Computational Phenotyping of Obstructive Airway Diseases: A Systematic Review

Introduction

Chronic obstructive airway diseases, such as asthma and COPD, are heterogeneous conditions that exhibit diverse clinical presentations due to a variety of endogenous and exogenous factors.^1,2 Obstructive airway diseases have distinct mechanistic pathways and heterogenous clinical presentations known as phenotype.³ Identification of specific phenotypes of airway diseases is important as this will help better to target therapies, personalize clinical interventions, and improve diagnostic accuracy.⁴

Over the past two decades, there has been an increase in the use of data-driven approaches in identifying phenotypes of chronic obstructive airway diseases.⁵ These approaches rely on unsupervised methods to extract latent patterns of the disease that are not known beforehand.⁶ This allows for the identification of disease subgroups that are more reflective of natural disease phenomena and that can guide clinical decision-making^1,6. However, studies employing these methods, and the resulting phenotypes have been challenging to compare, perhaps due to differences in participants’ profiles, study settings, phenotyping methods employed, and number and types of variables used.^2,6 To gain clear appreciation of the landscape of computational phenotyping of chronic obstructive airway diseases, a systematic synthesis of the underlying evidence is valuable. Through this, the methodological underpinning of studies can be ascertained, and the quality of evidence appraised, thus helping to identify potential research gaps in moving the field forward.

This review aimed at identifying, critically appraising, and synthesizing data from studies that have utilized computational approaches to phenotype chronic obstructive airway diseases in both children and adults. The review set out to characterize and compare the populations included in studies, assess and compare the criteria used to select participants, evaluate and compare the variables used to derive phenotypes of chronic airway diseases across studies, and assess the choices informing inclusion of variables. Additionally, the review described and compared the computational approaches used across studies and described and assess the number and characteristics of phenotypes derived across studies in terms of their clinical interpretation.

Methods

Protocol and Registration

We developed a protocol that outlined the review processes and methods before undertaking this work, which was registered in PROSPERO (CRD42020164898) and published.⁷

Eligibility Criteria

Table 1 shows the full information on inclusion and exclusion criteria of studies into the review based on aspects of study design, setting, outcome, method of phenotyping, participants’ age, study year and language.

Table 1 Inclusion and Exclusion Criteria

Information Source

To identify relevant studies for the review, we searched PubMed, Embase, Web of Science, Scopus, and Google Scholar. For unpublished materials, such as conference proceedings, we searched databases of proceedings of conferences and databases of the literature, such as Open Grey. We also contacted experts in the field to request for any paper that was missed from our database searches. Finally, we screened the reference lists of included studies to identify any additional papers.

Search Strategy

We developed search strategies for all the databases to identify relevant studies for the review. The search strategies (Supplementary file 1) were first developed in PubMed and then adapted in searching the other databases.

Study Records

Data Management and Selection Process

The search results from the different databases were exported to EndNote for screening. The first stage of the literature review involved removal of duplicates from the database searches; then, we performed title and abstract screening. Two reviewers independently screened the studies on the basis of the review inclusion and exclusion criteria; any discrepancies were resolved by discussion, or a third reviewer arbitrated if a consensus was not reached. The final stage involved full-text screening of the studies potentially meeting the eligibility criteria on the basis of the titles and abstracts. We documented the screening process using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart.⁸

Data Collection Process

Reviewers, in pairs, independently extracted relevant data from included studies onto a data extraction form that was developed for the review; any discrepancies were resolved by discussion, or a third reviewer arbitrated if a consensus was not reached. We developed a data extraction form specifically designed for this review. The form was initially piloted on three included studies; any amendment was undertaken prior to using the form on all included studies.

Data Items

Information on the following data items were collected from included studies into the data extraction form: general information (author’s name, publication year and study time, aim of the study, and data source); information describing populations characteristics (population size, recruitment characteristics, sample size, children/adults, inclusion and exclusion criteria); type of chronic obstructive airway disease and how was the outcomes defined; information about the variables selected for phenotyping (number and description of variables, rational of selection, variable measurement and definition); type and features of computational approach used; and information of the derived phenotypes (number of phenotypes, characteristics of each phenotype, and clinical interpretation).

Outcome and Prioritization

We included studies focusing on computational phenotyping of the following chronic obstructive airway diseases:

Asthma

COPD and asthma and COPD overlap

Rhinitis

Emphysema

Quality Assessment of Included Studies

We appraised the general quality of included studies using an in-house developed checklist. Since, to our knowledge, there are no standard tools for assessing the quality of studies on computational disease phenotyping, we developed a checklist that enabled us to assess the quality of reporting specific aspects of the studies as they relate to performing a computational phenotyping. The aspects assessed were subjects’ selection and inclusion in the phenotyping sample; missing data; outcome definition; variables included for the phenotyping; clinical and scientific relevance of the derived phenotypes; and reproducibility of the phenotyping process. To evaluate reproducibility, we examined aspects such as the disclosure of detailed information on methods used for phenotyping, computational aspects of data processing, and the utilization of software and tools for reproducible research frameworks. Detailed information and form of quality assessment can be found in the supplementary material.

