Background

Chronic obstructive pulmonary disease (COPD) is currently the third leading cause of death in the United States [1]. COPD is characterized by airflow limitation that is incompletely reversible [2], and thus it is identified by the ratio of forced expiratory volume in 1 s over forced vital capacity (FEV1/FVC) at post bronchodilator. The severity is further distinguished by FEV1% predicted values by COPD guidelines [3]. The ratio of FEV1/FVC is used as an indicator to identify COPD patients in diagnosis of the disease [3], but it may not be sensitive enough to differentiate heterogeneous alterations characterized by multiple pathologies [4]. In contrast, quantitative computed tomography (QCT) can distinguish emphysema-predominant and airway-predominant diseases [5] and help link structural and functional variables [6, 7]. Individual imaging-based metrics have been derived from both CT and MRI studies of the lungs in both COPD and asthma [8]. With recent advances in unsupervised clustering of subject populations [9,10,11], there is an increased effort to employ these methods for grouping sub-populations of subjects within both the asthma [12] and COPD communities [13,14,15,16,17].

With the introduction of novel structural and functional imaging-based metrics [6] and corrections for inter-site and inter-subject variabilities [18], Choi et al. [7] recently integrated all of the imaging-based metrics measured at multi-scales to derive imaging-based clusters of subjects from an asthma population. These clusters were significantly associated with clinical characteristics. In the present work, we utilize the same approach, but with an expanded set of variables that include an image matching-based quantification of emphysema and functional small airways disease to derive imaging-based clusters in a COPD population with meaningful associations to clinical characteristics. For this purpose we investigated a subject population from within the Subpopulations and Intermediate Outcome Measures in COPD Study (SPIROMICS) [19] which was initiated to provide robust criteria for sub-classifying COPD participants and further identify biomarkers and phenotypes for efficient conduct of treatment trials.

Methods

Human data and QCT imaging

From the first 1000 subjects recruited into SPIROMICS [19] we performed image matching and identified 700 subjects in whom total lung capacity (TLC) to residual volume (RV) matches were successful. From these subjects with matching data we chose to study current smokers falling within strata 2–4 [19] (N = 284) as well as healthy non-smokers (N = 130). SPIROMICS categorized subjects into four strata 1–4. The healthy non-smokers (stratum 1) were defined as FEV1/FVC > 0.7 with smoking status (pack-year) < 1. Smokers with (pack-year) > 20 and FEV1/FVC > 0.7 were grouped in stratum 2. Also smokers in strata 3 and 4 had FEV1/FVC < 0.7; those with FEV1 > 50% were grouped in stratum 3 whereas those with FEV1 < 50% were in stratum 4 [19]. The demographics of these populations are summarized in Table 1. The current smokers were employed to derive imaging-based COPD clusters and individual metrics were compared with the non-smoking healthy controls. We initially performed cluster analysis [20] including both former and current smokers, which resulted in less statistically stable clusters based on the Jaccard index [21] (90% and 70% for current and both former and current smokers, respectively). This suggested that smoking status introduced confounding variables, interfering with many metrics such as the emphysema index which is shifted by the effect of inflammation (associated with smoking status) on regional lung density [22].

Table 1 Demography, baseline (Pre-bronchodilator) and maximal (Post-bronchodilator) pulmonary function tests for 130 Stratum 1 (healthy), 114 Stratum 2, 131 Stratum 3 and 39 Stratum 4 subjects
Full size table

Multiscale imaging-based variables

Volumetric CT imaging was carried out during coached breath holds at TLC and RV [23], and image analysis was carried out with use of the Apollo software (VIDA Diagnostics, Coralville, Iowa).

Sixty nine post-processed imaging-based variables were employed at both segmental and lobar levels, which is an expanded set of existing 57 variables used for asthma cluster analysis [7] utilizing our multiscale imaging-based clustering approach (MICA). The four structural variables at the pre-segmental and segmental levels were extracted from ten local regions to reflect the regional characteristics [6]. These structural variables included bifurcation angle (θ), airway circularity (Cr), wall thickness (WT) and hydraulic diameter (Dh), where each variable indicated alteration of airway geometry, alteration of luminal shape, wall thickening and luminal narrowing, respectively. The dimensions of WT and Dh were normalized by predicted trachea WT and Dh from healthy controls denoted by WT* and Dh* [6]. The normalization was used for eliminating inter-subject variability due to sex age and height.

