Skip to main content

Stratification of COPD patients towards personalized medicine: reproduction and formation of clusters



The global initiative for chronic obstructive lung disease (GOLD) 2020 emphasizes that there is only a weak correlation between FEV1, symptoms and impairment of the health status of patients with chronic obstructive pulmonary disease (COPD). Various studies aimed to identify COPD phenotypes by cluster analyses, but behavioral aspects besides smoking were rarely included.


The aims of the study were to investigate whether (i) clustering analyses are in line with the classification into GOLD ABCD groups; (ii) clustering according to Burgel et al. (Eur Respir J. 36(3):531–9, 2010) can be reproduced in a real-world COPD cohort; and (iii) addition of new behavioral variables alters the clustering outcome. Principal component and hierarchical cluster analyses were applied to real-world clinical data of COPD patients newly referred to secondary care (n = 155). We investigated if the obtained clusters paralleled GOLD ABCD subgroups and determined the impact of adding several variables, including quality of life (QOL), fatigue, satisfaction relationship, air trapping, steps per day and activities of daily living, on clustering.


Using the appropriate corresponding variables, we identified clusters that largely reflected the GOLD ABCD groups, but we could not reproduce Burgel’s clinical phenotypes. Adding six new variables resulted in the formation of four new clusters that mainly differed from each other in the following parameters: number of steps per day, activities of daily living and QOL.


We could not reproduce previously identified clinical COPD phenotypes in an independent population of COPD patients. Our findings therefore indicate that COPD phenotypes based on cluster analysis may not be a suitable basis for treatment strategies for individual patients.


The severity of chronic obstructive pulmonary disease (COPD) is defined by forced expiratory volume in 1 s (FEV1), divided into four stages of severity (Global initiative for chronic obstructive lung disease; GOLD I-IV) [1]. Nonetheless, the heterogeneity of the clinical presentation and disease development among patients within the same GOLD stage is substantial [2]. GOLD 2011 introduced the ABCD assessment tool to classify stable COPD patients on the basis of airflow limitation, number of exacerbations per year and questionnaires to measure the severity of symptoms: modified Medical Research Council (mMRC) scale and COPD Assessment Test (CAT) [1]. This approach is more comprehensive than airflow limitation alone, but it is still based on a limited number of variables. In GOLD 2017 a refinement of the ABCD assessment tool was suggested, which separated the spirometric GOLD I-IV grades from the ABCD groups and introduced pharmacotherapy recommendations per ABCD group.

An alternative method facilitating subgroup-specific treatment might be established by the identification of phenotypes on the basis of prognostic, demographic, clinical, pathophysiological or therapeutic characteristics. Han et al. proposed the following definition of phenotypes in the context of COPD: “a single or combination of disease attributes that describe differences between individuals with COPD as they relate to clinically meaningful outcomes” [3].

Phenotyping can be aided by using descriptive statistics, such as cluster analysis to identify separate patient groups according to preselected variables [4]. With regards to these variables, patients within a certain cluster are more similar to each other than to patients in different clusters [5]. The identification of coherent clusters may lead to the recognition of phenotypes, which could be a valuable step towards tailored treatment strategies per subgroup.

Several attempts have been made to develop a useful classification of phenotypes of COPD patients [4]. To be potentially useful in clinical practice, the identity of the defined clusters needs to be confirmed in different, independent cohorts of COPD patients, but to the best of our knowledge such replication studies have not been performed yet. Burgel et al. [6] performed more extensive phenotyping of COPD patients based on the clinical variables age, cumulative smoking, airflow limitation, body mass index (BMI), exacerbations per year, dyspnea, health status and depressive symptoms. Hereby, four clinical phenotypes were defined: phenotype 1 were relatively young subjects (median 58 [IQR 55–63] years old) with predominantly severe to very severe respiratory disease, frequent exacerbations and low BMI; phenotype 2 were older patients (median 68 [IQR 60–74] years old) with mild symptoms; phenotype 3 were younger subjects (median 59 [IQR 50–65] years old) with moderate to severe airflow limitation. In the fourth phenotype older patients (median 72.5 [IQR 67–77] years old) with moderate to severe airflow limitation were included. Compared to phenotype 3, these patients had a higher prevalence of depressive symptoms, higher BMI and more severe dyspnea. Patients with comparable FEV1 were assigned to different phenotypes [6]. Longitudinal 2-year follow-up showed that phenotype 2 is associated with a very low risk of mortality and that patients with phenotype 1 had the highest mortality rates and died at a younger age [7].

Moreover, in the current literature, cluster classifications are largely based on clinical variables, while behavioral variables are rarely used. Whereas most attention is drawn to smoking behavior [6, 8], other behavioral aspects such as coping, physical activity and quality of life (QOL) are not included. Nevertheless, these are important variables because they influence the impact of self-management interventions and can interfere with active participation [9]. For example, a high rate of physical activity is known to increase shortness of breath and therefore it is avoided by most COPD patients. On the other hand, in the long run physical exercise in COPD is associated with a reduction of shortness of breath [10]. Shortness of breath during physical activity can be mechanically influenced by air trapping, which makes this an interesting physiological parameter to add to phenotyping [11].

