Abstract
Background
Heterogeneity in chronic obstructive pulmonary disease (COPD) challenges the identification of optimal treatment populations. Subgroups derived from real-life cohorts using classification and regression trees (CARTs) have shown prognostic value for mortality, but their applicability to clinical trial populations to predict treatment response needs to be further investigated. We sought to evaluate whether previously identified CART-based subgroups are relevant for predicting outcomes and treatment response in the IMPACT trial, and to assess the stability of these subgroups using de novo clustering.
Methods
Post-hoc analysis of the IMPACT trial, a randomized controlled trial comparing inhaled corticosteroid (ICS)-containing dual or triple therapy to dual long-acting bronchodilation in patients with COPD. CART-based classification from prior real-life cohorts was applied to trial participants. Additionally, de novo clustering was performed using factor analysis of mixed data to identify alternative subgroup structures. Outcomes included mortality, exacerbation rates, lung function, dyspnea, and health status. Subgroups were compared descriptively in terms of baseline characteristics and on-treatment outcomes, with particular attention to blood eosinophil count and treatment response.
Results
Among 10,355 patients, both CART-based (5 classes) and de novo (5 clusters) classifications identified subgroups with distinct baseline profiles and variable on-treatment outcomes. Exacerbation and mortality rates differed markedly across subgroups, with numerically greater differences between treatment arms in higher-risk groups. Most clusters showed heterogeneous patterns of outcomes, while one cluster, characterized by elevated blood eosinophils (median 930/mm³), showed a numerically lower exacerbation rate with triple versus dual bronchodilation. Improvements in lung function and symptom scores were also more pronounced in this group. Despite limited concordance between the two methods, both consistently identified subgroups with higher event rates and greater numerical separation between treatment arms, supporting their potential value for clinical trial enrichment and personalized treatment strategies.
Conclusions
Application of a CART-based classification derived from real-life cohorts to a clinical trial population revealed subgroups with distinct baseline characteristics and differential treatment outcomes. These findings should be interpreted as exploratory and hypothesis-generating, and may inform future work on trial enrichment strategies and personalized approaches to COPD management.
Trial registration
The IMPACT trial was registered on ClinicalTrials.gov under number NCT02164513, first submitted on 12 June 2014.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12890-026-04312-0.
Keywords: COPD, Exacerbations, Mortality, Phenotype, Eosinophils
Background
COPD is a highly heterogeneous disease. Several patients’ characteristics are associated with clinically relevant outcomes including prognosis and therapeutic response. These characteristics span multiple domains encompassing clinical, functional, biological and imaging features. These characteristics can be used to identify biomarkers and define endotypes, phenotypes and treatable traits [1] guiding individually targeted management. Several clinical studies mostly based on observational cohorts have been performed to identify clinical phenotypes of COPD, using multivariate regression analyses [2–6] and clustering [7–10]. Applying clustering techniques to two distinct cohorts, we identified distinct subgroups of patients including two at higher risk of reduced survival [7, 11, 12]. The first corresponds to patients with severe airflow limitation occurring at an early age, the second to older patients with high rates of obesity, diabetes and cardiovascular comorbidities. In a large multi-cohort study, we used classification and regression trees (CART) to develop a decision tree to allocate individual patients to these clinical phenotypes (Fig. 1) [13], and confirm their relevance using mortality and age at death. CART is a statistical machine learning method used to break a dataset into smaller, meaningful groups. It builds a decision tree that predicts outcomes for both classification and regression tasks. It works by splitting data based on binary rules that reduce error at each node. Considering the highly different characteristics between subgroups, it may be hypothesized that they also respond differently to treatments. This classification approach could thus help develop personalized management strategies.
Fig. 1.
Initial classification tree
One central question regarding standard COPD pharmacological treatment relates to the best target population for inhaled corticosteroids (ICS)-containing therapy. Based on data from randomized clinical trials [14–18] the latest version of the GOLD strategy proposes to use these agents only as part of triple (ICS+ long-acting beta2 agonist [LABA]+long-acting antimuscarinic agent [LAMA]) combinations, as first-line therapy in GOLD E patients with blood eosinophil counts (BEC) > 300/mm3, or as step-up therapy in patients experiencing residual exacerbations despite LABA+LAMA therapy and blood eosinophils > 100/mm3.
