ARTICLE IN PRESS

Original Investigation

Machine Learning Algorithms Utilizing

Functional Respiratory Imaging May
Predict COPD Exacerbations
Maarten Lanclus, MSc, D10X XJohan Clukers, D1X XMD, D12X XCedric Van Holsbeke, D13X XPhD, D14X XWim Vos, D15X XPhD,
D16X XGlenn Leemans, D17X XMSc, D18X XBirgit Holbrechts, D19X XMD, D120X XKatherine Barboza, D12X XPhD, D12X XWilfried De Backer, D123X XPhD,
D124X XJan De Backer, D125X XPhD

X26D1 X
Rationale and Objectives: Acute chronic obstructive pulmonary disease exacerbations (AECOPD) have a significant negative impact on the
quality of life and accelerate progression of the disease. Functional respiratory imaging (FRI) has the potential to better characterize this dis-
ease. The purpose of this study was to identify FRI parameters specific to AECOPD and assess their ability to predict future AECOPD, by use
of machine learning algorithms, enabling a better understanding and quantification of disease manifestation and progression.
Materials and Methods: A multicenter cohort of 62 patients with COPD was analyzed. D127X X FRI obtained from baseline high resolution CT data
(unenhanced and volume gated), clinical, and pulmonary function testD128X w
X ere analyzed
D129X X and incorporated into machine learning algorithms.
Results: A total of 11 baseline FRI parameters could significantly distinguish ( p D130X X < 0.05) the development of AECOPD from a stable period.
In contrast, no baseline clinical or pulmonary function testD13X p
X arameters allowed significant classification. Furthermore, using Support Vector
Machines, an accuracy of 80.65% and positive predictive value of 82.35% could be obtained by combining baseline FRI features such as
total specific image-based airway volume and total specific image-based D132X X airway resistance, measured at functional residual capacity.
Patients who developed an AECOPD, showed significantly smaller airway volumes and (hence) significantly higher airway resistances at
baseline.
Conclusion: This study indicates that FRI is a sensitive tool (PPV 82.35%) for predicting future AECOPD on a patient specific level in con-
trast to classical clinical parameters.
Key Words: Pulmonary disease, chronic obstructive; Disease progression; Support vector machine; Patient-specific modeling; Radio-
graphic image interpretation, Computer-assisted.
© 2018 The Association of University Radiologists. Published by Elsevier Inc. All rights reserved.

INTRODUCTION
purulence, surpassing the daily variation, for which treatment
is initiatedD135X X (2 4).

C
hronic obstructive pulmonary disease (COPD) has a
high clinical and economic burden in the Western
worldD13X X (1). Acute COPD exacerbations (AECOPD)
often complicate the disease, which is already characterized
by a chronic and progressive natureD134X X (2). AECOPD are clini-
cally defined as episodes of increased respiratory symptoms,
particularly dyspnea, cough, and sputum production or
Acad Radiol 2018; &:1 9
From the FluidDA nv, Groeningenlei 132, 2550 Kontich, Belgium (M.L., C.V.H.,
W.V., G.L., K.B., W.D.B., J.D.B.); Faculty of Medicine and Health Sciences,
University of Antwerp (UAntwerpen), Antwerpen, Belgium (J.C., B.H., W.D.B.).
Received July 6, 2018; revised October 23, 2018; accepted October 28, 2018.
Contributions Conception and design: ML, JC, JDB, and WDB; Analysis and
interpretation: ML, JC, CVH, WV, GL, WDB, and JDB; Drafting the manuscript
for important intellectual content: ML, JC, WDB, JDB, BH, and KB. Take home
message: FRI captures disease progression in COPD. Address correspon-
dence to: M.L. e-mail: maarten.lanclus@fluidda.com
© 2018 The Association of University Radiologists. Published by Elsevier Inc.
All rights reserved.
https://doi.org/10.1016/j.acra.2018.10.022
AECOPD have a negative impact on the quality of life of
COPD patients, as they accelerate disease progression, and
can result in hospital admissions and deathD136X X (4,5). Therefore
an increased interest originated in the clinical phenotype of
patients with frequent exacerbations of COPD and the path-
ophysiology of these exacerbations. The unravelD137X Xing of this
mechanism will have a major impact on the prognosis and
treatment of COPD patientsD138X X (6 8).
Patients with a history of frequent exacerbations demon-
strate a higher risk of future exacerbationsD139X X (9,10). In the
ECLIPSE-study, exacerbations were followed longitudinally
during 3 years, showing that 71% of the frequent exacerbators
still exacerbated after the observation periodD140X X (9), 25% of the
frequent exacerbators stopped exacerbating and 25% of the
14X X
nonDfrequent exacerbators had an exacerbation within 3 years
(11). Therefore, therapeutic interventions for this patient
group (in clinical trials) are considered clinically meaningful
in the short termD142X X (10).