Employing a mass-preserving image registration technique [24, 25], lobar/global functional variables were further derived to describe the alterations of lung deformation between inspiration and expiration. The variables at lobar levels included fractional air volume change (ΔVairF), the determinant of Jacobian matrix (Jacobian) [26] and anisotropic deformation index (ADI) [26, 27], indicating regional contribution of ventilation (lobar fraction of air volume change between TLC and RV), regional volume change, and the degree of preferential deformation, respectively. In this study, we also employed three new variables; fraction-based small airways disease (fSAD%) to characterize small airway, fraction-based emphysema (Emph%) for emphysematous diseases as well as tissue fraction at TLC (βtissue). Emph % and fSAD% were defined based upon a variation of the image-matching-based parametric response map used by Galban et al. [28]. In our implementation, Emph% (98.5% air-fraction as the threshold) and fSAD% (90% air-fraction as the threshold) were used instead of using the density threshold identifying voxels < − 950 HU, to account for scanner variability [18]. βtissue indicates the portion of tissue volume in each voxel to assess a possible alteration of local tissue. Also, related global (whole lung) variables were included; fSAD% (Total) and Emph% (Total), apical-basal distance over ventral-dorsal distance at TLC (lung shape), the ratio of air-volume changes in upper lobes to those in middle and lower lobes between TLC and RV (U/(M + L)|v), Jacobian (Total) and ADI (Total) [26]. Therefore, we obtained 32 local structural as well as 30 lobar and 7 global variables, giving 69 imaging-based variables. These comprehensive imaging variables were then used for a cluster analysis. Full names of each variable are described in Abbreviations used section.

Clustering and statistical analysis

We compared three general clustering methods including K-means, Hierarchical [29] and Gaussian finite mixture model-based [30] fed by principal components [31, 32]. Three methods (Kaiser/Harris, Cattel Scree Test and Parallel Analysis [33]) were performed to retain an optimal number of principal components (Fig. 1). Internal properties of clusters (Connectivity, Dunn index and Silhouette indices) were used in order to find the best clustering method. The K-means clustering method showed more stability and an optimal number of clusters fitted for the structure of the imaging data (Additional file 1). The results of clustering for K-means and hierarchical are shown in Fig. 2. K-means clustering with could achieve more clear separation of the cluster membership compared to hierarchical clustering.

Fig. 1
figure 1

A scree plot for determining the optimal number of principal components

Full size image
Fig. 2
figure 2

a Clustering membership of K-means clustering on 2-D projected coordinates; (b) Clustering membership of Hierarchical clustering on 2-D projected coordinates

Full size image

Next, we performed association tests of imaging-based clusters with demographic and clinical variables to investigate the clinical relevance of current clusters. The data analysis was performed by R software (version 3.1.1). Kruskal-Wallis and chi-square tests were performed to compare differences of continuous and categorical variables, respectively. P = 0.05 was taken as the significant level in all tests. The validation of the cluster analysis was assessed by dividing the data set into training and validation sets (see the Additional file 1).

Results

Four clusters and imaging-based characteristics

The K-means clustering method produced four unique clusters, containing 96, 45, 88 and 55 subjects respectively (Table 2). Figure 3 shows the percentages of emphysema and small airway disease (Emph% and fSAD%) for the different clusters and the healthy group. Figure 4 summarizes the imaging-based characteristics of the four clusters. The major variables which best describe the four clusters were selected with a stepwise forward variable selection technique using Wilk’s λ criterion [34]. Ten major variables with higher Wilk’s λ values are presented to explain structural and functional alterations associated with each cluster (Table 2). We then performed a decision tree analysis to construct a simple predictive model (Fig. 5). The model comprising 7 discriminant variables achieved 89% accuracy in classification. These variables were Jacobian (Total), Dh* (sLLL), Dh* (sRLL), WT* (sRUL), WT* (sRML), βtissue (LLL) and fSAD% (Total).

