Artificial intelligence and machine learning in respiratory medicine

Abstract

Introduction: The application of artificial intelligence (AI) and machine learning (ML) in medicine and in particular in respiratory medicine is an increasingly relevant topic.
Areas covered: We aimed to identify and describe the studies published on the use of AI and ML in the field of respiratory diseases. The string “(((pulmonary) OR respiratory)) AND ((artificial intelligence) OR machine learning)” was used in PubMed as a search strategy. The majority of studies identified corresponded to the area of chronic obstructive pulmonary disease (COPD), in particular to COPD and chest computed tomography scans, interpretation of pulmonary function tests, exacerbations and treatment. Another field of interest is the application of AI and ML to the diagnosis of interstitial lung disease, and a few other studies were identified on the fields of mechanical ventilation, interpretation of images on chest X-ray and diagnosis of bronchial asthma.
Expert opinion: ML may help to make clinical decisions but will not replace the physician completely. Human errors in medicine are associated with large financial losses, and many of them could be prevented with the help of AI and ML. AI is particularly useful in the absence of conclusive evidence of decision-making.

Full text

Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=ierx20
Expert Review of Respiratory Medicine
ISSN: 1747-6348 (Print) 1747-6356 (Online) Journal homepage: https://www.tandfonline.com/loi/ierx20
Artificial intelligence and machine learning in
respiratory medicine
Evgeni Mekov, Marc Miravitlles & Rosen Petkov
To cite this article: Evgeni Mekov, Marc Miravitlles & Rosen Petkov (2020) Artificial intelligence
and machine learning in respiratory medicine, Expert Review of Respiratory Medicine, 14:6,
559-564, DOI: 10.1080/17476348.2020.1743181
To link to this article: https://doi.org/10.1080/17476348.2020.1743181
Accepted author version posted online: 13
Mar 2020.
Published online: 17 Mar 2020.
Submit your article to this journal
Article views: 97
View related articles
View Crossmark data

PERSPECTIVE
Artificial intelligence and machine learning in respiratory medicine
Evgeni Mekov
a
, Marc Miravitlles
b
and Rosen Petkov
a
a
Medical Faculty, Department of Pulmonary Diseases, Medical University - Sofia, Sofia, Bulgaria;
b
Pneumology Department, Hospital Universitari
Vall d´Hebron/Vall d’Hebron Institut de Recerca, CIBER de Enfermedades Respiratorias (CIBERES), Barcelona, Spain
ABSTRACT
Introduction: The application of artificial intelligence (AI) and machine learning (ML) in medicine and in
particular in respiratory medicine is an increasingly relevant topic.
Areas covered: We aimed to identify and describe the studies published on the use of AI and ML in the
field of respiratory diseases. The string ‘(((pulmonary) OR respiratory)) AND ((artificial intelligence) OR
machine learning)’was used in PubMed as a search strategy. The majority of studies identified
corresponded to the area of chronic obstructive pulmonary disease (COPD), in particular to COPD
and chest computed tomography scans, interpretation of pulmonary function tests, exacerbations and
treatment. Another field of interest is the application of AI and ML to the diagnosis of interstitial lung
disease, and a few other studies were identified on the fields of mechanical ventilation, interpretation of
images on chest X-ray and diagnosis of bronchial asthma.
Expert opinion: ML may help to make clinical decisions but will not replace the physician completely.
Human errors in medicine are associated with large financial losses, and many of them could be
prevented with the help of AI and ML. AI is particularly useful in the absence of conclusive evidence
of decision-making.
ARTICLE HISTORY
Received 20 September 2019
Accepted 12 March 2020
KEYWORDS
Artificial intelligence;
machine learning;
respiratory diseases; COPD;
pulmonary fibrosis
1. Introduction
The application of artificial intelligence (AI) in medicine and in
particular in respiratory medicine is an increasingly relevant topic.
Machine learning (ML) is part of AI, where computers use statistical
methods for self-learning without being explicitly programmed [1].