1
 ARTICLE IN PRESS
LANCLUS ET AL
Functional respiratory imaging (FRI) is a postDprocessing 143X X
technology that utilizes multiD14X Xslice D145X Xhigh-resolution computed
tomography D146X Xscans and computational fluid dynamics (CFD)
to assess the overall lung health and function in a regional
manner by quantifying endpoints as airway volume and resis-
tanceD147X X (12,13). FRI is considered a more sensitive method for
observing changes in airway volume and resistance than clas-
sical lung function tests (eD148X XgD149X X, forced expiratory volume in 1
second [FEV1])D150X X (12,14,15). This image-based method can
also be used to provide a comprehensive assessment of airway
tree changesD15X X (14,16).
The optimal management of COPD will continue to
evolve and become more complicated as benefits of therapy
are balanced with its limitations, costs, and the individual
needs of each patientD152X X(1). Predicting future COPD exacerba-
tions (eD153X XgD154X X, frequent exacerbators) with high accuracy, will
allow clinicians to tailor treatments at an early stage to pre-
vent these exacerbations, and accompanying morbidity and
mortalityD15X X (7,17,18). The aim of this study was to, retrospec-
tively, evaluate and compare the power of pulmonary func-
tion tests (PFT) and FRI (parameters) in predicting future
AECOPD.D156X X
This study explores whether both PFT and FRI parame-
ters, at baseline, can predict the eventual exacerbation situa-
tion on a D157X Xpatient-specific level.
METHODS
We retrospectively D158X Xanalyzed data from two studies mainly
performed at the Antwerp University Hospital. One study,
“Analysis of airway responses in severe COPD patients to
DaxasD159X X, using CT based functional respiratory imaging
(DAXAS; NCT01480661),” explored the effects of roflumi-
last on airways via FRID160X X (19). The second study, “Correlation
of functional respiratory imaging parameters with lung func-
tion parameters and patient reported outcome measures dur-
ing exacerbation of COPD (EXA; NCT01684384),”
correlated FRI to COPD symptoms and PFT during exacer-
bations of COPDD16X X (20). Permission from the local Ethical
Committee for analysis of these data was obtained.
For all patients enrolled in these studies, demographic data
D162X Xwere collected such as age, sex, smoking history (pack years),
body mass indexD163X X, and GOLD stage (2017 classification (3)).
CAT-scores at the start and end of each study were regis-
tered. Amount of exacerbation was documented from the
patient and the available study files. A patient was defined as a
frequent exacerbator in the past year, if there were more than
two exacerbations, or one leading to hospitalization, as per
GOLD guidelinesD164X X (3). AECOPD was defined as (new) use of
antibiotic and/or corticosteroid treatment during study fol-
low-up. Follow-up time was respectively 168 days for the
DAXAS study and 43 days for the EXA study. The 23
patients, recruited from the DAXAS study, were all treated
with triple inhalation therapy, as per study protocol. Regis-
tration of stable maintenance therapy was impossible for the
2
Academic Radiology, Vol &, No &&, && 2018
EXA study due to the enrolment of patients with AECOPD
at that time.
Across both studies, FEV1, vital capacity (VC), functional
respiratory capacity (FRC), and inspiratory capacity D165X Xwere
recorded (both absolute and in percentage predicted) for the
PFT’s. The examined FRI parameters included (specific)
image-based airway resistance for the central, distal, and total
airways at total lung capacity (TLC) and FRC ((s)imaging air-
way resistance [iRaw]); image-based lobe volumes at TLC
and FRC (iVlobe); and (specific) image-based airway volume
for the central, distal, and total airways, at TLC and FRC ((s)
imaging airway volume [iVaw]). A total of 94 parameters was
considered in this work, iD16X XeD167X X, 90 FRI parameters and 4 clinical
parameters (FEV1, FVC, FEV1 percent predicted, FVC per-
cent predicted).
The level of inspiration during computed tomography
(CT) was monitored with a commercially available spirome-
try system that enables real-time monitoring of the breathing
cycle. (Spirostik; Geratherm Respiratory GmbH, Kissingen,
Germany) Different CT-scans were used (multicenter studies)
although this does D168X Xnot influence the results as images were
acquired during randomized controlled trials with a standard-
ized scanning protocol, including phantom testing and other
quality insurance tests, to allow for robust FRI calculations.
The baseline prediction of COPD exacerbation included
62 patients from the original population, based on complete-
ness of datasets. Patients with missing values for certain
regional parameters, due to patient-specific pathologies were
not included for the feature selection methods of the super-
vised machine learning classifier. Demographic data for the
cohort can be found in Table 1.
Regarding quantitative CT Imaging Processing, CT scan
data D169X Xwere converted into 3D models of airways and lung
lobes using Mimics (Materialise, Leuven, Belgium) a com-
mercially available validated software package (Food and
Drug Administration, K073468; Conformite Europeenne
certificate, BE 05/1191.CE.01). Other software used
included; TGrid 14.0 (Ansys Inc, Canonsburg, PA) for 3D
meshing and Fluent 14.0 (Ansys Inc, Canonsburg, PA) for
CFD simulations.
Segmentation of the tracheobronchial tree was done using
directional thresholding with automated leakage detection.
Automatic airway segmentation was performed up to the
point where no distinction could be made between the intra-
D170X Xluminal and alveolar air. Following automated segmentation
of the bronchial tree, the airways were manually checked.
Missing branches were added to the bronchial tree and incor-
rect branches were deleted when necessary. The respiratory
tract was reconstructed down to the level of airways with a
diameter of 1 2 mm, beyond this point, the high-resolution
computed tomographyD17X X resolution is insufficient to distinguish
alveolar from intraluminal air. The segmented airway tree
was converted into a 3D model that was smoothed using a
volume compensation algorithm. The smoothed model was
trimmed perpendicular to the airway centerD172X Xline at the trachea
(using the middle point of the superior side of the sternum as
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2018 MACHINE LEARNING ALGORITHMS