Data Synthesis

Data was narratively synthesized. We used tables and figures to summarize the results and different aspects of the studies, including study characteristics, methods of phenotyping, variables considered in phenotyping, counts of number of phenotypes, and description, as well as the results of the quality assessment.

Deviation from the Study Protocol

None of the identified studies addressed emphysema as an outcome to be phenotyped. Instead, emphysematous changes as features of phenotyping obstructive airway diseases were reported within studies on COPD. Further, studies that included subjects with asthma and COPD overlap (ACO) were reported as separate outcome.

Results

Study Selection

A total of 3320 records were identified from the literature searches. After removal of duplicates, 2619 records were screened by title and/or abstract, of which 2460 records were excluded for not being eligible. A total of 159 records were considered for full-text review, of which 39 were excluded for different reasons, summarized in Table S1 in the supplementary material. Finally, 120 studies were included in this review analysis. Figure 1 shows the screening and selection of studies for this review.

Figure 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram illustrating the studies’ selection process. Note: This figure was adapted from Page, Matthew J., et al. ‘The PRISMA 2020 statement: an updated guideline for reporting systematic reviews.’ bmj 372 (2021). https://doi.org/10.1136/bmj.n71.

Study Characteristics

Asthma

A total of 60 studies were on asthma.^9–68 The average number of subjects included in these studies was 1251, ranging from 50 to 9651 participants per study. The majority of studies were conducted among adults (n = 31)^{10,12–14,18–20,24,25,27,29–32,35–38,42–44,46–48,50,52,56,59–62} and the remaining (n = 9)^{21–23,28,33,34,40,45,51} in children, with remaining in mixed sample. Most studies were of cohort design (n = 30),^{12,14,15,17,20,23,24,28,30–33,36–38,41,42,47,50,51,54–60,62,65,66} while the rest were mostly cross-sectional (n = 17).^{10,11,13,16,18,21,22,25,27,29,33,35,40,43,45,46,48} Most studies were conducted in a clinical setting (n = 33)^{10–13,15,19–22,24,25,31,33–38,42,43,45,50,52–58,60–62,65} with patients variously recruited from hospitals, pulmonary rehabilitation centers, and primary or tertiary care respiratory or general clinics. Studies with subjects selected from general population were 14,^{14,18,23,32,36,40,41,44,46,47,51,59,63,68} while 8^{19,28,29,31,35,40,44,53} studies did not report on the source of their participants. Full information on characteristics of studies in children and adults using unsupervised computational methods to phenotype asthma is presented in Table 2.

Table 2 Characteristics of Studies in Children and Adults Using Unsupervised Computational Methods to Phenotype Asthma and Severe Asthma

Severe Asthma

A total of 19 studies^69–87 were on severe asthma. The average number of participants included in each study was 230, ranging from 40 to 1424 subjects per study. Most studies were conducted in a clinical setting (n = 17).^{69–72,74–77,79–87} One study⁷⁸ in a general population setting and another study without a clear indication of setting.⁷³ Most were cohort studies (n = 11),^{71,74–76,78–81,86,87} while the remaining were cross-sectional studies. Characteristics of studies in children and adults using unsupervised computational methods to phenotype severe asthma are presented in Table 2.

COPD

A total of 28 studies^4,88–114 were on COPD. The average number of subjects per study was 5218, ranging from 46 to 104143 subjects per study. Most studies were conducted within a clinical setting (n = 17),^{4,88–93,95,97,98,100,107,109–113} with cohort studies^{88,89,92–96,98,100,101,109–111,114} being the most reported study design (n = 14), while the second common were of cross-sectional design (n = 7).^{4,90,91,97,107,112,113} Full information on characteristics of studies using unsupervised computational methods to phenotype COPD and ACO is given in Table 3.

Table 3 Characteristics of Studies Using Unsupervised Computational Methods to Phenotype COPD

Asthma and COPD Overlap (ACO)

Four of the included studies were on asthma and COPD overlap. The average number of participants included in these studies was 255, ranging from 47 to 435 participants per study. All were cross-sectional studies. Three studies were conducted in a clinical setting,^115–117 while one was conducted in a general population setting.¹¹⁸ Characteristics of studies in adults using unsupervised computational methods to phenotype COPD and ACO are presented in Table 3.

Rhinitis

A total of 9 studies were on rhinitis.^119–127 The average number of participants included in each study was 516, ranging from 115 to 1831 participants per study. Most studies were conducted in a clinical setting (n = 6),^{119–121,123,126,127} while three studies^122,124,125 were conducted in the general population. Most were cohort studies (n = 7),^{119,121,123–127} one was cross-sectional,¹²⁰ and one case–control study.¹²² Characteristics of studies using unsupervised computational methods to phenotype rhinitis are presented in Table 4.