We hypothesize that the classification techniques such as those developed in our cohort studies could be of particular interest to determine the best combination of variables (including clinical and lung function characteristics as well as eosinophil count) to identify subgroups with different outcomes and treatment response. To explore this hypothesis, we applied clustering and CART-derived classification techniques in a large randomized controlled trial (IMPACT) of triple versus dual inhaled therapy [14, 15, 19]. The main outcomes of interest were exacerbations and mortality. We explored two approaches: the application of previously identified subgroups and de novo development of new clusters.
Patients and methods
Study population, endpoints and analytic strategy outline
The study design is illustrated in Fig. 2. The study population was the ITT population of the IMPACT study (N = 10,355) [14]. The endpoints used for (1) the description of subgroups and decision trees and (2) the assessment of treatment effects in the subgroups were mortality, lung function (trough FEV1), dyspnea (TDI), health status (SGRQ, CAT), moderate and severe exacerbations (annualized rate, time-to-first) and rescue medication use. A moderate exacerbation was defined as an exacerbation treated with antibiotics or systemic glucocorticoids, a severe exacerbation was one resulting in hospitalization or death. Mortality was defined in IMPACT as adjudicated all-cause death occurring on-treatment (from treatment start date to treatment stop date plus 7 days). Time-to-death was analysed using the Kaplan–Meier method, and the cumulative probability of death was estimated accordingly. Reported percentages correspond to the Kaplan–Meier estimated cumulative probability of all-cause death at the end of the on-treatment follow-up period. These estimates reflect the probability of death over time, whereas mortality rates expressed per 1000 patient-years represent incidence rates accounting for person-time at risk. Analyses of annual exacerbation rates were performed using a generalized linear model assuming a negative binomial distribution and covariate of treatment group. Analyses of FEV1 were performed using a repeated measures model with covariates of cluster, treatment group, visit, baseline, baseline by visit, cluster by visit, cluster by treatment group, visit by treatment group and cluster by visit by treatment group interactions.
Fig. 2.
Study design
Two complementary approaches were explored to identify subgroups with different outcomes and treatment responses, especially regarding mortality. First, the previous CART-developed algorithm was applied to allocate IMPACT patients to previously identified clinically relevant subgroups (phenotypes) [13]. Second, a new set of factorial analyses was used for de novo unsupervised identification of phenotypes using variables and patients from IMPACT. This allowed to determine (1) the interest of this type of classification procedure to categorize clinical trial patients in relevant subgroups that could be used to predict outcomes and treatment responses; (2) the applicability of previously identified phenotypes to do so and, as a correlate, the need to update phenotypic identification before applying them to different settings and/or periods. To avoid confusion, subgroups identified by the original CART classification tree are referred to as “classes”, those identified by de novo clustering are referred to as “clusters”.
Identification of subgroups (classes) using the existing classification tree
The following variables were used in the original classification tree (Fig. 1): age, BMI, post-bronchodilator FEV1 (% predicted), mMRC dyspnea grade and presence/absence of cardiovascular conditions at baseline.
Since mMRC (used with a cutoff of 2 in the original classification tree) was not collected in IMPACT, we first tested whether it could be replaced by CAT with a cutoff of 10 following the GOLD strategy document, or if another cutoff was preferable (see online supplement).
Considering data available in IMPACT, cardiovascular conditions and risk factors considered for patients’ classification were angina pectoris, coronary artery disease, arrhythmia, congestive heart failure, hypertension, stroke, peripheral artery disease and diabetes mellitus.
For each class globally and by treatment arm, patient characteristics at baseline were summarized descriptively and summaries of the outcomes of interest (exacerbations, mortality, lung function, health status) were produced, but with no formal efficacy comparison.
De novo clustering to identify new subgroups
A new categorization scheme was developed using factorial analysis followed by hierarchical clustering. The selection of the variables used for the factorial analysis was guided by their relevance and availability (rate of missing values). These variables were: age, sex, BMI, post-bronchodilator FEV1 (% predicted), CAT, number of moderate (treatment with systemic corticosteroids and/or antibiotics) and of severe exacerbations (leading to hospitalization) during the 12 months prior to screening, smoking status, cumulative smoking (pack-years), presence/absence of cardiovascular conditions or risk factors (see above) and eosinophil count (as continuous or categorical variable). A descriptive summary of these variables was produced by treatment arm.
Factor analysis for mixed data (FAMD) [13, 20] was then used as a filter to extract the meaningful information, followed by classification of patients using Ward’s agglomerative hierarchical cluster analysis [7, 11–13, 21]. The factors retained for hierarchical clustering were those with an eigenvalue > = 1, i.e., those contributing more to explaining the relationships between original variables. A dendrogram and an Elbow’s plot were built to determine the optimal number of clusters in the data.