Table 2 Major structural and functional imaging-based variables in four imaging-based clusters and heathy subjects
Full size table
Fig. 3
figure 3

a Percentage of emphysema (Emph%) for four clusters and the healthy control group (green). † P > 0.05 between clusters 1, 2, 3 and the healthy group. P < 0.05 between Cluster 4 and other groups for all pairwise comparisons (b) Percentage of small airway disease (fSAD%) for four clusters and the healthy control group (green). ‡ P < 0.05 for comparisons between four clusters 2, 3, 4 and the healthy group for all pairwise comparison. P > 0.05 for between Cluster 1 and the healthy group

Full size image
Fig. 4
figure 4

A summary of imaging and clinical variables for four clusters

Full size image
Fig. 5
figure 5

Predicting imaged-based cluster using only 7 important variables with a classification tree (“simple” imaging-based clustering). Variables are Jacobian (Total), Dh* (sLLL), Dh* (sRLL), WT* (sRUL), WT* (sRML), βtissue (LLL) and fSAD% (Total) with 89% accuracy compared with “original” imaging-based clusters using 69 variables

Full size image

Associations with demography and PFT

Association of clusters with demography and PFTs are tabulated in Table 3. Cluster 1 with normal airway structures was mostly populated by GOLD 0 and stratum 2 subjects who were relatively younger and presented with a lower BODE index compared to other clusters (P < 0.05). Unlike Cluster 1, Cluster 4 was mostly populated by GOLD 2, 3, 4 and strata 3, 4, respectively with relatively older subjects. Cluster 1 included most subjects from group A and B from ABCD assessment while subjects in Cluster 4 shifted towards group D. Cluster 2 was associated with relatively higher BMI subjects. Also Cluster 3 was associated with subjects who exhibited a relatively low BODE index. Cluster 4 subjects showed higher BODE index and were relatively older males.

Table 3 Demography, baseline (Pre-bronchodilator) and maximal (Post-bronchodilator) PFTs, in four imaging-based clusters
Full size table

Both pre-bronchodilator and post-bronchodilation PFT-derived lung function values are tabulated in Table 3. FEV1/FVC showed a consistent, decreasing pattern from Cluster 1 to Cluster 4. Subjects in Cluster 4 demonstrated significant decreases in FEV1/FVC both pre-and post- bronchodilation, while Cluster 1 showed a mean FEV1/FVC of 0.74 that is above the normal cut off value of 0.7. A similar decreasing pattern from Cluster 1 to Cluster 4 was found for FEV1 and FVC % predicted values, with the highest and lowest values associated with Cluster 1 and Cluster 4, respectively.

Associations with symptoms and disease histories

Symptoms and disease histories were collected from the SPIROMICS [19, 23] data set and are summarized in Table 4. Cluster 4 showed a higher history of chronic bronchitis, emphysema, wheezing and whistling compared to Clusters 1, 2, and 3. The prevalence of symptoms in Clusters 1, 2 and 3 was less likely than Cluster 4. Cluster 2 showed an increased history of sleep apnea diagnosed at baseline compared to other clusters. Cluster 4 had higher smoking pack-years at baseline (P < 0.05) compared to the other clusters.

Table 4 Associations with Symptoms and Disease Histories
Full size table

CAT score, activity limitation and exacerbation histories

Blood biomarkers, baseline CAT score, exacerbation histories as well as activity limitation (6-min walk) are tabulated in Table 5. While clusters did not show significant difference in blood biomarkers, there was a significant difference for CAT score between clusters (P < 0.05). The CAT score for all clusters were more than 10, suggesting respiratory symptoms (symptomatic) in even the cluster 1 subjects, in agreement with the findings of Woodruff et al. [35]. While Clusters 1, 2 and 3 showed relatively similar CAT scores, Cluster 4 showed a higher CAT score than other clusters. Severe (since entering the study), total (since entering the study), and total (at baseline) exacerbations showed significant differences between clusters with Cluster 4 having the most severe exacerbations. There was no significant difference in the number of exacerbations between Clusters 1, 2 and 3. Also subjects in Clusters 2 and 4 were more likely to have activity limitations, as their 6-min walk distance and oxygen desaturation were lower than other Clusters.