Table 4 Characteristics of Studies Using Unsupervised Computational Methods to Phenotype Rhinitis

Phenotypes of Respiratory Diseases

Asthma

In total, 251 phenotypes were reported in studies on asthma with considerable degree of overlap between them. In characterizing asthma phenotypes, atopy was the most common feature included in most studies,^{13–15,18,20,22,28,30,31,33,35,37,38,40,41,43,44,47,48,53,55,57,59–62,65,68} resulting in differentiation of atopic and non-atopic asthma phenotypes (reported in 29 studies and featured in 100 of the reported asthma phenotypes). Atopic status was defined mostly based on skin prick test, serum IgE levels or subjects’ report of familial atopy. Atopic asthma phenotype was reported in 28 studies,^{13–15,18,20,21,28,30,31,33,35,37,38,40,41,43,44,47,48,53,55,57,59–62,65,68} while non-atopic asthma was reported in 22 studies.^{14,15,18,20–22,30,35–38,40,41,43,44,53,55,57,59,61,65,68} The second feature was lung function measures, which featured in 85 of the reported phenotypes and was considered in 30 studies.^{10,14–16,18–22,24,27,28,30,33,36–38,40–44,50,52,55,60–62,65,68} Time at asthma onset featured in 74 phenotypes and reported in 27 studies.^{11,13,15,18,21–23,28,30,31,35–38,43,44,48,50,52,53,55,59–61,65} The definition of early and late onset asthma varied among different studies. When studying both children and adolescents, asthma that developed during childhood and adolescence was referred to as early onset while adulthood developed asthma as late-onset asthma. However, when examining only adults or children, researchers measured the average age at which asthma onset and the standard deviation across different phenotypes. In these cases, the terms early and late onset asthma were defined differently, with the groups having younger individuals labeled as early onset and those with older individuals classified as late-onset asthma. Early onset asthma phenotype was reported in 19 studies^{15,18,21,22,28,35,37,38,43,44,48,50,52,53,55,59–61,65} while late – onset asthma was reported in also 18 studies.^{11,13,18,21–23,30,31,35,37,38,43,50,53,55,59,60,65} Level of asthma control was also a commonly reported feature, occurring in 45 of phenotypes and reported in 17 studies.^{11,13,18,21,24,27,33,42–46,48,50,53,56,60} Well-controlled asthma phenotype was reported in 11 studies,^{21,24,42–46,48,53,56,60} while uncontrolled asthma was reported in 16 studies.^{11,13,18,21,24,27,33,42–44,46,48,50,53,56,60}

Sex featured in 65 of the reported phenotypes. Female asthma phenotype was reported in 20 studies,^{10,11,13,16,18,20,21,30,31,35,42,43,45,48,52,53,55,60,61,63} while male asthma phenotype featured in 19 studies.^{10,11,13,18,20,22,28,30,40,42,43,45,47,52,53,55,60,61,63} Eleven studies reported on obesity-related asthma phenotypes.^{10,11,13,19,30,31,33,43,48,56,60}

Disease activity was characterized variously across asthma studies, based on either symptoms’ activity or disease severity. Frequency of symptoms and rate of exacerbation featured phenotypes of high or low symptoms’ activity, while disease severity defined using standard criteria of asthma severity characterized phenotypes of mild, moderate or severe asthma. Severe asthma phenotypes as indicated by investigators were reported in 12 studies,^{10,16,18,20–22,28,34,35,48,56,67} while 20 studies^{10,14,15,18,21–24,33,41–43,50,55,57,59,60,62,66,68} reported on asthma phenotypes with high symptoms or exacerbation rates. Across identified studies, labeling a phenotype as severe was not entirely based on standard GINA criteria for defining severe asthma or physician decision, although some studies applied such approach.^21,33,34,56 Otherwise, most investigators identified severity of phenotypes based on symptom frequency, need for high dosage of treatment and disease impairment of daily life, with no clear reporting on how severity was defined.^{10,16,20,28,35,48,67}

Inflammation was considered in deriving asthma phenotypes using different indicators like inflammatory cell counts in peripheral or sputum induced samples, fractional exhaled nitric oxide (FeNO),^10,18,44 or measure of inflammatory cytokines.⁵⁸ A total of 36 phenotypes were described based on high or low levels of eosinophilic inflammatory cells in sputum or peripheral blood. Some of those were reported in 17 studies^{10,11,15,16,20,21,30,33,40,41,43,44,48,50,52,58,65} as asthma with high eosinophilia. Variants of neutrophilic asthma phenotypes, in turn, were less commonly reported in 10 studies.^{11,13,30,33,42–44,50,52,58} See full results on number of derived phenotypes and their descriptions for studies on asthma and severe asthma in Table 5.

Table 5 Number of Derived Phenotypes and Their Descriptions for Studies on Asthma

Severe Asthma

The total number of reported severe asthma phenotypes was 61 with considerable degree of overlap between them. The most reported features that differentiated severe asthma phenotypes were atopy, featuring in 28 phenotypes; age at disease onset, featuring in 25 phenotypes; treatment defined as medication dosage or treatment step; inflammation measures, featuring in 14 phenotypes; disease activity as frequency of symptoms and exacerbations, featuring in 14 phenotypes; and age and sex that featured 13 phenotypes.