The aim of this study is threefold. First, to investigate whether the results of our cluster analyses match the ABCD groups defined by the GOLD criteria, which are either based on CAT and exacerbation frequency or on mMRC and exacerbation frequency. Second, to address whether the four COPD phenotypes previously identified by Burgel et al. can be reproduced by cluster analysis in another real-world COPD cohort. Third, to determine whether the addition of six new variables: QOL, fatigue, satisfaction relationship, air trapping and steps per day and activities of daily living improves the classification into distinct subgroups.


Study design

We performed three independent cluster analyses of COPD patient characteristics. In the first analysis, we aimed to identify clusters that corresponded with the ABCD groups [1]. Secondly, we investigated the reproducibility of Burgel’s clusters in our study population [6]. Thirdly, we added six new variables, QOL, fatigue, satisfaction relationship, air trapping and steps per day and activities of daily living to the parameter of the second analysis.

Setting and participants

Data were part of a registry study of patients with asthma and COPD, who were newly referred to the Franciscus Gasthuis and Vlietland Hospital in Rotterdam, the Netherlands, a center of excellence for asthma and COPD. All referred COPD patients (n = 155) who completed a previously published [12] comprehensive assessment during the period December 2012 till December 2017 were included. The diagnosis of COPD was based on an assessment by a pulmonologist and confirmed by spirometry (FEV1 / forced vital capacity (FVC) < 0.7). In this study, we used pseudonymized assessment data. Ethics approval for this study was waived by the Institutional Research Board of the Franciscus Gasthuis & Vlietland, Rotterdam, the Netherlands, because routinely collected health care data were used after pseudonymization.

Data collection

The following variables were collected for all patients:

Lung function. FEV1, FVC, and static and dynamic hyperinflation were performed according to the ATS/ ERS taskforce “standardization of spirometry” [13, 14]. Values for post-bronchodilation dynamic hyperinflation were measured by metronome-paced tachypnea after bronchodilation (400 µg of inhaled salbutamol) [15]. Lower levels of air trapping (dynamic hyperinflation after bronchodilation measured in liters decreasing inspiratory capacity) reflect poor outcome. All tests were performed with the Vmax Sensor Medics Viasys, type 6200 Encore.

Pack years. A pack year is defined as twenty cigarettes smoked per day for 1 year.

Body mass index (BMI): BMI is defined as the body mass divided by the square of the body height, expressed in units kg/m2 [16].

Exacerbations. The number of antibiotic courses and/or systemic steroids for their respiratory disease in the previous year (0, 1, 2 or 3 +).

Symptoms and health status. Modified Medical Research Council (mMRC) is a five-item questionnaire to score the dyspnea of COPD patients [17, 18] (Table 1). The Clinical COPD Questionnaire (CCQ) is a ten-item questionnaire about symptom severity [19]. A higher score indicates a worse health status. The minimal clinically important difference is 0.4 [20]. The Beck Depression Inventory for primary care (BDI-PC) was used to score symptoms of depression independently of physical function [21, 22].

Table 1 Scoring range of questionnaires

BOD-score. BOD-score includes the variables BMI, airflow obstruction and dyspnea in COPD (Table 1). Higher BOD scores for indicate a greater risk of death [23].

Physical activity. Physical activity was measured by an activity tracker (McRoberts© Triaxial accelerometer) during 1 week. The mean number of steps in 24 h over 7 days was used for analyses.

ABCD groups. Group ‘A’ includes patients with mMRC 0–1 or COPD Assessment Test (CAT) < 10 and 0–1 exacerbation per year; group ‘B’ includes patients with mMRC ≥ 2 or CAT ≥ 10 and 0–1 exacerbation per year; group ‘C’ includes patients with ≥ 2 exacerbations or ≥ 1 exacerbation leading to hospital admission with mMRC 0–1 or CAT < 10 and group ‘D’ includes patients with ≥ 2 exacerbations or ≥ 1 exacerbation leading to hospital admission with mMRC ≥ 2 or CAT ≥ 10 [1].

Nijmegen Clinical Screening Instrument (NCSI). Four NCSI domains were included; QOL, satisfaction relationship, behavioral impairment (termed activities of daily living) and fatigue [24]. The minimum and maximum scores are shown in Table 1.


Following the methodology applied by Burgel et al. [6], we used Principal Component Analysis (PCA) and Ward’s hierarchical cluster analysis. In Ward’s method, the analysis starts with each subject forming its own cluster [25]. Step by step, the number of clusters is reduced until all subjects are in one cluster. In each step, the two most similar clusters from the previous steps are combined, based on the variables that are have been selected to describe the clusters. These two clusters are selected in such a way that the total of the variances of all variables within the new clusters is as small as possible. Before clusters were formed, PCA was used to reduce the number of variables by replacing them by newly created uncorrelated variables (‘components’) with minimal information loss [26].

Dendrograms were used to graphically represent the hierarchical relationship between the clusters and the distance between them. The resulting cluster solutions were described and compared to determine the optimal number of clusters. All statistical analyses were performed using Stata/SE 15.1. Following Burgel et al. [6], variables were standardized (i.e. using Z-score or PCA) before they were included in the cluster analysis and categorical or dichotomous variables were expressed numerically. PCA were performed to reduce interaction between the variables included in the cluster analyses. Components with an eigenvalue > 1 were used.