Machine training is often likened to a ‘black box’–data is
entered on one side, the other being a solution, but the
process itself is incomprehensible (Figure 1). This is especially
true for the so-called ‘neural networks’, where data can
undergo complex transformations of many layers in the algo-
rithm, and the model could be too complex and work in
unpredictable ways. Some algorithms for ML are designed to
continue learning from newly introduced data (online train-
ing). Ensemble methods are combining statistics and ML by
using multiple learning algorithms to obtain better predictive
performance than could be obtained from any of the consti-
tuent learning algorithms alone. Unlike a statistical ensemble
in statistical mechanics, which is usually infinite, a machine
learning ensemble consists of only a concrete finite set of
alternative models but typically allows for a much more flex-
ible structure to exist among those alternatives.
On the one hand, this is an opportunity to increase accuracy
over time, but on the other, the risk of deviation from the gold
standard increases the possibility of learning from bad examples
when making incorrect decisions. The same risk exists also in the
presence of statistical errors or incorrect human data interpreta-
tion. Human error occurs randomly, and its effects are usually at
the local level. Errors in AI systems could be repeated over a long
period and affect all centers that use this technology while being
hard to find in practice.
AI studies are gaining popularity. In 2016, investment in
health projects involving AI is at the top of any other sector in
the global economy [2]. Interest in artificial intelligence in health-
care soared in 2019 with investors pouring 4 USD billion into the
sector which is up from nearly 2.7 USD billion invested in health-
care AI in 2018 [3]. In the field of respiratory medicine, several
studies have been published with promising results [4–16]
(Table 1).
Making an informed clinical judgment based on existing
data is the basis of evidence-based medicine [17]. Typically,
this task is solved using statistical methods that establish
patterns in the data and express them as mathematical equa-
tions (e.g. linear regression). Through the ML, artificial intelli-
gence can establish complex relationships that could not
easily be expressed through an equation. Neural networks,
for example, present the data through a large number of
interconnected neurons in a similar to the human brain way.
This allows the ML systems to approach the problem in
a similar to the clinician way –reaching motivated conclusions
by carefully analyzing the evidence. In contrast, however,
these systems could simultaneously monitor and process vir-
tually unlimited incoming data. Also, these systems can learn
from every new case and could process more cases within
minutes than a doctor will see throughout his practice. The
disadvantage of the method is its dependence on data quality
in terms of reliability and variability.
CONTACT Marc Miravitlles marcm@separ.es Pneumology Department, Hospital Universitari Vall d’hebron,P.Valld’Hebron 119-129, Barcelona 08035, Spain
EXPERT REVIEW OF RESPIRATORY MEDICINE
2020, VOL. 14, NO. 6, 559–564
https://doi.org/10.1080/17476348.2020.1743181
© 2020 Informa UK Limited, trading as Taylor & Francis Group

One of the relevant aspects of ML is the need to feed the
process with a huge amount of data to elaborate on the
analysis. There are different cohorts of patients followed pro-
spectively, but they usually involve hundreds or a few thou-
sands of patients. However, ML technology could deal with
larger datasets. One strategy developed to provide large data-
sets derives from the information that physicians write in
Electronic Health Records (EHR). In many medical institutions,
thousands and millions of EHR are written every day, this
information can be downloaded in an anonymized format
and analyzed using classical statistical methods to respond
to a specific research question [18,19]. However, more recently
there are initiatives developed as platforms for clinical deci-
sion support based on real-time dynamic exploitation of all
the information contained in EHRs.
To take advantage of the information contained in the EHRs, it
is necessary to combine computational skills with Natural
Language Processing (NLP), which is a subfield of linguistics,
computer science, information engineering, and artificial intelli-
gence specialized in processing and understanding text written in
natural language. This NLP could ‘read’and extract the requested
information on a large scale. One example of this AI methodology
has already been developed and is being applied in clinical prac-
tice and research [20]. All these new developments in AI and ML
will help to make clinical decisions, but will not replace the phy-
sician completely. Thus the ultimate responsibility for decisions is
inthehandsoftheclinician.
The string ‘(((pulmonary) OR respiratory)) AND ((artificial
intelligence) OR machine learning)’was used in PubMed as
a search strategy. The included studies represent the most
recent achievements or what we judged as having
a significant impact in this field to outline the examined topic.