Patient characteristics at screening/baseline and outcomes were summarized by cluster and treatment arm within each cluster, again without performing any formal efficacy analysis.
Finally, the concordance between original CART-based classification and de novo-identified clusters was analyzed as well as the correspondence between patients’ characteristics and outcomes between CART-based and de novo identified subgroups.
Results
Adaptation of the original classification tree
Analyses presented in the online supplement determined that the best CAT threshold to replace mMRC dyspnea grade with a cutoff of 2 in the classification tree was 20. Table 1 shows the adapted rules to allocate patients to CART-defined classes.
Table 1.
Allocation rules used in the modified CART-defined classification tree
| Condition | Class |
|---|---|
| Patients with no cardiovascular risk factor | |
| If CAT > 20 and FEV1% ≤35 then | Class 4 |
| Or if CAT ≤ 20 and FEV1% ≥60 then | Class 5 |
| Otherwise | Class 2 |
| Patients with cardiovascular risk factor(s) | |
| If CAT > 20 then | Class 1 |
| Otherwise: | |
| If age ≤ 70 then | |
| If BMI > 30 then | Class 3 |
| Otherwise | Class 2 |
| Otherwise | |
| If FEV1% ≥50 | Class 3 |
| Otherwise | Class 1 |
Study population: application of the CART-based allocation rules
The IMPACT population (n = 10,355) has been extensively described previously. The main baseline characteristics are summarized in Table 2 by original CART-based class. This table also presents a narrative summary of the main differentiating features for each class. Classes 3 and 4 were proportionally underrepresented.
Table 2.
Patients’ characteristics in the IMPACT trial by original CART-based class
| Class 1 | Class 2 | Class 3 | Class 4 | Class 5 | |
|---|---|---|---|---|---|
| n = 3576 | n = 4364 | n = 1318 | n = 632 | n = 426 | |
| Age, years | 67 [61; 74] | 64 [58; 68] | 69 [63; 74] | 62 [57; 67] | 63 [58; 70] |
| Male sex | 2346 (66%) | 2885 (66%) | 930 (71%) | 402 (64%) | 279 (65%) |
| Current smokers | 1207 (34%) | 1630 (37%) | 299 (23%) | 266 (42%) | 172 (40%) |
| BMI, kg/m2 | 27.1 [23.3;31.5] | 24.5 [21.5;27.6] | 31.6 [26.7;34.8] | 22.5 [19.6;26.5] | 25.5 [22.4;28.7] |
| Post BD FEV1% pred | 42 [33; 51] | 44 [36; 54] | 56 [47; 65] | 28 [23; 32] | 68 [64; 72] |
| GOLD | |||||
| 1 | 9 (< 1%) | 7 (< 1%) | 3 (< 1%) | 0 (0%) | 3 (< 1%) |
| 2 | 931 (26%) | 1432 (33%) | 924 (70%) | 0 (0%) | 423 (99%) |
| 3 | 2002 (56%) | 2373 (54%) | 336 (25%) | 257 (41%) | 0 (0%) |
| 4 | 634 (18%) | 552 (13%) | 55 (4%) | 375 (59%) | 0 (0%) |
| CAT score at screening | 24 [21; 27] | 17 [14; 20] | 16 [13; 19] | 25 [23; 29] | 15 [13; 18] |
| ≥ 2 moderate exacerbations or ≥ 1 severe in the past year | 2400 (67%) | 2962 (68%) | 1123 (85%) | 367 (58%) | 424 (> 99%) |
| Any CV risk factor | 3576 (> 99%) | 1959 (45%) | 1318 (> 99%) | 78 (12%) | 49 (12%) |
| Treatment arm | |||||
| FF/UMEC/VI | 1391 (39%) | 1759 (40%) | 557 (42%) | 251 (40%) | 175 (41%) |
| FF/VI | 1468 (41%) | 1719 (39%) | 511 (39%) | 255 (40%) | 170 (40%) |
| UMEC/VI | 717 (20%) | 886 (20%) | 250 (19%) | 126 (20%) | 81 (20%) |
| Narrative summary of classes’ main features | Older, CV comorbidities, symptomatic | Average / no highly specific features | Older, higher BMI, few current smokers, CV comorbidities, exacerbations | Symptomatic, lower FEV1 | Exacerbations, less impaired FEV1 |
Data are median [Q1-Q3] and n (%)
CAT COPD assessment test, BD bronchodilator, BMI body mass index, CV cardiovascular, FEV1 forced expiratory volume in one second, FF fluticasone furoate, UMEC umeclidinium, VI vilanterol
Factorial analysis and clustering to identify new subgroups
For these analyses, data from 10,294 patients were used after exclusion of those with missing data for key variables. FAMD using 11 variables identified 5 factors based on the selected variables, which explained 60.59% of the variance (see supplementary Figure S1). Figure S2 presents the variables contributing to the 2 main dimensions.