In the first cluster analysis, in which we aimed to match ABCD groups, two times two variables (exacerbations per year and mMRC versus exacerbations per year and CAT) were used. We used two types of symptom questionnaires (CAT and mMRC) as a marker of disease burden, because GOLD uses either one of these to form the ABCD groups.

In the second cluster analysis, in which we investigated whether the four clinical phenotypes of Burgel could be reproduced, the following eight variables were used: age, packyears, FEV1, BMI, mMRC, CCQ, BDI-PC and the number of exacerbations per year. Given the availability of data, CCQ analysis was used instead of the St. George Respiratory Questionnaire [27], and BDI-PC instead of the Hospital Anxiety and Depression Scale [28]. For comparative purposes in Fig. 3, we projected the CCQ and BDI-PC scores of our study onto the SGRQ and HADS (Burgel’s study) as follows. We divided the 50th, 25th and 75th percentiles of the CCQ and BDI-PC by the score range (which is 6 and 21 respectively), and multiplied this with the score range of the SGRQ and HADS which is 100 and 42 respectively).

In the third cluster analysis, six new variables, NCSI QOL, NCSI satisfaction relationship, NCSI fatigue, air trapping, steps per day and NCSI activities of daily living were added (Fig. 1).

Fig. 1
figure 1

Flowchart of patient enrollment in the different cluster analyses


Patient characteristics and ABCD classification

COPD Patients were divided into ABCD groups, based on the mMRC definition (n = 155) or the CAT definition (n = 107) (Shown in Table 2 and Additional file 1: Table S1, respectively).

Table 2 Cluster analysis (ABCD groups) using the variables exacerbation number and mMRC

Using the mMRC-based classification, patients in group A (n = 61) were ~ 53% male with a median age of ~ 62y [IQR 54–68y] and scored best on CCQ total score, BOD score, QOL, activities of daily living and fatigue. Patients in group B (n = 37) were ~ 62% male with a median age of ~ 66y [IQR 60–71y]. All four GOLD stages of airflow limitation were represented in group B, ~ 11% of the patients were classified in GOLD stage I and ~ 6% in GOLD stage IV. In group C (n = 31), the patients were ~ 55% female with a median age of ~ 62y [IQR 55–67y]. They had the lowest number of smoked pack years (PY) with a median value of 28 [IQR 19–50 PY], lowest depression score and most steps per day (median value 5743 [IQR 4473–6904]). Patients in group D were ~ 58% female with a median age of ~ 61y [IQR 52–70y]. All four GOLD stages of airflow limitation were represented in group D with ~ 12% GOLD stage I and ~ 24% stage IV. They scored worst on CCQ, depressions score, BOD score, QOL, activities of daily living and fatigue.

Using the CAT-based definition, patients in group A (n = 15) were ~ 60% female with a median age of ~ 63 [IQR 59–68y]. They had the lowest FEV1 with a median value ~ 49% of predicted [IQR 42–66] and only GOLD stage II and III were represented, resp. ~ 53% and ~ 47%. Patients in group B (n = 51) were ~ 67% male with a median age of ~ 64 [56–68y]. They scored worst on activities of daily living, QOL, satisfaction relationship and had the highest number of PY with a median value of ~ 62 [IQR 51–74 PY]. The four GOLD stages of airflow limitation were represented in group B with ~ 18% GOLD stage I and ~ 4% stage IV. Patients in group C (n = 2) included one male and one female, both GOLD stage II. Patients in group D (n = 39) were ~ 54% female with a median age of ~ 62y [IQR 55–69y] and scored worst on CCQ total score, depression score and fatigue. The four GOLD stages of airflow limitation were represented in group D with ~ 21% GOLD stage I and ~ 15% stage IV.

First cluster analysis: relation to ABCD groups

We performed cluster analysis based on exacerbation numbers in conjunction with mMRC (n = 155 patients) or CAT values (available in n = 107 patients).

Using the mMRC scale questionnaires, four clusters (n = 37, n = 22, n = 42, and n = 54) were identified, which showed only limited parallels with the ABCD groups (Table 2; Fig. 2). For the individual clusters, the largest contributing fraction of one of the ABCD groups was ~ 59–87%. By contrast, the four CAT-based clusters displayed a high level of similarity with the ABCD groups: in each of the four clusters ~ 90–100% of the patients were classified as a single ABCD group (Fig. 2, Additional file 1: Table S1): in clusters 1, 3 and 4 all patients fit in group D (CAT ≥ 10 and exacerbations ≥ 2), group A (CAT < 10 and exacerbations < 2) and group B (CAT ≥ 10 and exacerbations < 2), respectively. Only for cluster 2 we found patients classified in two different ABCD groups: group D (~ 90%) and group C (~ 10%; (CAT < 10 and exacerbations ≥ 2).