D5X X
TABLE 1. Patient CharacteristicsD 6X X

Sex Male 39 (62D.9%) 7X X

Female 23 (37D.1%) 8X X
D9X X
COPD stage GOLD II 24 (38.D7%) 10X X GOLD A 1X X
5 (8D.1%)
GOLD III 31 (50%) GOLD B 12 (19D.4%) 12X X
GOLD IV 13X X
7 (11D.3%) GOLD C 14X X
1 (1D.6%)
GOLD D 44 (71%)
Exacerbation during follow-up AECOPD 15X X
23 (37D.1%)
Stable 16X X
39 (62D.9%)
Total AECOPD Stable D17X X 18X v
pD X alue
Sex
Male 19X X
39 (62D.9%) 20X X
12 (19D.4%) 21X X
27 (43D.5%) 0.179
Female 2X X
23 (37D.1%) 23X X
11 (17D.7%) 12 (19.D4 24X X %)
Study
EXA 25X X
39 (62D.9%) 26X X
12 (19D.4%) 27X X
27 (43D.5%) 0.179
DAXAS 28X X
23 (37D.1%) 29X X
11 (17D.7%) 30X X
12 (19D.4%)
GOLD (FEV1)
II 24 (38D.7%) 31X X 32X X
6 (9D.7%) 18 (29%) 0.291
III 31 (50%) 14 (22D.6%) 3X X 17 (27D.4 34X X %)
IV 35X X
7 (11D.3%) 36X X
3 (4D.8%) 37X X
4 (6D.5%)
GOLD (Stage)
A 38X X
5 (8D.1%) 4 (6D.5%) 39X X 40X X
1 (1D.6%) 0.151
B 12 (19D.4%) 41X X 42X X
3 (4D.8%) 9 (14D.5%) 43X X
C 4X X
1 (1D.6%) 45X X
1 (1D.6%) 0 (0%)
D 44 (71%) 16 (25D.8%) 46X X 28 (45D.2%) 47X X
Exacerbation history
48X X
NonDfrequent 17 (27D.4%) 49X X 7 (11D.3%) 50X X 10 (16,3%) 0.683
Frequent 45 (72D.6%) 51X X 16 (25D.8%) 52X X 29 (46,8%)
Age (years) 67 § 8D.6 53X X (45D54X X 84) 65D.7 5X X § 9D.1 56X X 67,7 § 8,3 0.385
BMI (kg/m2) 26D.8 57X X § 7D.5 58X X (15D.9D 59X X 60X X 50D.6) 61X X 27D.5 62X X § 8 26,4 § 7,2 0.59
Pack years (years) 52D.2 63X X § 27D.5 64X X (16D65X X 135) 44 § 24D.4 6X X 57,1 § 28,4 0.071
CAT score (points) 21D.5 67X X § 7D.5 68X X (5D69X X 37) 21D.3 70X X § 7D.8 71X X 21,6 § 7,5 0.894
FEV1 (% predicted) 46D.1 72X X § 14D.1 73X X (15D.9D 74X X 75X X 74D.4) 76X X 42D.2 7X X § 12D.7 78X X 48,4 § 14,5 0.093
FRC (% predicted) 149D.6 79X X § 38D.1 80X X (79D81X X 252) 151 § 38D.1 82X X 148,9 § 38,5 0.841
IC (% predicted) 83X X § 22D.1
69D.4 84X X (29D85X X 119) 86X X § 28D.5
76D.3 87X X 65,3 § 16,4 0.057

AECOPD, Acute chronic obstructive pulmonary disease exacerbations; BMI, body mass index; COPD, chronic obstructive pulmonary dis-
ease; FEV1, forced expiratory volume in 1 second; FRC, functional respiratory capacity; IC, inspiratory capacity.