2. Artificial intelligence and respiratory diseases
2.1. AI and COPD
COPD is a common, preventable and treatable disease that is
characterized by persistent respiratory symptoms and airflow
limitation that is due to airway and/or alveolar abnormalities
usually caused by significant exposure to noxious particles or
gases. Currently, COPD is the fourth leading cause of death in
the world. Globally, the COPD burden is projected to increase
in the coming decades because of continued exposure to
COPD risk factors and aging of the population. Spirometry is
the gold standard for COPD diagnosis.
2.1.1. COPD and chest CT
Artificial intelligence algorithms search for patterns in a set of
data that could be used to predict a clinical outcome or to
detect obstructive phenotypes [16]. The ML is successfully
used for automatic interpretation of pulmonary function
tests and differential diagnosis of obstructive pulmonary dis-
eases. CT is an advanced imaging technique for the detection
and characterization of emphysema and the assessment of
airway diseases. The neural network model is state-of-the-art
for obstructive pattern recognition in CT [16]. Thus, Gonzalez
et al. use chest CT on COPDGene and ECLIPSE participants to
determine if this methodology could be used to diagnose and
stage COPD, and to predict exacerbation and mortality [7]. In
the first stage, the authors trained the model using CT from
participants in COPDGene, then used the created algorithm in
1,000 nonoverlapping COPDGene patients and 1672 ECLIPSE
patients. The established concordance index is 0.856 for COPD
based on a prediction with a probability of >50% in the
COPDGene cohort. Also, about half of the participants are
correctly staged according to GOLD (based on FEV1) and
75% are staged in the correct or adjacent stage. The same
algorithm shows slightly inferior results in ECLIPSE –29%
correctly staged and 75% within the same or adjacent stage.
The ability to predict exacerbation has also been investigated.
COPDGene participants according to this model have a 2.15
times higher probability of exacerbation than those not at risk
(concordance index 0.64), but in ECLIPSE the model could not
identify patients with increased risk. The mortality estimate is
also good at COPDGene (concordance index 0.72), but not in
ECLIPSE. Overall, the use of AI shows promising results in the
diagnosis of obstructive pulmonary diseases.
2.1.2. COPD and PFT
Interpretation of Pulmonary Function Tests (PFT) for respira-
tory disease diagnosis is based on expert opinion, which
Article highlights
●Artificial intelligence could generate models incorporating huge
amounts of data that are incomprehensible to the physician
●Machine learning is part of artificial intelligence, where computers
use statistical methods for self-learning without being explicitly
programmed
●In many cases, machine learning would help to make clinical deci-
sions by a doctor, but would not replace him/her completely
●In the field of respiratory medicine, several studies focused on
obstructive conditions and pulmonary fibrosis with regards to diag-
nosis, staging, exacerbations, and survival.
●AI is particularly useful in the absence of conclusive evidence for
decision-making
●The availability of very large datasets and the increasing capability of
machine learning approaches may increase the clinical benefit and
minimize the patient risk
Figure 1. General scheme of machine learning.
560 E. MEKOV ET AL.

Table 1. Studies in respiratory medicine which include artificial intelligence.
First author, year Disease Methods Results
Ganzert S [4] ARDS 10 patients
4 measurement methods: LOW-FLOW, SLICE, SUPER-SYRINGE, SCASS;
CUBIST regression system
The inclusion of background features improves the accuracy of the models almost by one order
of magnitude. Normalization is beneficial for regression as well.
Lakhani P [5] Tuberculosis 1007 posteroanterior chest radiographs
AlexNet and GoogLeNet deep convolutional neural network architectures,
including pre-trained and untrained models
The AUCs of the pre-trained models were greater than that of the untrained models.
The best-performing ensemble model had an AUC of 0.99.
Kukreja S [6] Asthma A questionnaire, clinical data, and medical records
Machine learning approach
Back-propagation model, Context-Sensitive Auto-Associative Memory Neural
Network Model, C4.5 Algorithm, Bayesian Network, and Particle Swarm
Optimization
All algorithms have an accuracy of over 80%. 94.2% for the best algorithm.
González G [7] COPD 7,983 COPDGene participants for training, 1,000 nonoverlapping COPDGene
participants and 1,672 ECLIPSE participants for the evaluation
Convolutional neural network
Logistic regression (C statistic and the Hosmer-Lemeshow test) for COPD
diagnosis and ARDE prediction.
