. Considering these concerns, the following 3 biomarkers
were selected for use based on their sensitivity in assessing the progression of AD.
FDG: From figure ***, we can see that FDG as an indicator of synaptic dysfunction
and neurodegeneration in AD, suggests the hypo-metabolism in regional cerebral
appearing in the MCI stage and exacerbating in AD stage. In 2008, Mosconi L
implicated the brain hypo-metabolism is related to the progression of synaptic pathology in
AD (2008, Mosconi L).

Hippocampus: From the figure ***, we know that the decreasing of the hippocampal
volume appearing in the MCI stage and deteriorating in the AD stage.

\\
\\
Transition probability
The primary interest of this study is the pattern of progression of latent states of AD
in two year period. Interestingly, from the transition probability of latent AD states,
we can clearly see there are two profiles of AD progression and one was determined
by FDG-PET and another was led by the volume of hippocampus. The figure of
transition probability demonstrates the pattern of AD progression in 2 years, for
instance, if an individual has been classified in Normal state at the beginning, 2
years later, he or she has

chance transit to F1 state and

change transit to H1

state. However, if an individual was classified as a member of H1 state at baseline

time, he or she has

probability to transit to H2 state but doesnt have chance to

transit into F2. This transition pattern suggests few interactions between these two
profiles.

Additionally, it is impossible for the individuals who have been classified as AD to

reverse back to other latent state which indicates that AD is irreversible.

Two profiles:
From figure , compared with clinical results, the latent states from our model are
clearly distinct. Based on the distributions of FDG and hippocampus, we named the
six latent state as normal, F1, F2, H1, H2 and AD which F1 state represents the

hypo-metabolism in regional cerebral and F2 means the deterioration of hypometabolism than F1; while H1 implies the hippocampal abnormality and H2
suggests the worse abnormality in hippocampus than H1. Besides, from the figure
***, the density distributions of the latent states of FDG profile and Hippocampus
profile are clearly distinct which provides evidence of the existing of two progression
profiles.
Discussion

Two profiles:
1. Two profiles normal F1 f2 AD or normal H1 H2 and AD which advance our
understanding of AD:
The novelty of this study is these two profiles of the AD progression. Although
the exact mechanism for developing AD as well as triggers responsible for progressing to
advanced stage of disease, is still largely unknown, this study provides an advanced
understanding of the progression of AD by detecting the latent states of AD. Surprise, our model
revealed these two profiles. One is determined by FDG while another is by Hippocampus. It
implies the complexity of AD pathological progression. Specifically, our model reveals that some
patients have hypometabolism in regional cerebral but their volume of

hippocampus are still preserved while others with hippocampal atrophy but
dont have hypometabolism in cerebral. Recently, Hanseeuw B revealed that
cortical hypometabolism was observed in both A-MCI and A+ MCI whereas hippocampal
atrophy was mostly found in A+ MCI which implies that hippocampal atrophy is thus more
informative about A status than cortical hypometabolism ( 2016, Hanseeuw B). This finding

may explain the two existing profiles of AD progression. However, we should

conduct further studies to find more evidence about these two profiles.

Additionally, from this study, we noticed that some participants in normal

clinical stage was classified into F1 or H1 states with our model which implies
both of FDG-PET and the volume of hippocampus could be pre-symptomatic
biomarkers. Specifically, this study suggests that the hypo-metabolism in
cerebral and hippocampal atrophy could happen before clinical diagnosis. In
2008, Mosconi L suggested that hypometabolism and oxidative stress as crucial players in the
initiation and progression of synaptic pathology in AD. Later, in 2016, K. Abigail Andrews

provided evidence to support that there is pre-symptomatic, amyloidmediated acceleration of hippocampal atrophy which suggesting the
hippocampal atrophy could be one of pre-symptomatic biomarkers of AD
progression (K. Abigail Andrews, 2016). Both of these studies provided evidence
implying the hypo-metabolism and hippocampal atrophy could happen in presymptomatic phase of AD progression.
From the transition probability, few interaction between these two profiles
suggests these two pattern of AD progression are almost independent from
each other which provides advanced information of the relationship of the
hippocampal atrophy and the hypo-metabolism in cerebral of AD progression.
In 2015, a study suggesting a different pathological mechanism in the
particular area of the brain may explain the two profiles of AD progression
since the hippocampal atrophy is more structural while the hypo-metabolism
of cerebral is more functional in AD progression (2015, Rodriguez-Oroz MC).
In our further studies, we should focus on pre-screen AD patients with presymptomatic biomarkers since they are critical in early detection and
diagnosis.
Transition probability:
1. Two profile pattern (no interaction)