a landmark) and at each terminal bronchus. Remaining volume is defined as the segmented airway volume starting
D173X Xartifacts due to noise in the CTs were then manually removed from the D174X Xthird bifurcation (4th generation), which include
from the model. Finally a series of manual quality checks segmental bronchi (B1R, B2R, etcD175X X) and subsegmental bron-
were performed. Total time for the automated steps and chi that are discernible on the scan (bronchi with a diameter
manual quality checks varied from 2 to 6 hours per scan. > 1D176X X 2 mmD;17X X FigD178X X 1) (21). iVlobe from the CT images at FRC
FRI is comprised of a combination of airway segmenta- (eD179X XgD180X X, expiratory CT) and TLC (eD18X XgD182X X, inspiratory CT) were also
tion, lung volume segmentation by lobe, and airway resis- extracted (FigD183X X1). By means of application of CFD on the seg-
tance calculations based on computational flow simulation mented airway model, iRaw were calculatedD184X X (13). Lungs
using boundary conditions provided by the lobe expansions were split into lobes by identification of the fissure lines from
from FRC to TLC, allowing calculation of iVlobe, iVaw, the CT scan. This allowed determination of total lung vol-
and iRaw, as well as their specific values (corrected for lobe ume and of the volume of each lobe individually.
volume). The airways were subsequently divided in central FRI has been validated by comparison with gamma scin-
and distal regions. The central airway volume is defined as tigraphy and single-photon emission computerized tomogra-
the region from the trachea to the segmental bronchi (the phyD185X X (13,22).
combination of the trachea, main bronchus left, main bron- For the baseline prediction of future exacerbation, the stu-
chus right, truncus intermedius, right upper lobe bronchus, dent tD186X X test was initially applied to assess individual biomarkers
right middle lobe bronchus, right lower lobe bronchus, left with significant predictive power for eventual exacerbation.
upper lobe bronchus, superior division bronchus, lingular The significance level was set at 0.05. Next, supervised
bronchus, left lower lobe bronchus). The distal airway machine learning with support vector machines was used in

3
 ARTICLE IN PRESS
LANCLUS ET AL Academic Radiology, Vol &, No &&, && 2018

RESULTS

In this cohort of 62 patients, 23 patients developed an

Figure 1. Central and distal airways (left) and lobe volumes (right)
as defined by FLUIDDA’s segmentation process. The central airway
volume is defined as the region from the trachea until the segmental
bronchi (the combination of the trachea, left main bronchus, right
main bronchus, truncus intermedius, right upper lobe bronchus, right
middle lobe bronchus, right lower lobe bronchus, left upper lobe
bronchus, superior division bronchus, lingular bronchus, and left
lower lobe bronchus). The distal airway volume is defined as the seg-
mented airway volume starting from the third bifurcation (fourth gen-
eration), which includes segmental bronchi (B1R, B2R, etcD)1X X and
subsegmental bronchi that are discernible on the scan (bronchi with
a diameter of >1 2 mm).
order to examine the predictive power of multiple feature
combinations, using significant features derived from the stu-
dent tD187X X test. A Radial Basis Functional D18X Xkernel, with default
tuning parameters, was used (cost: 1 and gamma: 1/data
dimension, 6 tuning steps) to enable nonD189X Xlinear hyperplanes
for classification purposes. Classification was performed in 1-
dimensional, 2-dimensional, and 3-dimensional space. All
Support Vector Machine models were optimized for accu-
racy. The cohorts were trained with the leave one out cross
validation method. The accuracies, sensitivities, specificities,
and positive predictive values as discussed here are the results
from the training process.
All analyses were performed using the open-source statisti-
cal environment R version 3.2.5 or higher (The R Founda-
tion for Statistical Computing, Vienna, Austria).
AECOPD during follow up, and 39 patients were considered
stable.
All patient characteristics are summarized in Table 1.
Comparison between stable patients and patients with an
exacerbation showed no differences in baseline characteris-
ticsD190X X—especially regarding disease severity (symptoms,
PFT’sD—FEV 19X X 1, inspiratory capacity,D192X X andD193X X FRCD194X X—, exacerba-
tion history, and GOLD stages). A large number of patients
D195X Xare classified as a frequent exacerbator. This is due to the fact
that they were recruited at time of an acute exacerbation,
mostly leading to D196X Xhospitalization, and thus meeting the pre-
determined GOLD definition.
Table 2 and Figure 2 show the significant baseline predic-
tors of AECOPD development based on the student tD197X X test
with a significance level of 0.05. A total of 11 baseline FRI
parameters could significantly distinguish development of
AECOPD from a stable period. However, no baseline clini-
cal or PFT parameters could significantly distinguish between
the stable and exacerbating cohort. As shown in Figure 3,
default clinical parameters FEV1 and FEV1 percent predicted
(FEV1pp) were nonD198X Xsignificant (D19X Xp = 0.0841 and D20X Xp = 0.1007,
respectively) in terms of prediction of eventual exacerbation.
Table 3 summarizes the results (iD201X XeD20X X, accuracies) of the 1-, 2-
and 3-dimensional machine D203X Xlearning-based classification for
all significant feature combinations. Accuracies in classifica-
tion based on one feature (1-dimensional classification) range
from 61.29% to 72.58%. iRaw of the right upper lobe at
FRC is the best baseline predictor for eventual frequent
exacerbation with maximal accuracy of 72.58%. Accuracies
in classification based on two features (2-dimensional classifi-
cation) range from 62.90% to 80.65%. The maximal accuracy
of 80.65% is obtained by combining baseline FRI features,

D8X X
TABLE 2. Baseline Predictors D89X X
and Their D90X X 91X Values,
pD D92X X D93X X
Including D94X X
Significant D95X X
FRI Predictors D96X X PFT PredictorsD
and All D97X X 98X X