Cox regression (C index and the Greenwood-Nam-D’Agnostino test) for mortality.
COPDGene results:
C statistic for COPD diagnosis: 0.856
51.1% of participants were accurately staged and 74.95% were within one stage
C statistic for ARDE: 0.64
C statistic for mortality: 0.72
ECLIPSE results:
29.4% of participants were accurately staged and 74.6% were within one stage
C statistic for ARDE: 0.55
C statistic for mortality: 0.60
Fernandez-Granero
MA [8]
COPD Prediction of COPD exacerbations
Machine learning with non-knowledge based approach
Self-reported data by patients
Significant results were obtained for exacerbation prediction. The best accuracy was achieved
using a PNN classifier.
Fernandez-Granero
MA [9]
COPD 16 patients
Home telemonitoring of symptoms
Probabilistic Neural Network model
33/41 early detected exacerbations
mean 4.8 days prior beginning, 80.5% accuracy
Sanchez-Morillo
D[10]
COPD 16 patients
Electronic questionnaire via mobile app
Pattern recognition techniques
31/33 early detected exacerbations
mean 4.5 days before the onset of the exacerbation, accuracy 84.7%
Hardinge M [11] COPD 18 patients
algorithm with personalized physiological cutoffs, patient-reported symptoms,
and used medications
mobile application
40% of exacerbations had an alert signal during the three days before a patient starting
medication
Topalovic M [12] Respiratory
diseases
50 cases
Machine learning approach
AI correctly interprets all PFTs (100%) and made the correct diagnosis in 82% of cases
(p < 0.0001 for both versus respiratory specialists).
Walsh SLF [13] Fibrotic lung
diseases
1157 cases for training, validation, and testing Accuracy 76.4%; 92.7% of the diagnoses were being placed in the right or the adjacent
category. The algorithm distinguishes between UIP and non-UIP with comparable accuracy to
an imaging specialist (HR 2.88 versus 2.74).
Maldonado F [14] Idiopathic
pulmonary
fibrosis
55 patients
CALIPER
The algorithm shows predictive power in 2.4 years follow-up based on reticular abnormalities
(HR 1.91, p = 0.006), the severity of interstitial abnormalities (HR 1.70, p = 0.003), and
proportion of interstitial abnormalities (HR 1.52, p = 0.017).
Jacob J [15] Idiopathic
pulmonary
fibrosis
283 patients
CALIPER
A three-group staging system derived from this model was powerfully predictive of mortality
(2.23, p < 0.0001). CALIPER-derived parameters, in particular PVV, are more accurate
prognostically than traditional visual CT scores.
Paredes M [16]ICU patients 342 patients
AI-based model for mortality prediction
The machine learning model predicts with 78% accuracy the likelihood that a sepsis patient will
die after 30-days of ICU discharge
Abbreviations: AI –artificial intelligence, ARDE –acute respiratory disease event, ARDS –Acute respiratory distress syndrome, AUC –Area under the curve, CALIPER –Computer-Aided Lung Informatics for Pathology Evaluation
and Rating, COPD –chronic obstructive pulmonary disease, HR –hazard ratio, ICU –intensive care unit, PFT –pulmonary function test, PNN –Probabilistic neural network, PVV –pulmonary vessel volume, UIP –usual
interstitial pneumonia
EXPERT REVIEW OF RESPIRATORY MEDICINE 561

relies on model recognition and clinical context to identify
a specific disease. In one study, 120 pulmonologists from 16
European hospitals evaluated 50 cases consisting of PFTs
and clinical information (6000 independent interpretations)
[12]. The same information was also analyzed by an AI.
Respiratory specialists interpreted 74.4% of cases against
the gold standard (American Thoracic Society (ATS)/
European Respiratory Society (ERS) PFT interpretation guide-
lines, medical history, and additional studies) with
a variability (kappa = 0.67). The correct diagnosis was
madein44.6%ofthecases,alsowithhighvariability
(kappa = 0.35). AI correctly interprets all PFTs (100%) and
made the correct diagnosis in 82% of cases (p < 0.0001 for
both). The authors conclude that the interpretation of PFTs
by a pulmonologist is associated with high variability and
mistakes. AI provides a more accurate interpretation and
could be used as an aid in making decisions to improve
clinical practice.