Fig. 2
figure 2

Overview of the cluster analyses categorized in ABCD groups. The three different cluster analyses (resp. used variables: mMRC and exacerbation frequency, CAT and exacerbation frequency and Burgels eight clinical variables) are shown categorized in group A, B, C and D. The definition of the ABCD group in the reproduction phenotype is based on mMRC. Definition ABCD groups: Group ‘A’ includes patients with mMRC 0–1 or CAT < 10 and 0–1 exacerbation per year; group ‘B’ includes patients with mMRC ≥ 2 or CAT ≥ 10 and 0–1 exacerbation per year; group ‘C’ includes patients with ≥ 2 exacerbations or ≥ 1 exacerbation leading to hospital admission with mMRC 0–1 or CAT < 10 and group ‘D’ includes patients with ≥ 2 exacerbations or ≥ 1 exacerbation leading to hospital admission with mMRC ≥ 2 or CAT ≥ 10; mMRC: modified Medical Research Council; CAT: COPD Assessment Test

Second cluster analysis: reproducing four clinical phenotypes

In the second analysis we aimed to reproduce Burgel’s four clinical phenotypes and included 122 COPD patients from whom a complete set of the preselected variables (see Materials and Methods), was available (Fig. 1).

The PCA transformed the eight original variables (see Additional file 1: Table S2) into independent components, the first four of which contained ~ 71% of the information and had an eigenvalue > 1, which indicated that they contained more information than the average of the replaced variables. To PC1 the variables CCQ total score, BDI-PC, mMRC and exacerbations contributed the most. PC2 was predominantly based on age and BMI. PC3 reflected FEV1%pred and BMI and was independent of mMRC and number of exacerbations. Finally, PC4 was correlated with numbers of exacerbation and inversely correlated with smoking pack years and BDI-PC (see Additional file 1: Table S2).

Based on these four components, the four-cluster solution consisted of the following clusters: reproduction cluster 1 (n = 37) with ~ 54% female and median age of ~ 64y [IQR 57–71y]. These patients had the lowest FEV1% predicted ~ 39 [IQR 30–52] and lowest value of BMI ~ 21 [IQR 19–25]; reproduction cluster 2 (n = 14) with 71% male and median age ~ 55 [IQR 48–61], had the highest number of PY ~ 50 [IQR 30–66] and scored worst on depression score, CCQ and QOL. Reproduction cluster 3 (n = 32) was ~ 53% female with median age ~ 65 [58–72]. They had the lowest number of PY: ~ 22 [IQR 5–35] and the most exacerbations per year ~ 2 [IQR 1–3 PY]. Reproduction cluster 4 (n = 39) with ~ 59% male and median age ~ 63 [57–68], scored best on mMRC, CCQ, exacerbations per year, depression score and activities of daily living.

Only one of our four clusters was comparable with one of the phenotypes of Burgel: reproduction cluster three appeared to be similar to phenotype two of Burgel et al. [6]. The patient groups had a similar median age (~ 65y [IQR 58–72y] vs. 68y [IQR 60–74y]), mMRC score (1 [IQR 0.5–1.5] vs. 1 [0–1]), and BMI value (28 [25–31] vs 28.1 [25.2–31.9]). None of the three other clusters were comparable with the phenotypes of Burgel et al. All levels of mMRC and BOD scores were present across our groups. Our clusters were mainly separated by PY, BMI and depression scale (Table 3; Fig. 3). None of the four clusters we identified matched with any of the ABCD groups (Fig. 2).

Table 3 Cluster analysis reproducing four clinical phenotypes
Fig. 3
figure 3

Overview of cluster analysis replicating Burgel. In the first column Burgel’s four clinical phenotypes are shown, and in the second column our four reproduction clusters are visualized. Burgel used Hospital Anxiety and Depression (HAD) scale as depression scale whereas the reproduction cluster used the Beck Depression Inventory-Primary Care (BDI-PC) scale. Burgel used St. George’s Respiratory Questionnaire (SGRQ) to measure quality of life (QOL) and the reproduction cluster used CCQ. To improve optical comparison between the two cohorts, the BDI-PC median [IQR] of the replication clusters are re-calculated in the range of the HAD scale and the CCQ median [IQR] re-calculated in the range of the SGRQ. Re-calculation is outlined in the Methods section

Third cluster analysis: addition of six new variables based on behaviour

In the third cluster analysis, 98 patients with COPD were selected for which 14 variables were available: in addition to the eight variables described above (Table 2) we included the following six variables: QOL, fatigue, satisfaction relationship, air trapping, steps per day, and activities of daily living (Fig. 1). The first five components of the PCA contained ~ 68.5% of the information (eigenvalue > 1). Correlations between these five components and the fourteen variables are shown in Additional file 1: Table S3. Second, Ward’s cluster analysis was performed, resulting in the identification of four new clusters (n = 22, n = 26, n = 24, n = 26) that mainly differed from each other in the number of steps per day, NCSI QOL and NCSI activities of daily living (Table 4, Fig. 4). The variables age, PY, mMRC and NCSI satisfaction relationship and NCSI fatigue largely overlapped across the four clusters. Cluster 1 was defined by worse scores on NCSI QOL, NCSI satisfaction relationship, NCSI activities of daily living and NCSI fatigue. Cluster 2 was defined by the best scores on CCQ, depression scale, NCSI QOL, NCSI activities of daily living and most median steps per day 7010 [IQR 5063–9343]. Cluster 3 was characterized by the highest BMI values (33 [IQR 30–36]), low levels of NCSI QOL 11.5 [6.7–17.9] but the lowest levels of dynamic hyperinflation measured in liters decreasing inspiratory capacity after bronchodilation (− 0.12 [IQR − 0.27 to − 0.04L]). Cluster 4 was defined by the highest level of NCSI activities of daily living (20.9 [12.3–26.0]). The new clusters with extra variables were not comparable to the clusters from the second cluster analysis reproducing clinical phenotypes.