2. Irreversible
Further research
1. Prescreen AD with presymptomatic biomarkers
2. Put more biomarkers

The primary clinical interest in this study is in the pattern of progression of latent
states of AD, such as the probability of transiting from normal control state to F
H+ state in two years. Interestingly, this outcome suggested that the AD
progression is not simply from lower to higher 12 disease states, but a complicated
system which illustrated two profiles of the progression. One of them was primarily
determined by FDG-PET and the other by the volume of hippocampus. Compared
with the three traditional diagnoses states, this six states latent transition model
provided more details about the pattern of intermediary transitions between states.
These results give the progression profiles of FDG-PET and the volume of
hippocampus. In this AD progression model for two-year period, the FDG, as one of
the sensitive biomarkers for earliest clinically detectable evidence for brain
pathological changes, appeared to be abnormal earlier than the volume of
hippocampus since the decrease of FDG from normal state to F H+ in twoyear
period is dramatic and close to the FDG value of final AD state(Figure 2). This
pattern indicated that the FDG biomarker has potential to be a better biomarker for
detecting the AD in pre-symptomatic phase and contribute the preclinical screening
for treatment development since the abnormality appears earlier in FDG than the
volume of hippocampus. The transition probability in two years was estimated and
reported in Figure 3, which indicate AD irreversible. Two patterns of AD progression
emerge in this six-states latent transition model. The latent transition model

suggested six states of disease progression and two different pathological profiles.
One progression profile was mainly determined by the biomarker of FDG-PET and
the other by the volume of Hippocampus.

Reference:
1. batsch.. 2012. https://www.alz.org/documents_custom/world_report_2012_final.pdf
2. Elliott J. Mufson.. 2016
http://www.sciencedirect.com/science/article/pii/S0166432816303059
3. Lan Li. 2016 http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4829010/
4. Sperling et al., 2011
http://www.alz.org/documents_custom/diagnostic_recommendations_alz_preclinical.pdf
5. Li. C 2015 https://www.ncbi.nlm.nih.gov/pubmed/25892034
6. Fonteijn HM et al., 2011 http://www.ncbi.nlm.nih.gov/pubmed/22281676
7. Furiak NM, 2012 http://www.alzheimersanddementia.com/article/S15525260(11)02621-5/pdf
8. Nick Holford, 2014 http://onlinelibrary.wiley.com/doi/10.1111/bcp.12170/abstract

9. In 2012, Escudero J https://www.ncbi.nlm.nih.gov/pubmed/22886027

10. Iordanescu G, 2012 https://www.ncbi.nlm.nih.gov/pubmed/22189675
11. Madsen SK, 2015 https://www.ncbi.nlm.nih.gov/pubmed/26413208
12. Sukkar R, 2012 https://www.ncbi.nlm.nih.gov/pubmed/23366517

13. Holford N1, 2013 https://www.ncbi.nlm.nih.gov/pubmed/?term=Holford

%2C+2013+Alzheimer%27s+disease
14. Gomar et al., 2011 https://www.ncbi.nlm.nih.gov/pubmed/21893661
15. Biagioni et al., 2011 https://www.ncbi.nlm.nih.gov/pubmed/22076127
16. Liu, .. 2016 http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4763344/
17.

Estimating sample sizes for predementia Alzheimer's trials based on the Alzheimer's Disease

Neuroimaging Initiative.(Clinical report), Neurobiology of Aging. Jan, 2013, Vol. 34 Issue 1, p62, 11
18. https://en.wikipedia.org/wiki/Alzheimer%27s_Disease_Neuroimaging_Initiative#cite_note-4
19. http://www.ncbi.nlm.nih.gov/pubmed/20442496
20. http://alzdiscovery.org/cognitive-vitality/what-apoe-means-for-your-health
21. Mosconi L 2008
http://www.ncbi.nlm.nih.gov/pubmed/19076441
22. Hanseeuw B 2016. http://www.ncbi.nlm.nih.gov/pubmed/27232217

23. K. Abigail Andrews, 2016 http://www.neurobiologyofaging.org/article/S01974580(15)00517-5/pdf

24. 2015, Rodriguez-Oroz MC http://www.ncbi.nlm.nih.gov/pubmed/26008810

http://www.ncbi.nlm.nih.gov/pubmed/26984944