Predictor Type D9X X

Baseline Predictor Zone Level D10X X 10X Value
pD D102X X

FRI iRaw DISTAL FRC 0.031438571

iRaw RUL FRC 0.037807489
iRaw TOTAL FRC 0.016851996
iVaw CENTRAL FRC 0.037735049
iVaw TOTAL FRC 0.03760724
siVaw CENTRAL FRC 0.015711939
siVaw RUL FRC 0.017558065
siVaw TOTAL FRC 0.012632473
siRaw DISTAL FRC 0.049747328
siRaw RUL FRC 0.030134498
siRaw TOTAL FRC 0.03054237
PFT FVC 0.68832987
FVC (percent predicted) 0.80575310
FEV1 0.08409363
FEV1 (percent predicted) 0.10066426

FEV1, forced expiratory volume in 1 second; FRC, functional respiratory capacity; iRaw, imaging airway resistance; iVaw, imaging airway vol-
ume; siRaw, specific imaging airway volume; siVaw, specific imaging airway resistance.

4
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2018 MACHINE LEARNING ALGORITHMS

Figure 2. Significant baseline predictors of AECOPD development based on the student tD2X Xtest with a significance level p
D3X X < 0.05. A total of 11
baseline FRI parameters could significantly distinguish development of AECOPD from a stable period. No baseline clinical or PFT parameters
could significantly distinguish between the stable and exacerbating cohort. AECOPD, Acute chronic obstructive pulmonary disease exacerba-
tions; FRI, functional respiratory imaging; PFT, pulmonary function test.

total specific imaging airway volume (siVaw), and total spe-
cific imaging airway resistance (siRaw) at FRC. Patients who
developed AECOPD, show significantly smaller airway vol-
umes and, hence, significantly higher airway resistances at
baseline D204X Xaccuracies in classification based on three features (3-
dimensional classification) range from 61.29% to 80.65%.
The maximal accuracy of 80.65% is obtained by combining
baseline FRI features, iRaw of the right upper lobe, total
iVaw, and siRaw of the right upper lobe at FRC. It should
be noted that both for 2-dimensional and 3-dimensional
classification models the same accuracy level is reached;D205X X
therefore it is preferable to make the classification based on
the 2-feature model, as the parameters are more general. For
this model, a sensitivity of 60.87% and a specificity of 92.31%
were obtained. Furthermore, a positive predictive value of
82.35% could be acquired for this 2-dimensional case. Posi-
tives were assigned to exacerbators and negatives to stable
patients. In general, it can be concluded that future develop-
ers of AECOPD tend to present with smaller airway volumes
and higher airway resistances at FRC at baseline.

5
 ARTICLE IN PRESS
LANCLUS ET AL Academic Radiology, Vol &, No &&, && 2018

4X X
Figure 3. Default clinical parameters FEV1 and FEV1 percent predicted (FEV1pp) are nonDsignificant in terms of prediction of eventual
AECOPD. AECOPD, Acute chronic obstructive pulmonary disease exacerbations; FEV1, forced expiratory volume in 1 second.

DISCUSSION FRI parameters with the best accuracy (80.65%) of predicting

an AECOPD, when combined in an SVM model. Seeing
We report 11 FRI parameters that could significantly distin- only D206X Xtwo parameters are needed for eventual prediction of
guish between stable patients and ones that develop an AECOPD with SVM, a drastic reduction in processing time
AECOPD. Total siVaw and siRaw, at FRC, are the baseline could be expected when used in clinical practice. Indeed,