2.1.3. COPD and exacerbation
Early detection of exacerbation is one of the main goals for
COPD patients. A six-month study attempted to create
a model for predicting them through ML [8]. An AI approach
without prior algorithm training was used through patient-
reported data using a multimodal tool during the remote
monitoring. In this way, the model learns itself only from
entered participants’data. All models have a good predictive
value regardless of the exacerbation definition used, with the
Probabilistic Neural Network (PNN) model being the most
accurate (89.3% accuracy compared to 84.7% in the K-means
model and 82.8% in radial basis function neural network).
The probabilistic neural network may predict the occur-
rence of exacerbation 4.8 days before its beginning through
home telemonitoring of symptoms in another study with an
accuracy of 80.5% and only 3% false-positive results [9].
However, these results should be interpreted with caution
because of the limited sample size (15 patients completed
the trial with 41 exacerbation episodes). The same authors in
2015 reported similar results –31 out of 33 exacerbations in
15 patients were early identified with an average of
4.5 ± 2.1 days before the onset of the exacerbation that was
considered the day of medical attendance [10]. This method
could help in the early detection of exacerbations which will
aid both physicians and patients.
Hardinge et al. in 2014 tested an algorithm with persona-
lized physiological cutoffs, patient-reported symptoms and
used medications (inhalers, antibiotics, and oral steroids)
through a mobile application (my mhealth) [11]. 40% of
exacerbations had an alert signal during the three days before
a patient starting a medication.
2.1.4. COPD and treatment
The treatment of chronic obstructive diseases is mainly
mediated by inhalation devices. Their use involves several
steps, and many patients make mistakes when using them.
Monitoring of the correct inhalation technique and compli-
ance are essential to improve the effectiveness of therapy.
Amiko Respiro Amiko Respiro’s Sense Technology (available
as a separate inhaler or optional device) tracks the compliance
and inhalation technique and could send feedback to
a physician. AI in the sensors allows patients to be reminded
of the dose administration and ways to improve the inhalation
technique. Available devices improve adherence to therapy by
180% in children (Adherium’s Hailie™) and 58–59% in adults
(Adherium’s Hailie™and Propeller Health).
2.2. AI and fibrosis
One study uses ML for classification of lung fibrosis diseases
using high-resolution CT against current guidelines (2011
ATS/ERS/Latin American Thoracic Association (ALAT) and
Fleischner Society for idiopathic pulmonary fibrosis) [13].
The accuracy of the algorithm in the first series of cases
was 76.4%, with 92.7% of the diagnoses being placed in
the right or the adjacent category. The required time to
estimate the 150 four-slices images (each image represents
a separate case from the second series) was 2.31 seconds.
The mean accuracy of radiologists in the second series was
70.7%, while the accuracy of the algorithm was 73.3%,
a better result than 66% (60/91) of the thoracic radiologists.
The algorithm distinguishes between UIP and non-UIP with
comparable accuracy to an imaging specialist (HR 2.88 ver-
sus 2.74). According to the authors, the assessment of high-
resolution CT by ML is an accessible, reproducible and
almost instantaneous method for classification of pulmonary
fibrosis with accuracy comparable to that of a radiologist.
Maldonado et al. used CALIPER (Computer-Aided Lung
Informatics for Pathology Evaluation and Rating) to analyze
and evaluate the severity of parenchymal pulmonary abnorm-
alities (honeycombing, reticular abnormalities, ground-glass
opacities, and emphysema) and progression over time (med-
ian 289 days) in high-resolution CT [14]. The algorithm showed
predictive power in 2.4 years follow-up based on reticular
abnormalities (HR 1.91, p = 0.006), the severity of interstitial
abnormalities (HR 1.70, p = 0.003), and proportion of intersti-
tial abnormalities (HR 1.52, p = 0.017). Besides, CALIPER could
estimate the pulmonary vessel volume (PVV). Instruments such
as CALIPER have the potential to improve staging in idiopathic
pulmonary fibrosis [15].
2.3. Other applications
One of the first studies in this field as early as 2002 compared
different methods for measuring pressure-volume curves in
mechanically ventilated patients with ARDS [4]. With the
help of the ML, the introduction of patient data improved
the results of statistical regression significantly, creating an
opportunity for individualized treatment.