The dendrogram resulting from clustering is shown in Figure S3. Elbow’s plot (also shown in Figure S3) identified an optimal number of 3 clusters. However, it was decided to ultimately use 5 clusters to increase the amount of variance explained and increase coherence with the original number of classes. Table 3 depicts the population’s characteristics by cluster and the corresponding narrative descriptions for each cluster. Cluster 3 (characterized by markedly higher eosinophil counts) was notably underrepresented.
Table 3.
Patients’ characteristics in the IMPACT trial by new cluster
| Cluster 1 | Cluster 2 | Cluster 3 | Cluster 4 | Cluster 5 | |
|---|---|---|---|---|---|
| n = 2074 | n = 3198 | n = 449 | n = 2554 | n = 2019 | |
| Age, years | 66 [61;71] | 71 [67;75] | 65 [60;71] | 61 [56;66] | 61 [57;66] |
| Male sex | 1534 (74%) | 2773 (87%) | 394 (88%) | 1498 (59%) | 629 (31%) |
| Current smokers | 596 (29%) | 318 (10%) | 154 (34%) | 1485 (58%) | 1011 (50%) |
| BMI, kg/m2 | 26.3 [22.1;29.7] | 26.6 [23.4;29.5] | 25.6 [21.8;28.8] | 22.8 [19.8;25.4] | 32.0 [27.6;35.6] |
| Post BD FEV1% pred | 43 [31;54] | 47 [37;57] | 51 [39;65] | 39 [29;47] | 53 [42;64] |
| GOLD | |||||
| 1 | 1 (< 1%) | 7 (< 1%) | 1 (< 1%) | 1 (< 1%) | 12 (< 1%) |
| 2 | 681 (33%) | 1201 (38%) | 221 (49%) | 492 (19%) | 1108 (55%) |
| 3 | 953 (46%) | 1636 (51%) | 181 (40%) | 1380 (54%) | 805 (40%) |
| 4 | 439 (21%) | 354 (11%) | 46 (10%) | 681 (27%) | 94 (5%) |
| CAT score at screening | 20 [16;24] | 18 [14;21] | 18 [14;21] | 21 [16;25] | 23 [18;27] |
| ≥ 2 moderate exacerbations or ≥ 1 severe in the past year | 625 (30%) | 1971 (62%) | 315 (70%) | 1229 (48%) | 1487 (74%) |
| Any CV risk factor | 1560 (74%) | 2686 (84%) | 273 (61%) | 727 (28%) | 1720 (85%) |
| Eosinophil count /mm3 | 210 [100;270] | 190 [90;250] | 930 [620;1030] | 180 [80;240] | 190 [90;250] |
| Treatment arm | |||||
| FF/UMEC/VI | 829 (40%) | 1277 (40%) | 178 (40%) | 1047 (41%) | 794 (39%) |
| FF/VI | 846 (41%) | 1287 (40%) | 170 (38%) | 1000 (39%) | 811 (40%) |
| UMEC/VI | 399 (19%) | 634 (20%) | 101 (22%) | 507 (20%) | 414 (21%) |
| Narrative summary of clusters’ main features | Lower FEV1 | Male, older, less current smokers | Eosinophilic, men, exacerbations | Less CV comorbidities, lower FEV1, less exacerbations | More CV comorbidities, female, exacerbations |
Data are median [Q1-Q3] and n (%)
CAT COPD assessment test, BD bronchodilator, BMI body mass index, CV cardiovascular, FEV1 forced expiratory volume in one second, FF fluticasone furoate, UMEC umeclidinium, VI vilanterol
Correspondence between original classes and new subgroups
Table 4 shows the correspondence between original and new subgroups, which was rather limited overall.
Table 4.
