Cluster analysis identifies quantitative CT imaging-based subgroups in COPD subjects with pulmonary artery enlargement

Extract

Pulmonary hypertension (PH) affects up to 60% of patients with COPD [1, 2]. When present, COPD-PH is a high-risk finding with associated mortality of up to 40% at 5 years [1, 3]. This mortality risk is multifactorial and in part reflects significant disease heterogeneity, which has limited the ability to identify effective treatments [4, 5].

Shareable abstract

Disease heterogeneity in COPD-PH has limited the ability to identify effective treatments. Chest CT is a non-invasive tool to advance disease sub-phenotyping. In COPDGene, cluster analysis identifies CT-based subgroups with differential mortality risk. https://bit.ly/4jEy5RT

To the Editor:

Pulmonary hypertension (PH) affects up to 60% of patients with COPD [1, 2]. When present, COPD-PH is a high-risk finding with associated mortality of up to 40% at 5 years [1, 3]. This mortality risk is multifactorial and in part reflects significant disease heterogeneity, which has limited the ability to identify effective treatments [4, 5].

Quantitative chest computed tomography (CT) is a non-invasive and widely available tool poised to enrich COPD-PH phenotyping. The World Symposium on Pulmonary Hypertension (WSPH) recently emphasised its use to inform clinical trial enrolment [6, 7]. The ratio of the diameters of the pulmonary artery to aorta (PA/Ao) measured on CT is a commonly referenced metric of pulmonary arterial vasculopathy in COPD which can be used to screen for elevated mean pulmonary arterial pressure (mPAP) [8–11]. This image-based biomarker of PA/Ao >1 is currently used as a metric for inclusion in COPD-PH studies (e.g. NCT05937854) [6]. However, the heterogeneity of COPD for people with this PA/Ao threshold is unknown. To explore this, we leveraged clinical, radiological and outcome data from the COPDGene Study [12]. We hypothesised that quantitative CT metrics would identify novel imaging-based COPD subgroups with possible pulmonary vascular disease, with differential mortality risk and blood-based biomarker levels supportive of emphysema and/or pulmonary vasculopathy.

In this retrospective cohort study, we identified smokers from the phase 2 visit of the COPDGene Study (2013–2017) with COPD (defined based on a forced expiratory volume in 1 s (FEV₁) to forced vital capacity ratio <0.7) and volumetric CT scanning of the chest [12]. Measures of emphysema, defined as the percentage of low attenuating tissue less than −950 Hounsfield units (%LAA-950), and pulmonary vascular pruning, defined as the ratio of volume of small vessels with cross-sectional area <5 mm² to the total pulmonary arterial blood vessel volume (aBV5/aTBV), were collected [13–15]. PA/Ao measurements were manually performed by S.W. Johnson and F.N. Rahaghi for internal validation on 384 subjects from the phase 2 visit. These measurements were then used to externally validate automated PA/Ao measurements on all phase 2 subjects using the Total Segmentator Software Platform, followed by detection of the point prior to branching of the pulmonary artery to determine diameter [16].

We subsequently restricted our analysis to people with COPD and a CT-measured PA/Ao ratio >1. Within this cohort, we applied k-means cluster analysis to four select variables. Pulmonary function test variables were chosen based on expert consensus guidance regarding the identification of patients with pulmonary vascular disease “out-of-proportion” to COPD [2, 17]. Radiographic features were selected based on their hypothesised capacity to inform differential COPD phenotypes with presumed pulmonary vascular disease [18]. The four variables selected were: FEV₁ (% predicted), diffusing capacity of the lung for carbon monoxide (D_LCO) (% predicted), %LAA-950 and aBV5/aTBV (a CT pulmonary vascular metric correlated with mPAP) [19]. To ensure stability, clustering was repeated with k-medoids and hierarchical approaches; the k-means results are reported based on the strength of their internal validation statistics [20].

Clustering was performed in Python on a complete case set; there were no missing variables. The elbow plot method and internal validation metrics, including the within-cluster sum of squares, silhouette score and Davies–Bouldin index, were collectively used to determine cluster number, k [21]. Clustering was performed 1000 times to obtain consensus clustering. Random Forest modelling was used to determine the relevant importance of each variable on overall cluster assignment [22]. Within each cluster, SHapley Additive Explanations (SHAP) values were then used to quantify the average impact of each of the four variables on cluster assignment [23].