6
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2018
D103X X
TABLE 3. Results From 104X X
1- ,2- ,3-DDimensional D105X X
Classification
with Support Vector MachinesD106X X
Dimensions D107X X
Baseline Predictors Accuracy
1  iRaw RUL at FRC 72.58%
2  siVaw TOTAL at FRC 80.65%
 siRaw TOTAL at FRC
3  iRaw RUL at FRC 80.65%
 iVaw TOTAL at FRC
 siRaw RUL at FRC
FRC, functional respiratory capacity; iRaw, imaging airway resis-
tance; iVaw, imaging airway volume; siRaw, specific imaging airway
volume; siVaw, specific imaging airway resistance.
one would only have to look at post processing steps for those
D207X Xtwo parameters, instead of for the full spectrum of FRI
parameters.
D208X XFEV1D209X X, and pulmonary function tests as a whole, are widely
used as a guideline in the care for COPD patients (3). How-
ever, they exhibit significant shortcomings on the individual
patient level, as FEV1 does D210X Xnot correlate well to dyspnea,
exercise tolerance or health statusD21X X (3,23,24). FEV1 is of lim-
ited value in prognosis in terms of disease progression and
mortality and it is not sensitive enough for changes in the
small airways or other pathophysiological mechanisms in
COPDD21X X (11). Changes in pulmonary function, mainly FEV1
or peak expiratory flow rate (PEFR), are poorly sensitive in
the individual diagnosis of exacerbations, even when mea-
sured daily, possibly because the individual variability is larger
than the mean change that occurs during an exacerbationD213X X
(25,26).
A lower FEV1, especially if less than 50% predicted, is asso-
ciated with more AECOPDD214X X (5,27). Yet, a subset of these
patients with a low FEV1 does not experience frequent exac-
erbationsD215X X (23). A comparative study of different prediction
models (on exacerbations of COPD) identified patients at
risk of developing an AECOPD. FEV1 had no predictive
power on an individual patient level. Based on the results of
these prediction models, it is difficult to design interventions
Figure 4. Airway collapse at exhalation (FRC) in the AECOPD
patient (left) and no airway collapse for the stable patient. AECOPD,
Acute chronic obstructive pulmonary disease exacerbations; FRC,
functional respiratory capacity.
MACHINE LEARNING ALGORITHMS
or provide treatment advice to reduce the risk of an exacerba-
tionD216X X (28,29).
The exact pathophysiology of AECOPD remains unclear,
not in the least due to a lack of a widely accepted consensus
definition and a sensitive method to pinpoint the exact
moment of declineD217X X (30). Although many studies suggest that
the breathlessness during AECOPD is caused by airway nar-
rowing and increased ventilation-perfusion mismatchD218X X
(25,31). This decrease in airway caliberD219X X is indicated by reduc-
tion in FEV1 and PEFR during AECOPD (26) and the
increase of these measures during recoveryD20X X (32).
Our results suggest that a further reduction in airway vol-
ume and ensuing rise in airway resistance in the chronic nar-
rowed (small) airways may be the driver for an AECOPD,
which the patient perceives as being more symptomatic (eD21X XgD2X X,
dyspnoeic), beyond the daily variations in symptoms. This is
in line with a previous study, using FRI, where changes in
siVaw and iRaw at FRC show an excellent correlation with
patient reported outcomes D23X X(20).
Figure 4 shows typical airway volume renders at FRC for
both an AECOPD developer and a stable patient. From the
figure it shows that the AECOPD developer experiences air-
way collapse during exhalation, which results in a smaller spe-
cific airway volume and higher airway resistance. Seeing the
stable patient does D24X Xnot experience airway collapse at the time
of the scan, it might explain the absence of future exacerba-
tion.
Quantitative CT (qCT) has been used to phenotype
COPD exacerbators. Han M.K. et alD25X X reported an association
of more lung emphysema and airway wall thickness with
AECOPD, independent of the severity of airflow obstructionD26X X
(33). Our D27X Xtwo-feature FRI model proves to be highly accu-
rate in the prediction of an AECOPD on a patient-specific
level, in contrast to prediction models based on classical
PFT’s (27 29) or patient reported outcomesD28X X’D29X X (34), as stated
earlier. To our knowledge, total siVaw and total siRaw at
FRC are the first radiological parameters, described in the lit-
erature, with this kind of predictive powerD230X X (33,35,36). qCT
has added value for phenotyping in COPD and AECOPD,
especially with the recent determination that biomarkers
prove no added value to clinical information in the prediction
of COPD exacerbation frequencyD231X X (37,38).
This prediction model could prove useful in the manage-
ment of COPD patients. After a hospital admission for
AECOPD up to 20% of patients are reD23X Xadmitted within
30 days of discharge due to a new exacerbationD23X X (39,40). A
qCT, with FRI analysis, could provide the insight for a con-
sidered decision to discharge or more intensified treatment,
eventually including nonD234X Xinvasive ventilation (41), and fol-
low-up. A similar reasoning could be made when patients are
diagnosed with COPD, as it may help identify the subset of
patients at high-risk to develop AECOPD and thus guide the
initial treatment.
Ultimately, this is a retrospective study. Validation in a
prospectively studied cohort, with incorporation of treatment
decisions to investigate the above postulated hypothesis, is
7
 ARTICLE IN PRESS
LANCLUS ET AL Academic Radiology, Vol &, No &&, && 2018