Concordance between original classes and new subgroups
Each cell presents (1) in purple, the percentage of individuals from a given class that fall into a specific cluster (i.e., row percentage relative to the class total) and in red, the proportion of each cluster represented by individuals from a given class (i.e., column percentage relative to the cluster total)
Outcomes by treatment arm in original classes and new clusters
Figures 3 and 4 depict moderate-to-severe exacerbation rates and all-cause mortality rates, respectively, by treatment arm for each CART class (Figs. 3A and 4A) and for each new cluster (Figs. 3B and 4B). All main and secondary outcomes are presented in supplementary tables by treatment arm for each CART class (Tables S1-S10) and for each new cluster (Tables S11-S20). Regarding exacerbation rates, they were greater in class 4 and lower in class 5, with visually similar gradients by treatment group (UMEC/VI > FF/VI > FF/UMEC/VI), except for cluster 5 in which exacerbations rates were similar in FF/UMEC/VI and FF/VI arms. Moderate-to-severe exacerbation rates were more homogeneous between new clusters in patients on triple therapy, while on UMEC/VI they were notably higher in cluster 3 and, to a lesser extent, in cluster 1, suggesting greater differences between triple therapy and dual bronchodilation in these clusters. Regarding mortality, rates were numerically lower with triple therapy compared with dual bronchodilation in class 1 (older symptomatic patients with higher BMI and more frequent cardiovascular comorbidities) and cluster 1 (patients with exacerbations and more impaired lung function). In class 4, differences in lung function evolution between treatment arms were minimal compared to other classes, while in class 5 a more pronounced numerical difference in favor of triple therapy was observed versus the two other treatment arms (Table S7). Class 5 was associated with a greater proportion of TDI, SGRQ and CAT responders to triple therapy compared to the 4 other classes. Regarding new clusters, FEV1% predicted improved more in all treatment arms in cluster 3, with numerically greater improvements with triple therapy compared to dual treatments, as in other clusters. In this cluster, exacerbation rates were numerically lower with ICS-containing regimens compared to UMEC/VI. In cluster 5, FEV1 improvement was also greater in the triple therapy arm compared to other clusters. Regarding symptomatic responses, cluster 3 showed a distinct pattern, with larger numerical differences between treatment arm in terms of SGRQ and CAT.
Fig. 3.
Exacerbation rate by original CART class (panel A) and new cluster (panel B). Data are estimates (number of exacerbations per patient-year) and 95% confidence intervals
Fig. 4.
Probability of all-cause death by original CART class (panel A) and new cluster (panel B). Data are Kaplan–Meier estimated cumulative probabilities (%) of all-cause death at the end of the on-treatment follow-up period, with 95% confidence intervals. These estimates reflect cumulative risk over time and are not directly comparable to mortality rates per 1000 patient-years (Tables S3 and S13), which account for person-time at risk
Discussion
Summary of results
CART-based classification and de novo clustering using the IMPACT population identified populations with markedly different characteristics especially in terms of demographics, smoking status, lung function, symptoms, exacerbations, cardiovascular risk factors and comorbidities. The two classification approaches resulted in rather poorly concordant subgroups. Outcomes patterns of differences between treatment arms varied across subgroups identified using both CART and de novo clustering. The new clustering identified a relatively small group with markedly higher blood eosinophil count and more favourable outcomes. Both approaches identified subgroups at increased risk of exacerbations and mortality, and larger numerical differences between treatment arm in favour of triple therapy.
Interpretation of results
Several factors can explain the differences between classes identified by the original CARTs and the new clusters identified by de novo clustering: differences in populations are likely, the original real-life cohort population used to derive the CARTs being less selected than the IMPACT population, a typical randomized controlled trial population of patients at risk of exacerbations [22]. Possibly due to the greater sample size in IMPACT, more variables were selected to contribute to the new clustering (additional variables being exacerbations, gender, smoking status and blood eosinophil counts), compared to the variables used by the original CARTs. As a result, clusters differ more than classes in terms of smoking status and gender. Conversely, they differ less in terms of symptoms and reversibility (not shown). In the original phenotyping study that led to the subsequent CART analyses, eosinophil counts were not systematically available, compromising their use in classification trees. In the new clustering process, integration of eosinophil count in the analyses led to the identification of a small subgroup with high eosinophil count. Otherwise, the other clusters were similar with respect to eosinophil count. Although not formally assessed statistically, it does not seem that one classification is more discriminant than the other in terms of both baseline characteristics and outcomes. Both the original CART analyses and the new clustering process identified subgroups with different patterns of outcomes across treatment arms on exacerbations and mortality, the two main outcomes of interest in the present study. Cluster 3 displayed a distinct profile compared with other subgroups, characterized by markedly elevated blood eosinophil counts (mean 990 ± 600 cells/µL; median 930 [IQR 620–1030]), with 75% of patients above 600 cells/µL, suggesting that this pattern was not driven by extreme values. This subgroup was associated with larger numerical differences between treatment arms across exacerbations, lung function and symptoms. These findings remain descriptive and should be interpreted with caution. While such a profile may be compatible with a type 2 inflammatory signal, cluster 3 did not exhibit greater reversibility than other clusters (data not shown), and alternative explanations cannot be excluded. In particular, a higher proportion of patients originated from regions where parasitic infections are more prevalent, which may have contributed to elevated eosinophil levels.