Table 4 Cluster analysis behavior variables included
Fig. 4
figure 4

Overview of cluster analysis behavioral variables included. Six new variables; NCSI QOL, NCSI satisfaction relationship, activities of daily living, fatigue, air trapping and steps per day were added to the eight clinical variables of Burgel. These added variables altered the clusters substantially and led to the formation of four clusters that mainly differed from each other on non-physiological parameters. The four behavioral clusters are represented in the X-axis


In a real-world COPD cohort, we were essentially able to identify the GOLD ABCD groups based on CAT, but only one of the four clinical phenotypes described by Burgel et al. [6]. The addition of the six new variables QOL, fatigue, satisfaction relationship, air trapping, steps per day and activities of daily living to the clinical variables, resulted in the formation of four new clusters that did not match the original clinical phenotypes. The new clusters mainly differed in QOL and physical activity, while the previously formed clusters based on clinical variables were very heterogeneous in satisfaction relationship, fatigue, QOL, air trapping, steps per day and activities of daily living.

Burgel et al. suggested it is important to apply PCA and cluster methodology to other COPD cohorts to examine whether similar, or different COPD phenotypes can be identified in different populations [6]. In our cohort, we first attempted to reproduce clinical COPD phenotypes before we added new variables. At first, we clustered based on ABCD GOLD criteria. We noticed that CAT provided a better prediction of the ABCD groups than mMRC, which may be explained by the fact that CAT is a more extensive questionnaire for dyspnea. Second, when we aimed to reproduce the clinical COPD phenotypes as identified by Burgel et al. [6], our clusters were mostly separated by PY, BMI and depression scale. These variables are very different from those that differentiated the original clusters in the report by Burgel et al. (age, airflow limitation and symptoms). It remains unclear why we could not reproduce the clusters of Burgel et al., in our study. Possible explanations include (i) unmeasured confounding factors, (ii) greater heterogeneity in the COPD population, or (iii) the use of slightly different questionnaires.

Not only Burgel et al. used cluster analysis to identify clinical COPD phenotypes. Cluster analysis was previously applied to predict the first acute COPD exacerbation [29]; clusters were identified on the basis of lung function assessment [30] and comorbidity clusters related to inflammatory markers were formed [31]. Although these studies all use similar cluster methods, the clinical variables that were used differed. Clustering on exacerbation type or comorbidities can have clinical value when the clusters are reproducible or correlate with applicable longitudinal data. A longitudinal study based on comorbidity clusters [31] was performed, to associate the changes in exercise performance and health status after pulmonary rehabilitation [32]. This study showed that none of the comorbidity clusters influence the likelihood of clinically meaningful change in exercise performance and health status following pulmonary rehabilitation. The authors conclude that comorbidities in COPD patients should not preclude patients from following pulmonary rehabilitation. Clustering on lung function assessment resulted in seven different clusters [30]. However, based on health status these clusters could not be differentiated from each other because of small differences in mMRC and CCQ [30]. These cluster analyses illustrate the heterogeneity across individual COPD cohorts and the complexity of the identification of COPD phenotypes. In another study the variables used were comparable with our study; COPD phenotypes were clustered according to levels of physical activity, body composition, health related quality of life (HRQoL) and sedentary behavior [33]. Three groups were identified. Phenotype 1 was more physically active and less sedentary compared with phenotype 2 and 3. Phenotype 2 was older and phenotype 3 had worse HRQoL and body composition. Lung function did not differ across the three phenotypes. These results are in line with our behavioral clustering results (Table 2). However, inclusion of these variables in the cluster analysis changed the previously formed clusters based on the clinical variables of Burgel (Table 3), which demonstrates that these clusters are not stable. A study in a COPD population discriminating on asthma, emphysema and chronic bronchitis symptoms, the main phenotypes were recognized by easy to obtain clinical characteristics such as smoke exposure and questionnaires on complaints [34]. In parallel to Burgel’s study, we also excluded patients diagnosed with asthma-COPD overlap syndrome, in order to prevent clustering based on smoke exposure and symptom severity. We used the study of Burgel et al. as a reference to reproduce clinical phenotypes, because of a good matching with our clinical variables and because these phenotypes focus on treatable traits instead of future risk factors [35].

The first strength of the study is that we used the same method to form the hierarchical clusters as Burgel did. Second, as we did not use specific exclusion criteria for our real-world COPD patient cohort, our results are expected to have a good external validity for patients in secondary care. All patients with COPD were diagnosed by a pulmonologist. The data we collected were routinely available in daily practice.