Descriptive characteristics by cluster are reported as mean±sd or median (interquartile range (IQR)) based on their distribution. Kaplan–Meier curves were used to describe unadjusted survival risk by cluster. Multivariable Cox proportional hazard models were used to describe the association of cluster with all-cause mortality. These models were adjusted for covariates with known association with COPD mortality risk, including age, sex, 6-min walk distance (6MWD) and the modified Medical Research Council (mMRC) dyspnoea score [24]. The significance of common comorbidities, including congestive heart failure, hypertension and diabetes, were assessed, but were not significant and therefore not included. Survival analysis follow-up time was calculated from the date of the phase 2 visit. The proportional hazards assumption was assessed to ensure against violation. In an exploratory analysis, we performed t-testing of select protein expression from the SomaScan proteomic data by cluster [25]. Proteins were selected based on their clinical and prognostic relevance to pulmonary arterial hypertension (PAH) and COPD, and were as follows: sRAGE, PSMP, LKHA4, C1QR1, KISHB, PTPRD, netrin-4, thrombospondin-2, activin A, FSTL3, SVEP1, PXDN, renin, NRP1, PRDX4, NT-proBNP, IL-6, endostatin and galectin-3 [25, 26].

297 people with COPD had a CT-assessed PA/Ao >1. From this cohort, three subgroups were identified; percent emphysema was the most important variable for the determination of cluster assignment. Cluster 1 (C1) (n=100) had minimal emphysema (1%, IQR 3%) and minimal pruning (0.58±0.07). Cluster 2 (C2) (n=135) was characterised by pruning (0.45±0.07) and minimal emphysema (4%, IQR 10%). Cluster 3 (C3) (n=62) was characterised by a similar degree of pruning to C2 (0.44±0.08) but with severe emphysema (45%, IQR 14%) (figure 1). Accordingly, pruning was more important to cluster determination for C2 as compared to C3. Emphysema quantification was equally important to cluster determination for C2 and C3. Cluster 3 had the most severe expiratory airflow obstruction (FEV₁ 36.3% (IQR 15.4%) for C3 versus 78.7% (IQR 12.9%) for C1) and the lowest D_LCO (33.0% (IQR 9.6%) for C3 versus 75.1% (IQR 29.9%) for C1). People with pruning in the presence of emphysema (C3) were the oldest (69 (IQR 7) years) and had the lowest BMI (25.9±7.8 kg·m⁻² for C3 versus 28.5±0.6 kg·m⁻² for C1). This subgroup also had the lowest average 6MWD (826±535 feet) and greatest self-reported dyspnoea as assessed by mMRC dyspnoea score; 97% of subjects had mMRC ≥2.

FIGURE 1.

All-cause mortality by clusters of COPD subjects with pulmonary artery/aorta diameter ratio >1. a) In a univariate survival analysis, there was a significant difference in survival across Clusters 1 through 3 (log-rank p<0.0001) as well as between Clusters 2 and 3 (p=0.001) where loss of distal small arterial vascular volume occurred in the absence and presence of emphysema, respectively. b) Demographic and clinical features by cluster, including the four variables input into the k-means cluster analysis. BMI: body mass index; mMRC: modified Medical Research Council dyspnoea score; LABA: long-acting beta-agonist; ICS: inhaled corticosteroid; LAMA: long-acting muscarinic antagonist; FEV₁: forced expiratory volume in 1 s; FVC: forced vital capacity; D_LCO: diffusion capacity of the lung for carbon monoxide; aBV5/aTBV: ratio of small artery volume in cross section (<5 mm²) to total pulmonary artery volume; %LAA-950: percentage of low attenuating tissue less than −950 Hounsfield units. Descriptive variables are presented as mean±sd for normally distributed values and median (interquartile range) for non-normally distributed values. ^#: COPD medication use missing: Cluster 1: beta-agonist (n=1), LABA/ICS (n=3), LABA/LAMA (n=44), macrolide (n=20), phosphodiesterase inhibitor (n=20), oral steroids (n=4); Cluster 2: beta-agonist (n=1), LABA/ICS (n=1), LABA/LAMA (n=40), macrolide (n=14), phosphodiesterase inhibitor (n=14), oral steroids (n=4); Cluster 3: LABA/ICS (n=2), LABA/LAMA (n=22), macrolide (n=2), phosphodiesterase inhibitor (n=2), oral steroids (n=2).