Comparison with existing literature
The variables contributing to these classification approaches are well established prognostic factors in COPD, some being actually used in prognostic scores such as the BODE index or the CODEX [23, 24]. Conversely, their contribution as predictors of treatment response is less known. Here we observed some differences in treatment responses according to class or cluster, which warrant further prospective investigations. Importantly, exacerbation rates differed both between classes (as in the original CART analyses) [13], and between clusters, while being used for subgroup definition only in the new clustering approach. It is intriguing to observe that the two subgroups in which treatment effects on mortality were more marked, i.e. class 1 and cluster 1, were rather different in terms of differentiating characteristics: class 1 is characterized by older age, more cardiovascular risk factors and comorbidities and greater symptom burden, while cluster 1 is characterized by more exacerbations and lower lung function, translating a more “respiratory” phenotype.
Strengths
The main strength of this study is that it relies on a well characterized clinical trial population with very few missing data and high-quality follow-up. Moreover, the original classification scheme was based on a succession of concordant analyses of 3 real-life cohorts, followed by phenotypes validation and CART analyses in a large sample of 3? 651 patients from the 3CIA consortium of COPD cohorts. Finally, the de novo development of new clusters used a robust well-established factorial and clustering approach.
Limitations
The findings correspond only to populations similar to that recruited in the IMPACT trial, but this actually corresponds to the main objective of the analyses, i.e., the identification of patients’ subgroups of special interest relative to outcomes and treatment effects.
The analyses inherently rely only on available data. Some potentially relevant items such as asthma history, some biological or imaging variables are absent from IMPACT as from most randomized COPD trials. However, patients’ characterization is rather extensive in these studies. In addition, there are slight differences in the operational definition of the cardiovascular (CV) composite between the IMPACT dataset and the original CART derivation cohorts. In IMPACT, the CV composite included angina, coronary artery disease, arrhythmia, congestive heart failure, hypertension, stroke, peripheral artery disease, and diabetes, whereas the original CART analyses considered cardiovascular comorbidities (hypertension, coronary artery disease and/or left heart failure) and/or diabetes. These differences in variable definition and data capture may have influenced class allocation and limit strict comparability across settings, but their impact on patients’ classification is likely marginal, although available data do not allow to formally confirm this.
Future perspectives
The findings reported here may serve as bases for enriching clinical trials in patients with greater frequency of occurrence of the outcomes of interest, namely exacerbations and mortality. In line with results from many studies and trials, these classification approaches confirm that cardiovascular comorbidities play a major role as determinants of outcomes and treatment effects in patients with COPD.
Conclusions
These analyses show that applying a previously established classification developed using real-life cohorts to a clinical trial population identifies subgroups with different baseline characteristics. New clustering of this population identifies somehow different subgroups. Both the original CART classification and the new clustering identified subgroups with different on-treatment outcomes, particularly in terms of exacerbation frequency and mortality, as well as apparent differences in outcomes across treatment arms. This may be useful to enrich future clinical trials in patients with higher likelihood of events and response, and to guide clinical practice through the identification of the best target populations. The relevance of the findings will need to be prospectively tested in specific trials.
Supplementary Information
Abbreviations
- BMI
Body Mass Index
- BD
Bronchodilator
- BEC
Blood Eosinophil Count
- CAT
COPD Assessment Test
- CART
Classification and Regression Trees
- CV
Cardiovascular
- COPD
Chronic Obstructive Pulmonary Disease
- FEV₁
Forced Expiratory Volume in One Second
- FAMD
Factor Analysis for Mixed Data
- FF
Fluticasone Furoate
- GOLD
Global Initiative for Chronic Obstructive Lung Disease
- ICS
Inhaled Corticosteroids
- ITT
Intention-To-Treat
- LABA
Long-Acting Beta₂-Agonist
- LAMA
Long-Acting Muscarinic Antagonist
- mMRC
Modified Medical Research Council Dyspnea Scale
- SGRQ
St George’s Respiratory Questionnaire
- TDI
Transition Dyspnea Index
- TTF
Time-To-First
- UMEC
Umeclidinium
- VI
Vilanterol
Authors’ contributions
NR had full access to all of the data in the study and takes responsibility for the integrity of thedata and the accuracy of the data analysis. JLP, NR and PRB contributed substantially to the study design, data analysis and interpretation. LR contributed substantially to data analysis and interpretation, wrote the first draft of the manuscript that was revised and approved by all authors regarding important intellectual content. All authors approved the final version of the manuscript.