Regarding allergic status and time at disease onset, 10 studies^{72,74–76,78–81,86,87} reported phenotypes of atopic severe asthma, while non-atopic severe asthma phenotypes were reported by 8 studies.^{74–76,78–80,86,87} Early onset severe asthma phenotypes were reported in 8 studies,^{70,72,74,75,78,79,86,87} while late-onset severe asthma phenotype variants were reported in 7 studies.^{70,72,74,78,79,86,87} Defining age of disease onset in most studies was based on measuring the mean and standard deviation of age at disease onset and comparing phenotypes.^{72,74,79,81,87} Only one study defined more than 12 years as cutoff for late onset.⁷⁸

Disease activity in terms of symptoms differentiated phenotypes of severe asthma with high symptoms presentation in 6 studies,^{70,74,78,79,81,82} as well as in another 4 studies^70,74,79,81 with low symptoms. Based on medication usage, phenotypes of severe asthma that require extra higher treatment were described in 6 studies.^{69,74,76,78,79,81} Those were in form of extra higher doses of ICS, oral corticosteroids (OCS), additional controller, regular use of systematic CS, or more frequent need for OCS and short controllers. In turn, lower to more moderate medication usage or requirement that was reported in 5 studies.^{69,74,76,79,81} Although spirometry measures were not as commonly reported as other indicators of disease activity, highly obstructed variants of severe asthma phenotypes were reported in 7 reports,^{72,74,75,78,79,86,87} while moderate to mild obstructed severe asthma in 5 reports.^{69,72,74,78,79}

For demographic characteristics, variants of female severe asthma phenotypes were described in 4 studies,^74,75,78,81 and male severe asthma phenotypes in similar count of records.^74,75,78,80 Elderly related variants of severe asthma phenotypes were described in 4 studies,^74,75,78,81 and young age severe asthma phenotypes in 3 records.^75,78,82 Obesity-related variants of severe asthma phenotypes were reported in 5 studies.^{70,73,74,78,79} See full results on number of derived phenotypes and their descriptions for studies on severe asthma in Table 6.

Table 6 Number of Derived Phenotypes and Their Descriptions for Studies on Severe Asthma

COPD

The total number of reported COPD phenotypes was 57. The most reported feature for defining COPD phenotypes was lung function measured by spirometry that differentiated 44 phenotypes. Other commonly reported features were age, featuring in 26 phenotypes; symptoms and frequency of exacerbations, featuring in 24 phenotypes; sex, featuring in 17 phenotypes; and cardiovascular, metabolic, and psychiatric comorbidities, featuring in 14 −17 phenotypes.

Based on spirometry lung function measures, COPD phenotypes were classified as mild, moderate, or severely obstructed disease. Severe to moderately obstructed phenotypes of COPD were reported in 10 studies,^{4,88,91,93,95–97,102,105,109,111,112} while 5 studies^{91,93,96,102,112} reported mild obstructed COPD phenotypes. Other measures of lung function used for deriving COPD phenotypes were measures of accompanying emphysematous changes like lung diffusion capacity for carbon monoxide (DLCO),⁸⁸ computed tomography (CT) measure of lung density and airway wall thickness.^4,90,97,105 The latter identified COPD phenotypes with high, moderate to low emphysematous changes in 4 studies.^4,90,97,105

Demographic and social characteristics like age, sex, body mass index and smoking were also used to define COPD phenotypes. Elderly related COPD phenotype was reported in 9 COPD studies,^{88,91–93,95,96,98,102,113} while 7 studies^{88,91,93,95,96,98,113} described COPD phenotypes that were characterized by young age. Variants of female-related COPD phenotypes were reported in 3 records,^88,91,97 while male sex-related COPD was reported in 4 studies.^88,91,96,102 Both over- and underweight were associated with COPD phenotypes when considering BMI. Obesity-related COPD was reported in two studies by Burgel et al^4,91 while under or low weight-related COPD phenotypes were reported in 5 studies.^{4,93,95,111,113} Heavy, persistent, high rate or long duration smoking-related COPD phenotypes were reported in 3 studies,^91,97,105 while 2 studies reported low smoking-related COPD phenotypes.^91,102

Disease activity/severity was characterized in studies of COPD phenotyping variously using frequency of symptoms and exacerbations and level of treatment. COPD phenotypes with high frequency of symptoms and exacerbations were reported in 11 studies of COPD,^{4,91,93,95,97,100,102,107,109,111,112} while COPD phenotypes with low symptoms in 8 studies.^{4,91,93,97,100,102,107,112} Four studies^91,93,97,105 reported on COPD phenotypes with utilization of high treatment doses and 2 others with low dosage treatment.^91,97

Concerning comorbidities, the mostly reported ones to differentiate COPD phenotypes were cardiovascular diseases and diabetes and metabolic diseases,^{4,92,94,95,98} together with depression and anxiety.^{91,94,98,100,107} Additionally, features considered in characterizing COPD phenotypes were disease impairment on physical and daily activity, respiratory health, quality of life and mortality. COPD phenotypes with impaired quality of life were reported in 3 studies,^97,109,113 while high mortality-related COPD phenotypes were reported in 5 studies.^{92,93,96,112,113} See full results on number of derived phenotypes and their descriptions for studies on COPD in Table 7.