AI could be used in case of disagreement between experts.
According to one study, AI correctly interprets chest X-ray and
establishes pulmonary tuberculosis with 95% sensitivity and 100%
specificity [5]. With the help of AI, the radiographs from the
individual centers could be interpretedbyacentralizedsystem.
Kukreja et al. develop and evaluate algorithms for the
diagnosis of bronchial asthma according to a comprehensive
questionnaire, clinical data and medical records [6]. The algo-
rithms proposed in this study have an accuracy of over 80%,
and with reliable data training, the model with automatic
562 E. MEKOV ET AL.

associative memory to the neural network algorithm reaches
an accuracy of over 90% with 1% inconclusive results and the
developed mobile applications reach 94.2% accuracy.
Human errors in medicine are associated with large finan-
cial losses, and many of them could be prevented with the
help of AI and ML. One study reported a 10-18% reduction in
mortality and reduced stay in an intensive care unit by an
average of 5.9 to 8.4 days when diuretics were used in an ML
model [16]. AI is particularly useful in the absence of conclu-
sive evidence for decision-making.
Access to relevant, high-quality data and gold standards for
testing and validation is essential to build association-based
models [21]. Also, multi-scale, integrated computational mod-
els with medically relevant outputs on the level of the indivi-
dual and focused on specific medical questions are needed.
For example, U-BIOPRED (Unbiased BIOmarkers in PREDiction
of respiratory disease outcomes) is a research project using
information and samples from adults and children to learn
more about different types of asthma to ensure better diag-
nosis and treatment for each person [22]. It is taking an
integrated approach to the study of severe asthma. Given
the complex pathophysiology, underlying different pheno-
types of respiratory disease, the integration of data at different
levels may provide causal insights.
3. Conclusion
The application of AI in medicine and in particular in respira-
tory medicine is an increasingly relevant topic. Machine learn-
ing is part of AI, where computers use statistical methods for
self-learning without being explicitly programmed. In 2016,
investment in health projects involving AI is at the top of
any other sector in the global economy. In many cases, ML
would help to make clinical decisions, but would not replace
the physician completely. Human errors in medicine are asso-
ciated with large financial losses, and many of them could be
prevented with the help of AI and ML. AI is particularly useful
in the absence of conclusive evidence for decision-making. In
the field of respiratory medicine, several studies have been
published with promising results. They mainly focus on
obstructive conditions and pulmonary fibrosis with regards
to diagnosis, staging, exacerbations, and survival.
4. Expert opinion
Despite the tremendous advance in many aspects of understand-
ing and managing COPD, there is also a lot to demand from the
future. Currently, personalized medicine in COPD is still evolving
with searching for distinguished phenotypes, biomarkers, prog-
nostic indices, etc. which should guide the decision-making pro-
cess in the individual patient. This process could be facilitated by
AI. AI could generate models incorporating huge amounts of data
that are incomprehensible to the physician. Multi-scale, integrated
computer models with medically relevant individual outcomes
and specific medical questions (mortality estimates, risk of exacer-
bation, evaluating and reshaping treatment based on risk factors/
disease characteristics) are the next steps in COPD management.
Improving healthcare largely depends on proper analysis and
interpretation of data. As data become more important in
improving healthcare, we believe clinicians and healthcare sys-
tems will require mechanisms to understand them.
Relevant, high-quality data and goldstandards for testing and
validation are essential to building machine learning algorithms.
The access to large scale real-world and randomized controlled
trial data will be of great help in this case. Simultaneously,
findable, accessible, interoperable, reusable (FAIR) models are
required that bridge association-based (machine learning, statis-
tical) and mechanism-based modeling.
Theavailabilityofverylargedatasetsandtheincreasingcap-
ability of machine learning approaches may increase the clinical
benefit and minimize the patient risk with the development of the
so-called ‘dynamic clinical decision support system’. They should
provide direct aid to clinical decision-making (such as dashboards,
decision trees, tables, charts) that present information in
a comprehensive, deployable form and, therefore, help clinicians
to integrate and prioritize multiple, diverse evidence.