Table 5 Characteristics of biomarkers in four imaging-based clusters
Full size table

Cluster characteristics

Cluster 1: Relatively resistant smokers with preserved pulmonary function

Cluster 1 had increased smoking pack-years (41.79 ± 22.05) with no or minimal airway obstruction (FEV1/FVC = 0.74). Cluster 1 was mostly populated by GOLD stage 1 (66%) with low emphysema and low fSAD%. Cluster 1 showed that structural variables including WT*, Dh* and Cr are very close to those of healthy controls. The CAT score, BODE index and severe exacerbation history of this cluster were relatively low compared to other clusters. Cluster 1 can be considered to be relatively resistant smokers with preserved pulmonary function.

Cluster 2: Airway-wall-thickening fSAD-dominant subjects with obesity and activity limitation

Cluster 2 had increased smoking pack-years (42.89 ± 18.7) and a FEV1/FVC relatively close to the lower limit of normal, 0.7. This cluster had the highest BMI among all clusters and a higher BODE index than Clusters 1 and 3. Cluster 2 exhibited a decrease of Dh* and Cr compared to Cluster 1 and had the highest WT* and βtissue and the lowest Jacobian among all clusters. Cluster 2 also showed an increase of fSAD%, but with Emph% close to that of Cluster 1. Cluster 2 showed no significant difference in the number of exacerbations or CAT score. Cluster 2 had decreased 6-min walk distance and oxygen desaturation, similar to Cluster 4 but decreased compared to Cluster 1, (P < 0.05). Thus, Cluster 2 can be classified as thickened airway wall, narrowed airway lumen and fSAD-dominant subjects with associated obesity and activity limitations.

Cluster 3: Airway-wall-thinning fSAD-dominant subjects

Compared to Clusters 1 and 2, Cluster 3 with smoking pack-years (47.06 ± 19.39, P > 0.05) showed a continued increase of fSAD% (P < 0.05) with similar Emph% (P > 0.05). Dh* showed significant decrease as compared with Cluster 1, but not significant difference from that of Cluster 2. Also WT* decreased compared to Clusters 1 and 2 (P < 0.05). FEV1/FVC (=0.63) for Cluster 3 remained close to the normal range with no significant difference for the three categories of exacerbation (severe, total and total at baseline) between Clusters 1, 2 and 3. Cluster 3 had 58% of subjects in GOLD stages 2–4 and had a CAT score close to Clusters 1 and 2. While Cluster 3 did not show significant differences in 6-min walk distance or oxygen desaturation compared to Cluster 2, its oxygen desaturation decreased comparable to that of Cluster 2 (P < 0.05). Cluster 3 can be categorized as fSAD-dominant subjects with luminal narrowing and decreased wall thickness.

Cluster 4: Severe emphysema-fSAD-mixed subjects with severe airway luminal narrowing and wall thinning

Cluster 4 had significantly greater smoking pack-years (54.95 ± 21.03) compared to other clusters. It had a higher CAT score along with more exacerbations and greater activity limitations compared to the other clusters. Cluster 4 also showed significant elevation of emphysema and small airways disease (fSAD%↑↑ and Emph%↑↑), significant decreases in lung deformation (Jacobian↓↓ and ADI↓↓) and significant airway luminal narrowing (Dh*↓↓, P < 0.05) compared to Clusters 1, 2, and 3. Cluster 4 had significant decreases in airway wall thickness (WT*↓↓, P < 0.05) compared to Clusters 1 and 2. Cluster 4 also had a much lower FEV1/FVC for both baseline function and maximal post-bronchodilator lung function compared to the other Clusters. Cluster 4 had a higher BODE index compared to the other clusters. Lymphocyte% decreased in Cluster 4 and reached near statistical significance (P = 0.08). Therefore, Cluster 4 subjects can be classified as severe mixed emphysema-fSAD with severe luminal narrowing, decreased wall-thickness and lung function.