Table 7 Number of Derived Phenotypes and Their Descriptions for Studies on COPD and Asthma COPD Overlap (ACO)

Asthma and COPD Overlap (ACO)

A total of 21 phenotypes of ACO were identified. The most reported features considered for differentiating ACO phenotypes were smoking status, which identified 7 phenotypes; inflammation status which identified 9 phenotypes; atopy that identified 7 phenotypes; spirometry measures identifying 5 phenotypes and disease activity/severity as per symptoms identifying 5 phenotypes.

Regarding socio-demographic aspect, smoking-related ACO phenotypes were reported in two studies,^116,118 along with female ACO phenotype and obesity-related ACO,¹¹⁵ each one record. Lung function measures featured a highly obstructed ACO phenotype that was reported in two studies^115,116 and a high symptom phenotype of ACO was reported in 1 study.¹¹⁵ With respect to inflammation status, eosinophilic variants of ACO were reported in one study,¹¹⁵ as well as neutrophilic ACO phenotype.¹¹⁵

For other disease characteristics, early onset ACO phenotypes were reported in one study,¹¹⁸ while 3 records reported a variant of atopic ACO phenotype.^115,116,118 See full results on number of derived phenotypes and their descriptions for studies on COPD and ACO in Table 7.

Rhinitis

The total number of reported rhinitis phenotypes was 45. The most considered features for differentiating phenotypes of rhinitis were sex, which featured in 19 phenotypes; disease severity, which featured in 18 phenotypes; impairment on quality of life, which featured in 14 phenotypes and disease activity per symptoms that featured 10 phenotypes.

Considering socio-demographic characteristics, sex, age, and socio-economic status (SES) identified several rhinitis phenotypes. Variants of female-related rhinitis as well as male-related phenotypes of rhinitis were reported in near half of the reports (n = 5).^{119–121,124,126} Phenotypes of old age-related rhinitis were reported in 2 studies,^121,126 as well as young age-related ones.¹²⁶ SES featured phenotypes of high and low SES-related rhinitis which was reported by Lee et al.¹²⁵ Alcohol intake further identified high intake-related phenotypes of rhinitis that was reported by Soler et al.¹²⁶

Disease activity in terms of frequency of symptoms, classification of disease based on severity status as well as medication intake were commonly used to differentiate rhinitis phenotypes. High symptom phenotypes of rhinitis were reported in 3 studies.^121,123,125 Severe rhinitis phenotypes, in turn, were reported in 3 studies,^120,121,123 while Lee et al¹²⁵ reported rhinitis phenotypes which require high treatment doses.

Measures of airways or lung function that were used to feature rhinitis included CT scanning diagnostics of rhinitis, endoscopy score, and FeNO, in addition to spirometry, bronchodilator reversibility, bronchial hyperresponsiveness in subjects with accompanying asthma.^124,125 Two records described rhinitis phenotypes with highly obstructed airways.^124,125 Rhinitis with high endoscopic and CT score were reported by two studies.^119,126 One study reported on rhinitis phenotypes among asthmatics that is characterized by high inflammation indicated by FeNO.¹²⁴ Among asthmatic with rhinitis, phenotypes of rhinitis with low to moderate BDR and BHR were also reported.^124,125

Based on disease characteristics like time of onset and seasonality, atopy status and accompanying nasal polyposis, both early and late onset variants of rhinitis phenotypes were reported in two studies,^121,124 while seasonal rhinitis and accompanying nasal polyposis phenotypes were reported by in the same count of studies.^122,124 Variants of atopic rhinitis were reported in 3 studies,^123–125 while polysensitization rhinitis phenotypes were also reported in 3 studies.^120,122,123

The aspect of disease impairment on QOL and comorbidities was also frequently considered in featuring phenotypes of rhinitis. Phenotypes of rhinitis with impaired QOL were reported in 3 studies.^121,126,127 Rhinitis phenotypes with related comorbidities of depression, fibromyalgia, diabetes, and dermatitis were reported in one study by Soler et al.¹²⁶ See full results on number of derived phenotypes and their descriptions for studies on rhinitis in Table 8.

Table 8 Number of Derived Phenotypes and Their Descriptions for Studies on Rhinitis

Methods of Phenotyping

Various methods of unsupervised computational phenotyping of respiratory diseases were used across the reported studies (Figure 2). The most frequently implemented and reported unsupervised approaches for phenotyping of chronic airway diseases were hierarchical and non-hierarchical clustering^{10,32,38,40,46,54,70,81,82,90,95,101,110–112,114,117} with some records (n = 19) reported the implementation of the two approaches in the same study.^{17,19,21,24,35,38,45,47,62,75,76,79,80,86,93,100,107,113} In addition, latent class modelling^{10,13–16,20,23,31,41,42,44,48,50,58–60,123,125} was also frequently used. Other non-model-based methods of dimensionality reduction, such as factor analysis, principal component analysis, discriminant analysis and multiple correspondence analysis were also reported as methods for deriving phenotypes, albeit less frequently. Over years, hierarchical and non-hierarchical clustering were common particularly between 2010 and 2018. However, between 2015 and later, there was an increase in the use of other methods such as mixture-based model,^69,126 structural equation modelling,⁸⁸ and factor analysis with latent class modeling.¹²⁸ Figure 2 shows the count of studies reporting the methods applied for phenotyping.