desirable. The total studied population is rather small, how- 4. Miravitlles M, Ferrer M, Pont A, et al. Effect of exacerbations on quality of
ever, this is not unusual for studies using FRID235X X (12,14,15,20). life in patients with chronic obstructive pulmonary disease: a 2 year fol-
low up study. Thorax 2004; 59:387–395.
Recent prospective studies using qCT enrolled approxi- 5. Donaldson GC, Seemungal TAR, Bhowmik A, et al. Relationship between
mately 40 patientsD236X X (35,36). The follow-up time is short (43 exacerbation frequency and lung function decline in chronic obstructive
and 168 days), although considering that reD237X Xadmissions win- pulmonary disease. Thorax 2002; 57:847–852.
6. Hurst JR. Exacerbation phenotyping in chronic obstructive pulmonary
dows studied are typically 30 days, this period is sufficient. disease. Am J Respir Crit Care Med 2011; 184:625–626.
The influence of early (<D238X X30 days) or late (>D239X X30 days) 7. Wilkinson TM, Donaldson GC, Hurst JR, et al. Early therapy improves
AECOPD on the prediction significance of the PFT or FRI outcomes of exacerbations of chronic obstructive pulmonary disease.
Am J Respir Crit Care Med 2004; 169:1298–1303.
parameters was not investigated. Although half of the studied 8. Zhou A, Zhou Z, Zhao Y, et al. The recent advances of phenotypes in
population showed an AECOPD before day 43 whereby the acute exacerbations of COPD. Int J Chron Obstruct Pulmon Dis 2017;
presumption could be made that patients with early or late 12:1009–1018.
9. Hurst JR, Vestbo J, Anzueto A, et al. Evaluation of CLtIPSEI. Susceptibil-
AECOPD still show the same baseline FRI characteristics. ity to exacerbation in chronic obstructive pulmonary disease. N Engl J
Similar studies using PEFR and FEV1 showed the same Med 2010; 363:1128–1138.
AECOPD behavioD240X Xr with no difference in baseline PFTD241X X 10. Vestbo J, Agusti A, Wouters EF, et al. Evaluation of CLtIPSESI. Should
we view chronic obstructive pulmonary disease differently after
(26,42). ECLIPSE? A clinical perspective from the study team. Am J Respir Crit
Our population contains more patients with advanced Care Med 2014; 189:1022–1030.
COPD (iD24X XeD243X X, mean FEV1pp < 50% and overrepresentation of 11. Vestbo J, Anderson W, Coxson HO, et al. Evaluation of COPD Longitudi-
nally to Identify Predictive Surrogate End-points (ECLIPSE). Eur Respir J
GOLD stages III-IV D24X Xand D), potentially distorting the data as 2008; 31:869–873.
a resultD245X X (5,27). Even though our FRI model is significant 12. De Backer J, Vos W, Vinchurkar S, et al. The effects of extrafine beclo-
among these severe patients where PFT’s cannot, the ques- metasone/formoterol (BDP/F) on lung function, dyspnea, hyperinflation,
and airway geometry in COPD patients: novel insight using functional
tion remains if the predictive power is as high in patients with respiratory imaging. J Aerosol Med Pulm Drug Deliv 2015; 28:88–99.
mild COPD (FEV1pp > 50 % and GOLD stages I-II). This is 13. De Backer JW, Vos WG, Vinchurkar SC, et al. Validation of computational
a great incentive for expanding the current dataset with dif- fluid dynamics in CT-based airway models with SPECT/CT. Radiology
2010; 257:854–862.
ferently staged patients to make for more robust predictive 14. De Backer LA, Vos W, De Backer J, et al. The acute effect of budesonide/
246X X applications.
modelDing formoterol in COPD: a multi-slice computed tomography and lung func-
We were not able to study the influence of inhalation tion study. Eur Respir J 2012; 40:298–305.
15. De Backer LA, Vos WG, Salgado R, et al. Functional imaging using com-
treatment regimens in this cohort (approximately D247X Xone-third puter methods to compare the effect of salbutamol and ipratropium bro-
of patients were on triple therapy (inhalation corticosteroids, mide in patient-specific airway models of COPD. Int J Chron Obstruct
long acting beta-2-agonist and long acting muscarinic antag- Pulmon Dis 2011; 6:637–646.
16. De Backer J, Van Holsbeke C, Vos W, et al. Assessment of lung deposi-
onist)). On the other hand this retrospective COPD cohort tion and analysis of the effect of fluticasone/salmeterol hydrofluoroal-
resembles the heterogenic treatments that are used in our kane (HFA) pressurized metered dose inhaler (pMDI) in stable persistent
population. COPD guidelines have evolved over the last asthma patients using functional respiratory imaging. Expert Rev Respir
Med 2016; 10:927–933.
years (3), especially since time of recruitment of the original 17. Wedzicha JA, Brill SE, Allinson JP, et al. Mechanisms and impact of the
studies (before 2013) and regarding the treatment of the frequent exacerbator phenotype in chronic obstructive pulmonary dis-
exacerbation phenotype. ease. BMC Med 2013; 11:181.
18. Celli BR, Decramer M, Wedzicha JA, et al. An official American Thoracic
Society/European Respiratory Society statement: research questions in
COPD. Eur Respir Rev 2015; 24:159–172.
CONCLUSION 19. De Backer W, Vos W, Van Holsbeke C, et al. The effect of roflumilast in
addition to LABA/LAMA/ICS treatment in COPD patients. Eur Respir J
D248X XFRID249X X features, total D250X XsiVawD251X X and total D25X XsiRawD253X X, measured at D254X XFRCD25X X, 2014; 44:527–529.
have shown to be significant predictors for future AECOPD, 20. Vos W, Van Holsbeke C, Van Geffen W, et al. Changes in functional
respiratory imaging (FRI) endpoints correlate with changes in patient
in contrast to classical clinical parameters. Results showed reported outcomes (PRO) after recovering from acute COPD exacerba-
that smaller airway volumes and higher airway resistances at tion. Eur Respir J 2015; 46.
FRC tend to significantly predispose to AECOPD. An accu- 21. Backer JD. Method for determining treatments using patient-specific
lung models and computer methods. Google Patents 2014.
racy of 80.65 % and positive predictive value of 82.35% could 22. Vinchurkar S, Backer LD, Vos W, et al. A case series on lung deposition
be obtained by means of support vector machine models. analysis of inhaled medication using functional imaging based computa-
Future prospective validation is warranted. tional fluid dynamics in asthmatic patients: effect of upper airway mor-
phology and comparison with in vivo data. Inhal Toxicol 2012; 24:81–88.
23. Agusti A, Calverley PM, Celli B, et al. Evaluation of CLtIPSEi. Characterisation
REFERENCES of COPD heterogeneity in the ECLIPSE cohort. Respir Res 2010; 11:122.
24. Jones PW, Agusti AG. Outcomes and markers in the assessment of
1. Guarascio AJ, Ray SM, Finch CK, et al. The clinical and economic bur- chronic obstructive pulmonary disease. Eur Respir J 2006; 27:822–832.
den of chronic obstructive pulmonary disease in the USA. Clinicoecon 25. Papi A, Luppi F, Franco F, et al. Pathophysiology of exacerbations of
Outcomes Res 2013; 5:235–245. chronic obstructive pulmonary disease. Proc Am Thorac Soc 2006;
2. Wedzicha JAEC-C, Miravitlles M, Hurst JR, et al. Management of COPD 3:245–251.
exacerbations: a European Respiratory Society/American Thoracic Soci- 26. Seemungal TA, Donaldson GC, Bhowmik A, et al. Time course and
ety guideline. Eur Respir J 2017: 49. recovery of exacerbations in patients with chronic obstructive pulmonary
3. disease Gifcol. Global strategy for the diagnosis, managment and pre- disease. Am J Respir Crit Care Med 2000; 161:1608–1613.
vention of chronic obstructive pulmonary disease Available from: http:// 27. Marin JM, Carrizo SJ, Casanova C, et al. Prediction of risk of COPD
www.goldcopd.org; 2017. exacerbations by the BODE index. Respir Med 2009; 103:373–378.