Over a median follow up of 7.0 (IQR 3.2) years, 8% of subjects in C1 (8/100), 33.3% of subjects in C2 (45/135) and 54.8% of subjects in C3 (34/62) died. Survival was significantly worse in C3 as compared to C1 and C2 (figure 1). In survival analysis adjusted for age, sex, BMI, 6MWD and mMRC dyspnoea score, compared to C3, C1 experienced a significantly decreased risk of death (HR_C1 0.40, 95% CI 0.17–0.96; p=0.04). The mortality risk was non-significantly lower in C2 (HR_C2 0.94, 95% CI 0.57–1.57; p=0.82). In an exploratory analysis, protein tyrosine phosphatase receptor type D (PTPRD) expression was higher in C3 as compared to C2 (12.28 versus 12.02; p<0.01) and peroxiredoxin 4 (PRDX4) expression was higher in C2 as compared to C3 (10.21 versus 10.10; p=0.02).

In COPDGene, we identified three distinct subgroups of people with COPD and PA/Ao >1. These groups ranged from younger, more physiologically preserved subjects with less parenchymal lung disease (C1), to those with pruning and minimal emphysema (C2), to those with significant emphysema, more vascular pruning, and poorer functional status (C3). Clinical outcomes were in proportion to the disease severity of each cluster.

An elevated PA/Ao is a sensitive radiological metric to screen for PH in people with COPD, yet our findings suggest that there is marked heterogeneity in the people who meet that definition. These findings raise the possibility that quantitative chest CT metrics could be used to non-invasively assist in characterising phenotypes of COPD patients in cohorts with right heart catheterisation (RHC)-confirmed PH. In fact, identification of subgroups defined by radiographic and haemodynamic features may be useful for ongoing efforts to identify effective COPD-PH treatments [5, 6, 18].

We found that in people with COPD and an elevated PA/Ao, almost one-third had minimal radiological evidence of parenchymal or pulmonary vascular pruning. This group may represent pulmonary vascular disease unrelated to COPD. As nearly 50% of these subjects carried a diagnosis of systemic hypertension, pulmonary artery dilation may be driven by pulmonary venous hypertension not captured by aBV5/aTBV. Additionally, sleep apnoea was present in 18% of C1 subjects, which may cause PA dilation [27, 28]. Finally, as PA/Ao >1 is a sensitive metric to screen for pulmonary vascular disease but also determined based on ease of radiographic reporting, it is possible that our inclusion criteria captures subjects without pulmonary vascular disease who are represented in this cluster [10]. Future work inclusive of haemodynamic data is critical to interrogate these possibilities.

The remaining people fell into one of two clusters. These subgroups had similar amounts of distal vascular pruning, but differed markedly in their amount of emphysema. In fact, quantitative emphysema was the most important predictor of cluster assignment. C2 may represent a proposed COPD-PH “pulmonary vascular phenotype” characterised by muscularisation of pulmonary arterioles, whereas C3 is hypothesised to represent a subgroup characterised by emphysema-mediated vascular destruction or compression due to air trapping [17, 29, 30]. In line with this hypothesis, biomarker elevation of the emphysema-predictive protein PTPRD in C3 as compared to C2, may highlight a more COPD-centric phenotype whereas higher PRDX4 expression in C2 as compared to C3, which has been implicated in PAH, may reflect antioxidant response to abnormal vascular-centric pathology [25, 26].

In summary, these data are novel in their ability to highlight the strength of quantitative CT metrics to deconstruct COPD heterogeneity through the lens of an elevated PA/Ao ratio. Future work should not only validate these findings within similarly well-phenotyped COPD cohorts with available CT-derived metrics, but also incorporate invasive haemodynamic data within a RHC-confirmed COPD-PH population to identify subgroups defined by both haemodynamic and imaging features. In agreement with the 2024 WSPH documents, we believe quantitative CT imaging metrics are critical to future COPD-PH clinical research efforts and should be considered as a tool to enhance PH clinical trial enrolment to ultimately identify effective COPD-PH treatments.

Shareable PDF

This PDF extract can be shared freely online.

Shareable PDF

ERJ-00038-2025.Shareable