Some limitations need to be mentioned. It remains challenging to directly compare our clustering analysis to the clustering by Burgel. First, although the cluster methodology was identical, we did not have exactly the same variables as Burgel et al. We used two different questionnaires: the Clinical COPD Questionnaire (CCQ) instead of the St. George Respiratory Questionnaire (SGRQ), and the Beck Depression Inventory (BDI-PC) instead of the Hospital Anxiety and Depression Scale (HADS). However, the two questionnaires to measure depression—HADS and BDI-PC—are highly correlated [21, 28]. The CCQ and SGRQ are disease-specific questionnaires that measure shortness of breath, amongst others, and are also highly correlated [19, 27, 36]. Compared to the CCQ, the SGRQ is more extensive and includes QOL-related questions. It is possible that these differences in the questionnaires explain the inability to reproduce the clinical clusters, however we would expect more similarity between the clusters because the residual six variables were identical. A second limitation is the small sample size. In the second analysis, 8 variables were used for clustering in a sample of 122 patients, compared to 322 patients in the study of Burgel et al. [6]. Only patients with a complete set of variables could be included in our cluster analyses, which resulted in a small sample size. Phenotyping based on cluster analysis may improve when the number of included patients increase or more suitable variables are added. Third, there may be critical differences between the two cohorts. We included ~ 48% females, whilst Burgel et al. included ~ 23% females. Perez et al. showed that female COPD patients are younger, have lower pack-years, higher FEV1%, lower BMI and exacerbate more often [37]. Moreover, the clinical characteristics of the population of COPD patients may well differ across different medical centers. Newly formed clinical phenotypes need longitudinal validation to determine how they are associated with important clinical outcomes of disease progression or mortality, before conclusions on their clinical relevance can be drawn [3].

In our study, clinical characteristics were used in an attempt to identify clusters as a step towards tailored treatment strategies per subgroup of COPD patients. The inability to reproduce earlier reported clusters in our real-world COPD population questions the relevance of clustering approaches for clinical practice. Clinical practice seems to call for personalized medicine [38], given the heterogeneity of the COPD population even within clusters. Individual patient characteristics should be the main focus to improve clinical outcomes and minimize unnecessary side effects for individual patients with COPD [39]. In this context, it may be more productive to develop personalized medicine approaches based on treatable traits [40, 41], than on clinical phenotypic characteristics.


In this study we used statistical cluster analyses in a real-world COPD cohort to identify subgroups of patients. Hereby, patients could be divided into clusters that largely reflected the GOLD ABCD groups. By contrast, we could not reproduce the four clinical phenotypes identified by Burgel et al. in our cohort on the basis of a series of 8 variables that were essentially the same as those used by Burgel et al. [6] The addition of six new variables, air trapping, steps per day, QOL, satisfaction relationship, activities of daily living and fatigue, altered the clusters substantially and led to the formation of four clusters that were separated mainly by these behavioral parameters. We conclude that heterogeneity in the COPD population calls for a personalized medicine approach that is not based on the stratification of patients into subgroups but rather on individual characteristics.

Availability of data and materials

All data relevant to the study are included in the article or uploaded as supplementary information. The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.



Global initiative for chronic obstructive lung disease


Chronic obstructive pulmonary disease


Quality of Life


Modified Medical Research Council


Clinical COPD Questionnaire


COPD Assessment Test


Body Mass Index

FEV1 :

Forced Expiratory Volume in 1 s


Forced Vital Capacity




Interquartile range


Beck Depression Inventory for primary care


Hospital Anxiety and Depression


Beck Depression Inventory-Primary Care


St. George’s Respiratory Questionnaire


Body mass index, airflow Obstruction and Dyspnea score


  1. Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease (2020 report).; 2020.

  2. Agusti A, Calverley PM, Celli B, Coxson HO, Edwards LD, Lomas DA, et al. Characterisation of COPD heterogeneity in the ECLIPSE cohort. Respir Res. 2010;11:122.

    Article  Google Scholar 

  3. Han MK, Agusti A, Calverley PM, Celli BR, Criner G, Curtis JL, et al. Chronic obstructive pulmonary disease phenotypes: the future of COPD. Am J Respir Crit Care Med. 2010;182(5):598–604.

    Article  Google Scholar 

  4. Nikolaou V, Massaro S, Fakhimi M, Stergioulas L, Price D. COPD phenotypes and machine learning cluster analysis: a systematic review and future research agenda. Respir Med. 2020;171: 106093.

    Article  Google Scholar 

  5. Minnesota Uo. Cluster analysis: basic concepts and algorithms, chapter 8. In: Book KD, editor. Computer Science and Engineering.

  6. Burgel PR, Paillasseur JL, Caillaud D, Tillie-Leblond I, Chanez P, Escamilla R, et al. Clinical COPD phenotypes: a novel approach using principal component and cluster analyses. Eur Respir J. 2010;36(3):531–9.

    Article  Google Scholar 

  7. Burgel PR, Roche N, Paillasseur JL, Tillie-Leblond I, Chanez P, Escamilla R, et al. Clinical COPD phenotypes identified by cluster analysis: validation with mortality. Eur Respir J. 2012;40(2):495–6.