Currently, despite the initial fascinating results, some draw-
backs should be noted. The most important one is the selection
of used features, which may affect the results. Choosing features
(parameters, predictors) is a natural process when making
a forecast. Feature selection improves the performance of an ML
algorithm and needs to be used on an independent dataset rather
than the entire training data. Using feature selection on the entire
training set may yield an improved model by the selected features
over other models being tested to get seemingly better results
when there are selection biases. Unfortunately, this is not possible
in most COPD ML studies that have access only to small datasets.
Second, many studies did not report the use of an independent
test dataset. This means data that were not used in the training
procedure. A missing independent test dataset could overesti-
mate the performance of the ML algorithms and should always
be a step in the ML model development and evaluation. Also,
some studies are population-specific, have small sample sizes and
will need external or local validation in order to be clinically useful.
Last but not least, AI provides multidisciplinary challenges. Despite
infrastructural and methodological tasks, clinical modelers are
required to work cooperatively with other specialists such as
statisticians and mathematicians.
AI or ML shouldn’t necessarily preclude the use of traditional
hypothesis-testing approaches such as regression analysis. For
sure, AI/ML (especially the newer generations like XGBoost)
have the best combination of prediction performance and
processing time compared to other algorithms. However, AI
could be used to uncover patterns in a hypothesis-neutral
way, by using the training dataset to learn a hypothesis and
the test dataset to evaluate it. However, findings could be used
to generate specific hypotheses based on clinical expertise.
Given the medicine is not a pure mathematic and the patients
are not numbers, these methods should be complementary
rather than mutually exclusive.
The next five years should be devoted to solving the afore-
mentioned tasks. After dealing with the infrastructural and
methodological issues and with a multidisciplinary approach,
a well-designed and trained with real-life and randomized-
controlled data AI algorithm could improve the current man-
agement related to exacerbation reduction, and improving
survival, lung function, and quality of life and have the poten-
tial to move the COPD management to a whole new level.
EXPERT REVIEW OF RESPIRATORY MEDICINE 563

Funding
This paper was not funded.
Declaration of interest
M Miravitlles has received speaker or consulting fees from AstraZeneca, Bial,
Boehringer Ingelheim, Chiesi, Cipla, CSL Behring, Laboratorios Esteve, Ferrer,
Gebro Pharma, GlaxoSmithKline, Grifols, Menarini, Mereo Biopharma, Novartis,
pH Pharma, Rovi, TEVA, Verona Pharma and Zambon, and research grants from
GlaxoSmithKline and Grifols. unrelated to this manuscript. E Mekov has
received speaker or consulting fees from AstraZeneca and Chiesi, unrelated
to this manuscript. R Petkov has received speaker or consulting fees from
AstraZeneca, Boehringer Ingelheim and Chiesi, unrelated to this manuscript.
The authors have no other relevant affiliations or financial involvement with
any organization or entity with a financial interest in or financial conflict with
the subject matter or materials discussed in the manuscript apart from those
disclosed.
Reviewer disclosures
Peer reviewers on this manuscript have no relevant financial or other
relationships to disclose.
ORCID
Marc Miravitlles http://orcid.org/0000-0002-9850-9520
References
Papers of special note have been highlighted as either of interest (•)orof
considerable interest (••) to readers.
1. The Lancet Respiratory Medicine. Opening the black box of
machine learning. Lancet Respir Med. 2018;6(11):801.
2. CB insights research healthcare remains the hottest AI category for
deals; 2017 [cited 2020 Mar 09]. Available from: https://www.cbinsights.
com/research/artificial-intelligence-healthcare-startups-investors/
3. Investors poured $4B into healthcare AI startups in 2019. [cited
2020 Mar 09]. Available from: https://www.fiercehealthcare.com/
tech/investors-poured-4b-into-healthcare-ai-startups-2019
4. Ganzert S, Guttmann J, Kersting K, et al. Analysis of respiratory
pressure-volume curves in intensive care medicine using inductive
machine learning. Artif Intell Med. 2002;26(1–2):69–86..
5. Lakhani P, Sundaram B. Deep learning at chest radiography: auto-
mated classification of pulmonary tuberculosis by using convolu-
tional neural networks. Radiology. 2017;284(2):574–582.