Figure 2 Number of studies using each unsupervised phenotyping method for each respiratory outcome.

Quality Assessment of the Included Studies

Overall, the comprehensiveness of variables included in deriving the phenotypes was the best quality aspect reported in majority of studies on asthma,^{10,11,14–16,18,20,21,24,28,30,31,33,34,41,43–45,47,48,50,56,59–61,65,66} COPD,^{4,88,91–93,95,96,101,102,107,109–111,114} severe asthma,^{72,78,82,86,87} ACO^116–118 and rhinitis.^{120,122,124–126} Random sampling of study subjects, however, was less frequently performed among studies on asthma (n = 8)^{23,32,41,46,51,59–61}, severe asthma (n = 2),^78,81 COPD (n = 1),¹⁰⁵ ACO (n = 2)^117,118 and rhinitis (n = 1).¹²⁵ Majority of studies excluded subjects based on either clinical, social, or demographic characteristics. With respect to method of outcome definition, the most reported approach was usage of physician diagnosis assisted by clinical and biomarkers, which was reported in 33 of studies on asthma,^{10,12,13,15,16,18,21,24,27,30,31,33–38,40,42,44,46,48,52,54–63} 16 of studies on COPD^{4,88–91,93–95,97,101,102,109–112,114}, 12 of studies on severe asthma,^{69,71–73,75,76,78,80–82,86,87} four of studies on rhinitis,^{119,120,126,127} two of studies on ACO.^116,117 Overall reporting on reproducibility practices was uncommon, including how investigators handled noise and variation in data, rationale for selecting statistical methods for phenotyping, visualization techniques, and utilization of available tools for implementing reproducibility. With respect to clinical, biological or scientific relevance of the derived phenotypes, most studies reported on this aspect: 26 studies on asthma;^{10–24,27,28,30–37} 18 studies on COPD;^{4,88,91–98,100–102,107,109,110,112,113} 11 studies on severe asthma;^{69–73,75,76,78,81,87} three studies^115,116,118 on ACO and nine studies on rhinitis.^119–127 Full information on the quality assessment results can be found in Figure 3 and Table S2 in Supplementary Material.

Figure 3 Quality assessment items reporting for studies on asthma, severe asthma, COPD, asthma and COPD and rhinitis.

Endotypes and Phenotypes of Airways Diseases

Efforts to define phenotypic subtypes of airway diseases involved utilizing various approaches, including the assessment of serum and sputum-induced inflammatory cells as well as other biomarkers associated with inflammation and related processes. These biomarkers include cytokines, airway-inducible inflammatory mediators, and the composition of the airway microbiome.

Mastalerz et al⁴⁸ and Liang et al⁴² conducted studies investigating the role of airway-induced pro- and anti-inflammatory lipid eicosanoid mediators in asthma. In their research, Mastalerz et al⁴⁸ identified three distinct asthma phenotypes based on high levels of anti-inflammatory mediators. One phenotype exhibited chronic rhinosinusitis (CRS), good control, and mixed inflammation. The second phenotype showed atopy, no CRS, good control, and mixed inflammation. The third phenotype had poor control, aspirin sensitivity, and eosinophilia. Additionally, Mastalerz et al⁴⁸ identified a phenotype characterized by high levels of anti-inflammatory mediators among obese women with early-onset, atopic, and severe asthma.

Liang et al⁴² derived a phenotype that is characterized by a mixture of pro- and anti-inflammatory mediators. This phenotype exhibited low basophils, high functional activity, and poor control. Liang et al⁴² also identified a phenotype with low basophils, high anti-inflammatory mediators, and low control. Similarly, Caljwaska et al¹⁶ utilized induced sputum supernatants for phenotyping of non-steroidal anti-inflammatory drug-exacerbated respiratory disease (NERD). Seys et al⁵⁸ focused on the expression of inflammatory cytokines in airway sections to further subtype asthma patients. Their work reported an unexpected pattern of cytokine predominance among Th2-high asthmatics. It, further, identified subclusters with high levels of IL-5, IL-10, IL-25, and IL-17, associated with low lung function, high eosinophils, neutrophils, and fractional exhaled nitric oxide (FeNO). Another cluster consisted of IL-5 and/or IL-10 high asthmatics, while a separate cluster showed high levels of IL-6. A fifth cluster exhibited a normal pattern of high Th2 cytokines but with high eosinophils and low neutrophils. These findings indicate further heterogeneity among individuals with Th2-high asthma. However, this phenotyping approach did not include other clinical parameters such as symptoms, lung function, and outcomes. Instead, the derived clusters were modeled against clinical outcomes for further evaluation.