Discussion

In the present study, we applied MICA [7], which utilized an expanded set of 69 QCT imaging-based variables at both segmental and global scales, to derive four statistically stable clusters in SPIROMICS current smokers with unique structural and functional characteristics, and establish their associations with clinical metrics. Cluster 1 comprised relatively resistant smokers with preserved pulmonary function (FEV1/FVC > 0.7) and respiratory symptomatology (CAT> 10). Cluster 2 was characterized by airway wall thickening, fSAD-dominance, obesity and activity limitation. Cluster 3 exhibited airway wall thinning (in agreement with the findings of Smith et al. [36] and fSAD-dominance. Both Clusters 2 and 3 had FEV1/FVC close to the lower limit of normal, 0.7. Cluster 4 had mixed emphysema-fSAD with severe airway luminal narrowing, wall thinning and decreased lung function.

To better understand the differences between spirometry-based GOLD stages and imaging-based clusters, Fig. 6 shows the distributions of GOLD 0–4 stages and Clusters 1–4 of the current smokers on a parametric response map (PRM) [28]. Except Cluster 2, Clusters 1, 3 and 4 appear to align with the path of the five GOLD stages. Wan et al. [37] studied a cohort of GOPDGene subjects with post-bronchodilator preserved ratio impaired spirometry (PRISm), characterized by a reduced FEV1 (< 0.8) with a preserved FEV/FVC ratio (≥0.7). They reported that PRISm subjects exhibit increased BMI, reduced 6-min walk, increased segmental airway wall area percentage, and increased respiratory symptoms [37], resembling both imaging and clinical characteristics of our Cluster 2. Thus, although only ~ 3% of the current smokers in this study met the spirometry criteria for PRISm, Fig. 7 displays the distributions of GOLD 0–4 stages and Clusters 1–4 of the same subjects on a post-bronchodilator FEV1-FEV1/FVC map. Cluster 2 is located nearest to the PRISm quadrant defined by the above spirometry criteria, as compared to GOLD 1 and 2. While a further study on a large PRISm cohort is needed to establish the link between imaging-based Cluster 2 and PRISm, the above analysis suggests that the current approach may be able to identify a clinically meaningful sub-population with COPD as compared with spirometric classification.

Fig. 6
figure 6

PRM based on GOLD stages and imaging-based derived clusters

Full size image
Fig. 7
figure 7

FEV1 and FEV1/FVC based of GOLD stages and imaging-based clusters. Dashed lines represent fixed threshold criteria (FEV1 = 0.8, FEV1/FVC = 0.7) used to distinguish possible PRISm subjects

Full size image

Castaldi et al. [14] classified four clusters in current and former smokers from the COPDGene study using four variables (features); FEV1% predicted, CT-quantified emphysema, segmental wall area% and emphysema distribution. Their four clusters are: relatively resistant smokers (i.e., no/mild obstruction and minimal emphysema despite heavy smoking), mild upper zone emphysema-predominant, airway disease-predominant and severe emphysema, with Clusters 2 and 4 having strong genetic associations. They included a PFT measure (FEV1% predicted) as one of the input features for cluster analysis, which is different from our MICA approach employing solely imaging-based variables to identify clusters and then establish associations of derived clusters with PFTs and other clinical measures and symptoms. Their clusters appear to overlap with our clusters. For example, our Cluster 1 (or 4) is similar to their Cluster 1 (or 4). Although our Cluster 2 and 3 had relatively lower FEV1/FVC (but being close to the cut-off threshold of 0.7) than that of Cluster 1, they exhibited a significantly increased fSAD% (P < 0.05) compared to Cluster 1 without a significant increase in Emph%. Thus, our Cluster 2, which exhibited increased fSAD%, thicker airway walls, the highest BMI, high BODE index and low Emph%, may correspond to their Cluster 3 being described as airway-predominant disease, thicker airway walls, lowest average emphysema of all clusters and high BMI. In addition, our Cluster 3 showed a relatively higher upper/lower emphysema ratio than others (Table 6); being similar to their Cluster 2 characterized by mild upper zone-predominant emphysema. Castaldi et al. [17] further investigated reproducibility of clustering analysis across multiple COPD cohorts using a set of common variables, suggesting that COPD heterogeneity may be characterized as a continuous trait.