Funding
The IMPACT study was sponsored by GSK. The present analyses were investigator-initiated, jointly designed by the authors and performed by GSK. Statistical input from Effi-Stat (JLP) was funded by GSK. The development of the original CART decision-trees applied in part of these analyses was funded by Boehringer-Ingelheim. The original phenotyping approaches inspiring the present work were performed by the Initiatives BPCO study group with the financial support of Boehringer Ingelheim and Pfizer.
Data availability
The data that support the findings of this study are available from GSK but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of GSK.
Declarations
Ethics approval and consent to participate
The IMPACT trial was performed in 37 countries from June 2014 through July 2017. It was conducted in accordance with Good Clinical Practice guidelines and the provisions of the Declaration of Helsinki and received approval from local institutional review boards or independent ethics committees. All the patients provided written informed consent.
Consent for publication
Not applicable.
Competing interests
NR reports personal fees from GSK, AstraZeneca, Sanofi, Chiesi, Pfizer, Austral, Biosency, Zambon, MSD, and Menarini for consulting or speaking engagements, and institutional support from Chiesi, GSK, and Pfizer. He also serves as Chair of the ERS Science Council. PRB reports personal fees from AstraZeneca, Boehringer Ingelheim, Chiesi, GSK, Insmed, MSD, Pfizer, Sanofi, Vertex, and Viatris, and institutional support from AstraZeneca and Chiesi. LR reports personal fees from AstraZeneca, Chiesi, GSK, and Sanofi, and institutional support for meeting attendance from AstraZeneca, Chiesi, and Sanofi. JLP is employed by Effi-Stat.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Pierre-Régis Burgel and Nicolas Roche contributed equally to this work.
References
- 1.Agusti A, Bel E, Thomas M, et al. Treatable traits: toward precision medicine of chronic airway diseases. Eur Respir J. 2016;47(2):410–9. 10.1183/13993003.01359-2015. [DOI] [PubMed] [Google Scholar]
- 2.Le Rouzic O, Roche N, Cortot AB, et al. Defining the Frequent Exacerbator Phenotype in COPD: A Hypothesis-Free Approach. Chest. 2018;153(5):1106–15. 10.1016/j.chest.2017.10.009. [DOI] [PubMed] [Google Scholar]
- 3.Burgel PR, Nesme-Meyer P, Chanez P, et al. Cough and sputum production are associated with frequent exacerbations and hospitalizations in COPD subjects. Chest. 2009;135(4):975–82. [DOI] [PubMed] [Google Scholar]
- 4.McGarvey L, Lee AJ, Roberts J, Gruffydd-Jones K, McKnight E, Haughney J. Characterisation of the frequent exacerbator phenotype in COPD patients in a large UK primary care population. Respir Med. 2015;109(2):228–37. 10.1016/j.rmed.2014.12.006. [DOI] [PubMed] [Google Scholar]
- 5.Vestbo J, Prescott E, Lange P, The Copenhagen City Heart Study Group. Association of chronic mucus hypersecretion with FEV1 decline and chronic obstructive pulmonary disease morbidity. Am J Respir Crit Care Med. 1996;153:1530–5. [DOI] [PubMed] [Google Scholar]
- 6.Hurst JR, Vestbo J, Anzueto A, et al. Susceptibility to exacerbation in chronic obstructive pulmonary disease. N Engl J Med. 2010;363(12):1128–38. 10.1056/NEJMoa0909883. [DOI] [PubMed] [Google Scholar]
- 7.Burgel PR, Paillasseur JL, Caillaud D, et al. Clinical COPD phenotypes: a novel approach using principal component and cluster analyses. Eur Respir J. 2010;36(3):531–9. 10.1183/09031936.00175109. [DOI] [PubMed] [Google Scholar]
- 8.Garcia-Aymerich J, Gómez FP, Benet M, et al. Identification and prospective validation of clinically relevant chronic obstructive pulmonary disease (COPD) subtypes. Thorax. 2011;66(5):430–7. 10.1136/thx.2010.154484. [DOI] [PubMed] [Google Scholar]
- 9.Paoletti M, Camiciottoli G, Meoni E, et al. Explorative data analysis techniques and unsupervised clustering methods to support clinical assessment of Chronic Obstructive Pulmonary Disease (COPD) phenotypes. J Biomed Inf. 2009;42(6):1013–21. 10.1016/j.jbi.2009.05.008. [DOI] [PubMed] [Google Scholar]