8
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2018
28. Guerra B, Gaveikaite V, Bianchi C, et al. Prediction models for exacerba-
tions in patients with COPD. Eur Respir Rev 2017; 26.
29. Hoogendoorn M, Feenstra TL, Boland M, et al. Prediction models for exacer-
bations in different COPD patient populations: comparing results of five large
data sources. Int J Chron Obstruct Pulmon Dis 2017; 12:3183–3194.
30. Rodriguez-Roisin R. Toward a consensus definition for COPD exacerba-
tions. Chest 2000; 117:398S–401S.
31. O'Donnell DE, Parker CM. COPD exacerbations ¢ 3: Pathophysiology.
Thorax 2006; 61:354–361.
32. Stevenson NJ, Walker PP, Costello RW, et al. Lung mechanics and dys-
pnea during exacerbations of chronic obstructive pulmonary disease.
Am J Respir Crit Care Med 2005; 172:1510–1516.
33. Han MK, Kazerooni EA, Lynch DA, et al. Chronic obstructive pulmonary
disease exacerbations in the COPDGene study: associated radiologic
phenotypes. Radiology 2011; 261:274–282.
34. Strassmann A, Frei A, Haile SR, et al. Commonly used patient-reported
outcomes do not improve prediction of COPD exacerbations. Chest
2017; 152:1179–1187.
35. Cheng T, Wan H, Cheng Q, et al. Computed tomography manifestation
of acute exacerbation of chronic obstructive pulmonary disease: a pilot
study. Exp Ther Med 2016; 11:519–529.
MACHINE LEARNING ALGORITHMS
36. Hackx M, Ghaye B, Coche E, et al. Severe COPD Exacerbation: CT fea-
tures. COPD J Chronic Obstr Pulm Dis 2015; 12:38–45.
37. Donaldson GC, Wedzicha JA. Prediction of chronic obstructive pulmo-
nary disease exacerbation frequency. clinical parameters are still better
than biomarkers. Am J Respir Crit Care Med 2017; 195:415–416.
38. Keene JD, Jacobson S, Kechris K, et al. Biomarkers predictive of exacer-
bations in the SPIROMICS and COPDGene cohorts. Am J Respir Crit
Care Med 2017; 195:473–481.
39. Bahadori K, FitzGerald JM. Risk factors of hospitalization and readmis-
sion of patients with COPD exacerbation—systematic review. Int J
Chronic Obstr Pulm Dis 2007; 2:241–251.
40. Guerrero M, Crisafulli E, Liapikou A, et al. Readmission for acute exacer-
bation within 30 days of discharge is associated with a subsequent pro-
gressive increase in mortality risk in COPD patients: a long-term
observational study. PLoS One 2016; 11:e0150737.
41. De Backer L, Vos W, Dieriks B, et al. The effects of long-term noninvasive
ventilation in hypercapnic COPD patients: a randomized controlled pilot
study. Int J Chron Obstruct Pulmon Dis 2011; 6:615–624.
42. Donaldson GC, Law M, Kowlessar B, et al. Impact of prolonged exacer-
bation recovery in chronic obstructive pulmonary disease. Am J Respir
Crit Care Med 2015; 192:943–950.