Table 6 Upper/lower zone Emph% and fSAD%
Full size table

Woodruff et al. [35] divided subjects (including both current and former smokers) from the SPIROMCS study into five categories A-E: (A) never smoked, preserved pulmonary function (B) CAT ≤10 (asymptotic); (C) CAT ≥10 (symptomatic), mild-to-moderate (GOLD stage 1 or 2); (D) CAT ≤10 and (E) CAT ≥10. The symptomatic subjects with preserved pulmonary function in category C had greater airway-wall thickness, but did not have higher Emph%, as compared with asymptotic subjects. These category-C subjects were younger with higher BMI and were more likely current smokers. These characteristics are strikingly similar to those of our Clusters 1 and 2 subjects. Cluster 1 included subjects that had thicker airway walls compared to Clusters 3 and 4, and had minimal-to-no emphysema. In addition, Cluster 2 exhibited several characteristics similar to Cluster 1, including lower symptomatology with CAT ≥10, thicker airway walls, minimal-to-no emphysema and FEV1/FVC = 0.68 (close to 0.74 for Cluster 1) as well as the highest BMI and βtissue among all clusters. Nonetheless, different from Cluster 1 but similar to Cluster 4, Cluster 2 exhibited severe activity limitations and had relatively higher fSAD% and lower Jacobian. The major difference between Clusters 2 and 4 is that Cluster 2 had the highest BMI and βtissue. This suggests that symptomatic current smoker subjects in category C with preserved pulmonary function may be further divided into two sub-groups (Clusters 1 and 2) with distinct characteristics.

Garcia-Aymerich et al. [10] identified three groups in a cohort of 342 subjects recruited for the Phenotype and Course of COPD (PAC-COPD) study in Spain, using a comprehensive set of clinical, functional, biological and imaging metrics. Groups 1, 2 and 3 had respective FEV1/FVC of 0.44, 0.57 and 0.61. In addition to milder airflow limitation, Group 3 exhibited high BMI (obesity), systemic inflammation, cardiovascular disease, diabetes and activity limitation. These characteristics appear to overlap with those of our Cluster 2. While both Clusters 2 and 3 were fSAD-dominant subjects, they were characterized by increased and decreased airway wall-thickness, respectively. Also, Sood et al. [38] suggested that higher BMI (obesity) might contribute to systemic inflammation.

Our study here has several limitations. It focused on current smokers and was a cross-sectional study. In the future, the analysis shall be extended to include former smokers and compared with the current analysis. Also, our analysis will be extended to longitudinal data and cross validation shall be performed to examine cluster transition and stability over time. As a preliminary study, we included in the Additional file 1 the cluster analysis of longitudinal data from a small COPD cohort, showing consistently four stable clusters. While our use of image matching is refined to the level of accounting for lobar slippage, it requires segmentation of the lobes at both inspiration and expiration.

Conclusions

In conclusion, using a K-means clustering method we found four distinct stable clusters of COPD subtypes. These are Cluster 1, non-severe COPD with normal airway structure (relatively resistant smoker); Cluster 2, a mix of non-severe and severe COPD with fSAD dominance, low emphysema percentage, high tissue fraction with wall thickening; Cluster 3, a mix of non-severe and severe COPD, fSAD dominance with decreased wall thickness and luminal narrowing; Cluster 4, a mix of severe fSAD and emphysema with significant alterations in functional and structural variables. A decision tree analysis with only 7 discriminant imaging-based variables allows classification with an accuracy close to the “original” cluster membership. The unique structural and functional characteristics observed in each cluster can help shed light on the existing heterogeneous nature of the disease.