Nagasaki et al⁵⁰ identified phenotypes characterized by high serum periostin. One phenotype exhibited high eosinophilia, early-onset disease with good control, while the other phenotype showed high periostin, mixed inflammation, severe disease with poor control, high IL-6, mixed inflammation, and multimorbidity.

In the context of COPD, Bafadhel et al⁸⁹ identified five distinct phenotypes of COPD exacerbations based on biological biomarkers. These phenotypes were characterized by different predominant factors, including bacteria, viruses, eosinophils, paucigranulocytemia, and elevated levels of Sputum IL-1b and serum CXCL10. Britani et al⁹⁰ employed exhaled breath condensate analysis to assess inflammatory biomarkers in COPD, revealing unique metabolite profiles such as the low proline phenotype and high serine, valine, lysine, acetate, alanine, isoleucine, and other metabolite phenotypes.

In severe asthma, Diver et al⁷¹ focused on airway microbiology and identified clusters of severe asthma with varying levels of Haemophilus and Moraxella sputum communities, as well as different ratios of Gammaproteobacteria (G) to Firmicutes (F). Gomez and colleagues⁷⁴ explored severe asthma phenotypes characterized by varying levels of the chitinase-like protein YKL-40 as an inflammatory mediator which was inversely associated with disease severity, control, and treatment response.

An endotype of high eosinophilic chronic rhinosinusitis with high nasal polyposis and low CT and endoscopy score was identified by Adnane et al,¹¹⁹ while Nakayama et al¹²⁷ reported a high eosinophil and basophil rhinitis that is characterized by comorbid asthma, high symptoms, high CT and endoscopy score. Low eosinophilic rhinitis reported by Adnane et al¹¹⁹ was variants among male subjects, and others among females but with high CRSnP and CT and endoscopy score. Nakayama and colleagues¹²⁷ variants of low eosinophilic rhinitis were non-differential in symptoms, endoscopy and CT score.

Discussion

Summary of Key Findings

Our review reveals a wide variation in the phenotypes derived across all obstructive airway diseases investigated, both in children and adults, as well as variations in the methods used for phenotyping, study participants and population settings from where they have been recruited, and variables included in deriving the phenotypes. For asthma, the most reported phenotypes related to atopic status, time at disease onset, sex differences, disease symptomatology, and severity. For severe asthma, lung function measures, atopic status and age at disease onset were the most characterizing features for defining phenotypes. COPD phenotypes were mostly characterized by lung function measures, as well as accompanying comorbidities, disease impairment on daily activity, and mortality and smoking status. Phenotypes of asthma and COPD overlap were mostly defined by smoking status, lung function measures, inflammation, and disease activity. Phenotypes of rhinitis were mostly defined by sex, disease severity, disease impairment on life and seasonality. The most reported unsupervised methods used for phenotyping were hierarchical and non-hierarchical clustering, particularly between 2010 and 2018. However, between 2015 and later, there was an increase in the use of other methods such as mixture-based model,^69,126 structural equational modelling,⁸⁸ and factor analysis with latent class modeling.^{10,13–16,20,23,31,42,44,48,50,58–60,123,125}

Results Interpretation

Study Setting and Reporting of Airway Diseases Phenotypes

The majority of studies across outcomes enrolled participants from clinical settings, including from primary care centers,¹⁵ tertiary hospitals,³⁸ pulmonary rehabilitation centers,²⁴ outpatient clinics,⁵⁸ and emergency departments.¹¹ Asthma studies that recruited participants from the general population reported commonly observed asthma phenotypes, such as mild early-onset asthma and mild atopic asthma^14,18,41, which were comparable to those observed in clinical settings.^{12,15,21,22,31,34,35,37,38,50} The phenotyping of asthma within clinical settings enabled the derivation of phenotypes mostly defined by measures easily obtained in clinics, compared to other epidemiological risk factors. For instance, phenotypes that were characterized by high response to treatment; low/high treatment adherence,^12,56 as well as courses of disease progression measured by symptoms or lung function were reported.¹⁹ However, two studies conducted in general population settings reported asthma phenotypes of persistent and gradually improved wheeze or lung function,^23,37 as well as varying trajectories of attack rates progression and remission.⁵¹ The work by Moore et al⁷⁸ is the only reported attempt of phenotyping severe asthma using a sample from general population. The characterized phenotypes of early-onset, atopic asthma, and late onset, non-atopic, severe obesity-related asthma with high utilization of health care were frequently reported in other studies conducted in clinical setting.^{79,81,82,86,87}

Similarly, Kim et al¹⁰¹ was the only reported work on phenotyping COPD based on a general population sample, it reported phenotypes of highly obstructed older males with mild COPD; highly obstructed with high symptoms and mild COPD among non-smoking obese with near normal functions. However, similar characteristics of such were reported in COPD variants reported from clinical setting.^{4,88,91–93,100} Studies on COPD phenotypes derived from clinical setting, however, distinctly reported phenotypes of COPD with high to low emphysematous and air-entrapment changes,^{4,88,93,96,97,100,105} as well as comorbidities.^{4,91,92,94,98,110,114} Three studies phenotyped rhinitis based on samples from the general population.^122,124,125