6. Kukreja S A comprehensive study on the applications of artificial intelli-
gence for the medical diagnosis and prognosis of asthma. [cited 2019
Jul 16]. Available from SSRN: https://ssrn.com/abstract=3081746
7. González G, Ash SY, Vegas-Sánchez-Ferrero G, et al. Disease staging
and prognosis in smokers using deep learning in chest computed
tomography. Am J Respir Crit Care Med. 2018;197(2):193–203..
8. Fernandez-Granero MA, Sanchez-Morillo D, Lopez-Gordo MA, et al.
A machine learning approach to prediction of exacerbations of
chronic obstructive pulmonary disease. In: Ferrández Vicente J,
Álvarez- Sánchez J, de la Paz López F, et al., editors. Artificial
computation in biology and medicine, 305–311. Springer
International Publishing Switzerland: IWINAC; 2015.https://www.
springer.com/gp/book/9783319189130
•- mobile app predicts exacerbation
9. Fernandez-Granero MA, Sanchez-Morillo D, Leon-Jimenez A, et al.
Automatic prediction of chronic obstructive pulmonary disease
exacerbations through home telemonitoring of symptoms.
Biomed Mater Eng. 2014;24(6):3825–3832.
10. Sanchez-Morillo D, Fernandez-Granero M, Leon A. Detecting COPD
exacerbations early using daily telemonitoring of symptoms and
k-means clustering: a pilot study. Med Biol Eng Comput. 2015;53
(5):441–451.
11. Hardinge M, Rutter H, Velardo C, et al. Using a mobile health
application to support self-management in chronic obstructive
pulmonary disease: a six-month cohort study. BMC Med Inform
Decis Mak. 2015;15:46.
12. Topalovic M, Das N, Burgel PR, et al., Artificial intelligence outper-
forms pulmonologists in the interpretation of pulmonary function
tests. Eur Respir J. 53(4): 1801660. 2019..
•• AI correctly interprets all PFTs (100%) and made the correct
diagnosis in 82% of cases.
13. Walsh SLF, Calandriello L, Silva M, et al. Deep learning for classify-
ing fibrotic lung disease on high-resolution computed tomogra-
phy: a case-cohort study. Lancet Respir Med. 2018;6(11):837–845..
•• - AI distinguishes between UIP and non-UIP with comparable
accuracy to an imaging specialist
14. Maldonado F, Moua T, Rajagopalan S, et al. Automated quantifica-
tion of radiological patterns predicts survival in idiopathic pulmon-
ary fibrosis. Eur Respir J. 2014;43:204–212.
15. Jacob J, Bartholmai BJ, Rajagopalan S, et al., Mortality prediction in
idiopathic pulmonary fibrosis: evaluation of computer-based CT
analysis with conventional severity measures. Eur Respir J. 49(1):
pii:1601011. 2017..
•AI predicts mortality in IPF patients.
16. Paredes M Can artificial intelligence help reduce human medical
error? Two examples from ICUs in the US and Peru; 2018 [cited
2019 Jul 16]. Available from https://techpolicyinstitute.org/wp660
content/uploads/2018/02/Paredes-Can-Artificial-Intelligence-help-
reduce-human661 medical-errors-DRAFT.pdf.
17. Buch V. Artificial intelligence in medicine: current trends and future
possibilities. Br J Gen Pract. 2018;68(668):143–144.
18. Monteagudo M, Roset M, Rodríguez-Blanco T, et al. Characteristics
of COPD patients initiating treatment with aclidinium or tiotropium
in primary care in Catalonia: a population-based study. Int J Chron
Obstruct Pulm Dis. 2017;12:1145–1152..
19. Barrecheguren M, Monteagudo M, Miravitlles M. Population-based
study of LAMA monotherapy effectiveness compared with LABA/
LAMA as initial treatment for COPD in primary care. NPJ Prim Care
Respir Med. 2018;28:36.
20. Hernandez Medrano I, Tello Guijarro J, Belda C, et al. Savana:
re-using electronic health records with artificial intelligence.
Int J Interact Multimedia Artif Intell. 2017;4(7):8–12.
21. Franssen F, Alter P, Bar N, et al. Personalized medicine for patients
with COPD: where are we? Int J COPD. 2019;14:1465–84.
22. [cited 2020 Feb 23]. Available from: https://www.europeanlung.
org/en/projects-and-research/projects/u-biopred/home
564 E. MEKOV ET AL.