    Article  Google Scholar 

  8. Siedlinski M, Cho MH, Bakke P, Gulsvik A, Lomas DA, Anderson W, et al. Genome-wide association study of smoking behaviours in patients with COPD. Thorax. 2011;66(10):894–902.

    Article  Google Scholar 

  9. Cannon D, Buys N, Sriram KB, Sharma S, Morris N, Sun J. The effects of chronic obstructive pulmonary disease self-management interventions on improvement of quality of life in COPD patients: a meta-analysis. Respir Med. 2016;121:81–90.

    Article  Google Scholar 

  10. Donesky D, Nguyen HQ, Paul SM, Carrieri-Kohlman V. The affective dimension of dyspnea improves in a dyspnea self-management program with exercise training. J Pain Symptom Manage. 2014;47(4):757–71.

    Article  Google Scholar 

  11. O’Donnell DE, Fluge T, Gerken F, Hamilton A, Webb K, Aguilaniu B, et al. Effects of tiotropium on lung hyperinflation, dyspnoea and exercise tolerance in COPD. Eur Respir J. 2004;23(6):832–40.

    Article  CAS  Google Scholar 

  12. van den Akker EF, ‘vantHul AJ, Chavannes NH, Braunstahl GJ, van Bruggen A, Rutten-van Molken MP, et al. Development of an integral assessment approach of health status in patients with obstructive airway diseases: the CORONA study. Int J Chron Obstruct Pulmon Dis. 2015;10:2413–22.

    Article  Google Scholar 

  13. Brusasco V, Crapo R, Viegi G, American Thoracic S, European RS. Coming together: the ATS/ERS consensus on clinical pulmonary function testing. Eur Respir J. 2005;26(1):1–2.

    Article  CAS  Google Scholar 

  14. Graham BL, Steenbruggen I, Miller MR, Barjaktarevic IZ, Cooper BG, Hall GL, et al. Standardization of Spirometry 2019 Update. An Official American Thoracic Society and European Respiratory Society Technical Statement. Am J Respir Crit Care Med. 2019;200(8):e70–88.

    Article  Google Scholar 

  15. Lahaije AJ, Willems LM, van Hees HW, Dekhuijzen PN, van Helvoort HA, Heijdra YF. Diagnostic accuracy of metronome-paced tachypnea to detect dynamic hyperinflation. Clin Physiol Funct Imaging. 2013;33(1):62–9.

    Article  Google Scholar 

  16. (WHO) WHO.

  17. Fletcher C. Standardised questionnaire on respiratory symptoms: a statement prepared and approved by the MRC Committee on the Aetiology of Chronic Bronchitis (MRC breathlessness score). BMJ. 1960;2:1665.

    Google Scholar 

  18. Bestall JC, Paul EA, Garrod R, Garnham R, Jones PW, Wedzicha JA. Usefulness of the Medical Research Council (MRC) dyspnoea scale as a measure of disability in patients with chronic obstructive pulmonary disease. Thorax. 1999;54(7):581–6.

    Article  CAS  Google Scholar 

  19. van der Molen T, Willemse BW, Schokker S, ten Hacken NH, Postma DS, Juniper EF. Development, validity and responsiveness of the Clinical COPD Questionnaire. Health Qual Life Outcomes. 2003;1:13.

    Article  Google Scholar 

  20. Kon SS, Dilaver D, Mittal M, Nolan CM, Clark AL, Canavan JL, et al. The Clinical COPD Questionnaire: response to pulmonary rehabilitation and minimal clinically important difference. Thorax. 2014;69(9):793–8.

    Article  Google Scholar 

  21. Steer RA, Cavalieri TA, Leonard DM, Beck AT. Use of the Beck Depression Inventory for Primary Care to screen for major depression disorders. Gen Hosp Psychiatry. 1999;21(2):106–11.

    Article  CAS  Google Scholar 

  22. De DA. Nederlandse versie van de Beck Depression Inventory-second edition (BDI-II-NL); handleiding. Enschede: The Psychological Corporation; 2002.

    Google Scholar 

  23. Celli BR, Cote CG, Marin JM, Casanova C, Montes de Oca M, Mendez RA, 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.

    Article  CAS  Google Scholar 

  24. Peters JB, Daudey L, Heijdra YF, Molema J, Dekhuijzen PN, Vercoulen JH. Development of a battery of instruments for detailed measurement of health status in patients with COPD in routine care: the Nijmegen Clinical Screening Instrument. Qual Life Res. 2009;18(7):901–12.

    Article  Google Scholar 

  25. Ward JH. Hierarchical grouping to optimize an objective function. J Am Stat Assoc. 1963;58:236–44.

    Article  Google Scholar 

  26. Jolliffe IT, Cadima J. Principal component analysis: a review and recent developments. Philos Trans A Math Phys Eng Sci. 2016;374(2065):20150202.

    Google Scholar 

  27. Jones PW, Quirk FH, Baveystock CM, Littlejohns P. A self-complete measure of health status for chronic airflow limitation. The St. George’s Respiratory Questionnaire. Am Rev Respir Dis. 1992;145(6):1321–7.