- 10.Vanfleteren LEGW, Spruit MA, Groenen M, et al. Clusters of comorbidities based on validated objective measurements and systemic inflammation in patients with chronic obstructive pulmonary disease. Am J Respir Crit Care Med. 2013;187(7):728–35. 10.1164/rccm.201209-1665OC. [DOI] [PubMed] [Google Scholar]
- 11.Burgel PR, Paillasseur JL, Peene B, et al. Two distinct chronic obstructive pulmonary disease (COPD) phenotypes are associated with high risk of mortality. PLoS ONE. 2012;7(12):e51048. 10.1371/journal.pone.0051048. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Burgel PR, Roche N, Paillasseur JL, et al. Clinical COPD phenotypes identified by cluster analysis: validation with mortality. Eur Respir J. 2012;40(2):495–6. 10.1183/09031936.00228511. [DOI] [PubMed] [Google Scholar]
- 13.Burgel PR, Paillasseur JL, Janssens W, et al. A simple algorithm for the identification of clinical COPD phenotypes. Eur Respir J. 2017;50(5). 10.1183/13993003.01034-2017. [DOI] [PubMed]
- 14.Lipson DA, Barnhart F, Brealey N, et al. Once-daily single-inhaler triple versus dual therapy in patients with COPD. N Engl J Med Published online April. 2018;18. 10.1056/NEJMoa1713901. [DOI] [PubMed]
- 15.Lipson DA, Crim C, Criner GJ, et al. Reduction in All-Cause Mortality with Fluticasone Furoate/Umeclidinium/Vilanterol in Patients with Chronic Obstructive Pulmonary Disease. Am J Respir Crit Care Med. 2020;201(12):1508–16. 10.1164/rccm.201911-2207OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Rabe KF, Martinez FJ, Ferguson GT, et al. Triple Inhaled Therapy at Two Glucocorticoid Doses in Moderate-to-Very-Severe COPD. N Engl J Med. 2020;383(1):35–48. 10.1056/NEJMoa1916046. [DOI] [PubMed] [Google Scholar]
- 17.Martinez FJ, Rabe KF, Ferguson GT, et al. Reduced all-cause mortality in the ETHOS trial of budesonide/glycopyrrolate/formoterol for COPD: a randomized, double-blind, multi-center parallel-group study. Am J Respir Crit Care Med Published online November. 2020;30. 10.1164/rccm.202006-2618OC. [DOI] [PMC free article] [PubMed]
- 18.Singh D, Fabbri LM, Vezzoli S, Petruzzelli S, Papi A. Extrafine triple therapy delays COPD clinically important deterioration vs ICS/LABA, LAMA, or LABA/LAMA. Int J Chron Obstruct Pulmon Dis. 2019;14:531–46. 10.2147/COPD.S196383. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Pascoe S, Barnes N, Brusselle G, et al. Blood eosinophils and treatment response with triple and dual combination therapy in chronic obstructive pulmonary disease: analysis of the IMPACT trial. Lancet Respir Med Published online July. 2019;4. 10.1016/S2213-2600(19)30190-0. [DOI] [PubMed]
- 20.Pages J. Multiple Factor Analysis by Example Using R. Boca Raton, FL. Chapman & Hall/CRC; 2014.
- 21.Burgel PR, Paillasseur JL, Dusser D, et al. Tiotropium might improve survival in subjects with COPD at high risk of mortality. Respir Res. 2014;15:64. 10.1186/1465-9921-15-64. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Pahus L, Burgel P, Roche N, Paillasseur J, Chanez P. Randomized controlled trials of pharmacological treatments to prevent COPD exacerbations: applicability to real-life patients. BMC Pulm Med. 2019;19:127. 10.1186/s12890-019-0882-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Celli BR, Cote CG, Marin JM, et al. The body-mass index, airflow obstruction, dyspnea, and exercise capacity index in chronic obstructive pulmonary disease. N Engl J Med. 2004;350(10):1005–12. [DOI] [PubMed] [Google Scholar]
- 24.Almagro P, Martínez-Camblor P, Miravitlles M, et al. External Validation and Recalculation of the CODEX Index in COPD Patients. A 3CIAplus Cohort Study. COPD. 2019;16(1):8–17. 10.1080/15412555.2018.1484440. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The data that support the findings of this study are available from GSK but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of GSK.