Burte et al¹²² reported rhinitis phenotypes that are characterized by high nasal symptoms overall, high nasal symptoms through the year with low sensitization and seasonal spring nasal symptoms with high sensitization, while Kurukulaaratchy et al¹²⁴ and Lee et al¹²⁵ reported variants of non-atopic rhinitis and atopic rhinitis with normal lung function. Rhinitis phenotypes derived from clinical setting reported similar variants of atopy-related rhinitis,^120,123 in addition to others related to high eosinophilia and nasal polyposis,^119,127 as well as quality of life and comorbidities.^121,126 Regarding asthma and COPD overlap, the reported study from general population sample was Fingleton et al¹¹⁸ and revealed variants of smoking related, onset related, as well as atopic related ACO, which were comparable to similar reporting from clinical setting report.¹¹⁶

Overall, the characteristics of phenotypes derived from the general population settings were not always consistent with those derived from clinical settings. The observed variations could be due to variation in the types of variables readily available for measurement in general versus clinical settings. Further, presentation of patients of specific degree of disease severity, control, and treatment regimens varied across settings, and this could contribute to the variations between studies.

Aspects Considered for Phenotyping Airway Diseases

In delineating the various aspects that contribute to the characterization of airway disease phenotypes, researchers have usually included variables that encompass etiological or risk factors, indicators of disease manifestation, and treatment behaviors, and prognostic indicators. Notably, among studies undertaken within the clinical setting, a significant focus has been placed on the physiological aspects of the disease, particularly lung function measures as vital parameters in phenotyping asthma and COPD. However, in the phenotyping of asthma, bronchodilatation (reversibility) is often not considered, which can potentially help in distinguishing different phenotypes with obstructive characteristic. Similarly, in the context of COPD, the consideration of reversibility is also limited, despite its importance in differentiating COPD from other phenotypes, such as asthma-COPD overlap (ACO).¹²⁹

Furthermore, the inclusion of bronchoprovocation tests in asthma phenotyping was seldom reported. Boudier et al¹⁴ identified two phenotypes characterized by bronchial hyperresponsiveness, with overlapping allergic status but differing in symptom severity. Similarly, only two studies have considered airway remodeling in asthma phenotyping. Considering the importance of bronchial hyperresponsiveness and airway remodeling in asthma severity and monitoring of effectiveness of treatments, their inclusion in phenotyping will provide valuable insights. Number and list of variables used for phenotyping airway diseases are listed in Tables S3–S7 in the supplementary material.

Comparison with Previous Work

Our finding that the most reported asthma phenotypes were the ones differentiated by atopy, age at disease onset, and disease severity is consistent with the results from the review by Cunha et al and colleagues.¹³⁰ Our review similarly noted hierarchical cluster analysis as the most commonly reported method of phenotyping. Our review also found that majority of studies recruited participants from specialized healthcare centers. Pinto et al¹³¹ reported similar results, indicating that hierarchical and non-hierarchical clustering were the most commonly reported methods to derive COPD phenotypes.

Strength and Limitations

The current review followed recommended rigorous systematic review processes, including a priori protocol development, registration, and publication, and a comprehensive search of the literature across five leading healthcare databases, supplemented by grey literature and expert consultations. This approach minimized the risk of missing important studies. However, the variations in the included studies – in study design, outcome disease, methodological approaches employed, variables used for phenotyping, and derived phenotypes – inhibited comparison between studies. In addition, the review highlighted consistent issues with poor reporting practices, particularly regarding reproducibility, emphasizing the need for methodological improvements to enhance research quality and comparability in computational phenotyping of obstructive airway diseases.

Future Research Implications

Harmonizing methodological approaches in computational phenotyping of obstructive airway diseases is essential. Developing a consensus on key variables for phenotyping and standardizing participant selection will enhance the comparability and interpretation of findings. The quality assessment tool created for this study addresses a significant gap, and broader application could lead to further improvements and consensus on evaluating study quality in this field. Many studies lack essential details that ensure validity, such as transparent reporting of data processing, handling of missing data, and the rationale for computational choices. Future research should prioritize full transparency and incorporate reproducibility tools, including code and data sharing, version control, and environment management systems, to improve documentation and sharing, ultimately advancing the quality and consistency of research in this area.

Conclusion

The use of computational data-driven methods to derive phenotypes of airway diseases such as asthma, COPD, severe asthma, and ACO has resulted in significant variation in derived phenotypes across studies. This variability may be attributed to differences in sample selection, outcome measures, definitions, and variable selection used for phenotyping. The infrequent use of reproducibility measures in computational phenotyping research hinders the possibility of investigating the causes behind such variation. To achieve a better understanding and validity of the derived phenotypes and their clinical and scientific utility, a consistent approach to outcome definition and variable selection, as well as reproducible methods for phenotyping airway diseases, is needed.