    Article  CAS  Google Scholar 

  28. Zigmond AS, Snaith RP. The hospital anxiety and depression scale. Acta Psychiatr Scand. 1983;67(6):361–70.

    Article  CAS  Google Scholar 

  29. Yoon HY, Park SY, Lee CH, Byun MK, Na JO, Lee JS, et al. Prediction of first acute exacerbation using COPD subtypes identified by cluster analysis. Int J Chron Obstruct Pulmon Dis. 2019;14:1389–97.

    Article  CAS  Google Scholar 

  30. Augustin IML, Spruit MA, Houben-Wilke S, Franssen FME, Vanfleteren L, Gaffron S, et al. The respiratory physiome: Clustering based on a comprehensive lung function assessment in patients with COPD. PLoS ONE. 2018;13(9): e0201593.

    Article  Google Scholar 

  31. Vanfleteren LE, Spruit MA, Groenen M, Gaffron S, van Empel VP, Bruijnzeel PL, 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.

    Article  Google Scholar 

  32. Mesquita R, Vanfleteren LE, Franssen FM, Sarv J, Taib Z, Groenen MT, et al. Objectively identified comorbidities in COPD: impact on pulmonary rehabilitation outcomes. Eur Respir J. 2015;46(2):545–8.

    Article  Google Scholar 

  33. Xavier RF, Pereira A, Lopes AC, Cavalheri V, Pinto RMC, Cukier A, et al. Identification of phenotypes in people with COPD: influence of physical activity, sedentary behaviour, body composition and skeletal muscle strength. Lung. 2019;197(1):37–45.

    Article  CAS  Google Scholar 

  34. Negro RWD, Carone M, Cuttitta G, Gallelli L, Pistolesi M, Privitera S, et al. Prevalence and clinical features of most frequent phenotypes in the Italian COPD population: the CLIMA Study. Multidiscip Respir Med. 2021;16(1):790.

    Google Scholar 

  35. Burgel PR, Paillasseur JL, Janssens W, Piquet J, Ter Riet G, Garcia-Aymerich J, et al. A simple algorithm for the identification of clinical COPD phenotypes. Eur Respir J. 2017;50(5):1701034.

  36. Stallberg B, Nokela M, Ehrs PO, Hjemdal P, Jonsson EW. Validation of the clinical COPD Questionnaire (CCQ) in primary care. Health Qual Life Outcomes. 2009;7:26.

    Article  Google Scholar 

  37. Perez TA, Castillo EG, Ancochea J, Pastor Sanz MT, Almagro P, Martinez-Camblor P, et al. Sex differences between women and men with COPD: a new analysis of the 3CIA study. Respir Med. 2020;171: 106105.

    Article  Google Scholar 

  38. Sidhaye VK, Nishida K, Martinez FJ. Precision medicine in COPD: where are we and where do we need to go? Eur Respir Rev. 2018;27(149):180022.

  39. Jameson JL, Longo DL. Precision medicine–personalized, problematic, and promising. N Engl J Med. 2015;372(23):2229–34.

    Article  CAS  Google Scholar 

  40. Agusti A, Bel E, Thomas M, Vogelmeier C, Brusselle G, Holgate S, et al. Treatable traits: toward precision medicine of chronic airway diseases. Eur Respir J. 2016;47(2):410–9.

    Article  Google Scholar 

  41. Faner R, Agusti A. Multilevel, dynamic chronic obstructive pulmonary disease heterogeneity. A challenge for personalized medicine. Ann Am Thorac Soc. 2016;13(Suppl 2):S466–70.

    Article  Google Scholar 

Download references


Not applicable.


JV and GB received unrestricted research grants from Teva, Chiesi, Boehringer Ingelheim, AstraZeneca.

Author information

Authors and Affiliations



CZ contributed to conceptualisation, statistical analyses and manuscript writing. LG contributed to conceptualisation, review statistical analyses and manuscript review. AW and GB contributed to manuscript review. MR and RH contributed to conceptualization and manuscript review. JV contributed to conceptualisation, study design, data collection and manuscript review. All authors have read and approved the manuscript.

Corresponding author

Correspondence to Cathelijne M. van Zelst.

Ethics declarations

Ethics approval and consent to participate

This study was performed after approval by the Institutional Research Board of the Franciscus Gasthuis en Vlietland, Rotterdam, the Netherlands. This research was declared as outside the scope of the Medical Research Involving Human Subjects Act, because routinely collected health care data were used after pseudonymization. This waives the requirement to obtain informed consent.

Consent for publication

Not applicable.

Competing interests

CZ: No competing interests related to the topic. LG: No competing interests related to the topic. AW: No competing interests related to the topic. GB: No competing interests related to the topic. RH: No competing interests related to the topic. MR: No competing interests related to the topic. JV: No competing interests related to the topic.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1: Table S1.

Cluster analysis (ABCD groups), using the variables exacerbation number and CAT. Table S2. Principal component analysis of clinical variables. Table S3. Principal component analysis behavior variables included.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

van Zelst, C.M., Goossens, L.M.A., Witte, J.A. et al. Stratification of COPD patients towards personalized medicine: reproduction and formation of clusters. Respir Res 23, 336 (2022).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: