Phenotypic Clustering of Idiopathic Pulmonary Arterial Hypertension: Insights Into Pulmonary Vascular and Cardiometabolic Co‐Morbidity Trajectories

ABSTRACT

Idiopathic pulmonary arterial hypertension (IPAH) exhibits significant clinical heterogeneity, necessitating a precision medicine approach. This study aimed to identify distinct IPAH phenotypes using machine learning‐based clustering and to evaluate their longitudinal therapeutic responses and long‐term survival. We analyzed 297 IPAH patients using hierarchical agglomerative clustering based on baseline demographics, hemodynamics, and comorbidities. The cohort was characterized by two distinctive yet overlapping phenotypic patterns: Cluster 1 (n = 198, “Pulmonary Vascular Phenotype”), comprising younger patients (median age: 40) with low comorbidity burden but severe hemodynamic impairment, and Cluster 2 (n = 99, “Cardiometabolic Phenotype”), consisting of older, predominantly female patients (median age: 65) with high prevalence of obesity (78%), diabetes (73%), and hypertension (57%).

Despite comparable baseline multiparametric risk scores, Cluster 1 exhibited more impaired echocardiographic right ventricular‐arterial coupling, lower left ventricular filling pressure and higher pulmonary vascular resistance (p < 0.001). Triple combination therapy was significantly more frequent in Cluster 1 (41.6% vs. 25.5%, p = 0.048). Linear mixed‐effects modeling over 12 months demonstrated that Cluster 1 achieved superior functional recovery (6‐min walk distance increase: 75 m vs. 42 m, p < 0.001) and more profound neurohormonal stabilization (Final NT‐proBNP: 55 vs. 145 pg/mL, p = 0.017). However, during a median follow‐up of 763 days, 5‐year survival was similar between groups (Log‐rank p = 0.452), and cluster membership was not associated with mortality in adjusted Cox regression (HR: 0.71; 95% CI: 0.47–1.07; p = 0.099). These findings suggest that while vascular phenotypes demonstrate more robust responses to therapy, mortality is driven by distinct factors—hemodynamic severity in younger patients versus metabolic frailty in the comorbid population.

1. Introduction

Idiopathic pulmonary arterial hypertension (IPAH), historically observed in young, otherwise healthy individuals, is increasingly diagnosed in older patients with multiple cardiometabolic and pulmonary comorbidities [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. In real clinical practice, comorbidities can influence the onset, progression, treatment response, and prognosis of PAH. Cardiometabolic comorbidities include systemic hypertension (HT), obesity, type 2 diabetes mellitus (DM), dyslipidemia, atrial fibrillation (AF), left atrial enlargement due to valvular disease or heart failure with preserved ejection fraction (HFpEF) and coronary artery disease (CAD) [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]. Metabolic syndrome, diabetes and obesity impairing nitric oxide signaling, chronic diseases and systemic inflammation promoting inflammatory cytokines may result in endothelial dysfunction and vascular remodeling. Furthermore, systemic hypertension, coronary artery disease, and metabolic abnormalities may trigger a maladaptive right ventricular (RV) response to pulmonary pressure mismatch [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]. This phenotype is sometimes referred to as “PAH with cardiometabolic phenotype” [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]. Moreover, respiratory diseases including chronic obstructive pulmonary disease, interstitial lung disease, sleep‐disordered breathing/obstructive sleep apnea or prior pulmonary embolism can coexist with or complicate PAH, and may cause controversies in differentiating true PAH from Group 3 PH due to lung disease [19, 20, 21, 22, 23, 24, 25]. Patients with systemic sclerosis–associated PAH often have more comorbidities and worse outcomes [1, 2, 3, 4]. However, currently available PH guidelines recommend initial monotherapy in PAH patients with cardiovascular comorbidities based on the limited evidence for combination therapy in this growing population [2, 3, 4, 5]. A precision medicine perspective with reliable treatment goals taking into account PAH etiology, non‐modifiable and modifiable risk factors including cardiovascular and pulmonary comorbidities on individual basis yet to be clearly identified, and advanced statistical methods such as Bayesian analytic or neural networks may provide new insights in tailoring the management strategies [2, 3, 4, 5].

In this single‐center study aimed to identify IPAH phenotypes using machine learning (ML)‐based cluster analysis and to evaluate their differences in terms of baseline clinical, neurohumoral, echocardiographic end hemodynamic characteristics, response to targeted therapies and long‐term survival along their phenotypic axes.

2. Methods

2.1. Study Population

This study retrospectively evaluated 297 patients with IPAH selected from a cohort of 1213 patients enrolled in the single‐center EvalUation of Pulmonary Hypertension Risk Factors AssociaTEd with Survival (EUPHRATES) study.

The diagnostic algorithms, hemodynamic confirmation, clinical sub‐classification of PH, and definitions of incident and prevalent PAH have been based on the recommendations of the European Society of Cardiology (ESC) and European Respiratory Society (ERS) 2015 and 2022 PH guidelines, depending on the time of the first hemodynamic confirmation [1, 2, 3, 4, 5]. For the hemodynamic definition of PH by right heart catheterization, cut‐off values of mean pulmonary arterial pressure (PAMP) > 25 mmHg and > 20 mmHg were adopted before and after the ESC/ERS 2022 PH guidelines, respectively [1, 2, 3]. For the diagnosis of pre‐capillary pulmonary hypertension, pulmonary arterial wedge pressure (PAWP) ≤ 15 mmHg and pulmonary vascular resistance (PVR) > 3 and > 2 Wood units were used as criteria before and after the ESC/ERS 2022 PH guidelines, respectively [1, 2, 3].

Longitudinal changes in WHO functional class (FC), 6‐min walk distance (6MWD), N terminal pro brain natriuretic peptide (NT‐proBNP), Multiparametric risk scores (MRS) and echocardiographic parameters including tricuspid regurgitation (TR), tricuspid annular planar systolic excursion (TAPSE) and tissue Doppler systolic velocity (S'), pulmonary arterial systolic and mean pressures (PASP, PAMP), TAPSE/PASP ratio, right atrial (RA) area and mean pressure estimates were evaluated throughout the follow‐up period [1, 2, 3, 4].

Multiparametric risk scores (MRS) included three‐strata risk prediction model from the 2022 ESC/ERS Guidelines for PAH, Swedish PAH Registry (SPAHR), the Comparative Prospective Registry of Newly Initiated Therapies for Pulmonary Hypertension (COMPERA) registry and COMPERA 2.0 four‐strata, and the French Pulmonary Hypertension Network (FPHN) registry low‐risk models, as well as the The Registry to Evaluate Early and Long‐Term PAH Disease Management (REVEAL), REVEAL 2.0 and its abridged 6‐component REVEAL Lite 2.0 scores [26, 27, 28, 29, 30, 31, 32, 33, 34].

All patients under regular follow‐up were informed. The study protocol was reviewed and approved by the Institutional Ethics Committee. This study was conducted in accordance with the Declaration of Helsinki.

2.2. Statistical Analysis

Continuous variables were expressed as mean ± standard deviation (SD) or median [interquartile range (IQR)] based on the normality of distribution, assessed via the Shapiro‐Wilk test. Categorical variables were presented as counts and percentages. To handle missing clinical data, a random forest‐based multiple imputation technique (missForest) was employed. Imputation was performed separately for numeric (missingness < 10%) and categorical (missingness < 5%) variables, strictly excluding the clinical outcome (mortality) from the process to prevent bias. Prior to clustering, a comprehensive correlation analysis was performed among all clinical, laboratory, and hemodynamic parameters to handle multicollinearity. For pairs of variables with a high correlation coefficient (Pearson's r or Spearman's rho > 0.7), the parameter with the highest clinical relevance and prognostic importance was prioritized for the model. Following this feature selection process, a total of 23 variables (comprising 14 continuous and 9 categorical parameters) were subjected to the final clustering analysis (Supporting material). All continuous variables were standardized using Z‐score normalization to ensure equal weighting across different units of measurement. Given the mixed‐type nature of the dataset, a multi‐algorithmic exploratory approach was used to determine the most stable clinical phenotypes. We evaluated and compared the performance of K‐prototypes, Partitioning Around Medoids (PAM), and Hierarchical Clustering.

The optimal number of clusters (k) was determined through a consensus of the Gap Statistic, Average Silhouette Method, and the Elbow Method (Supplementary Figure 1 and 2). To account for the mixed‐type nature of the clinical dataset (including both continuous and categorical variables), Gower's distance was utilized to calculate the dissimilarity matrix. Agglomerative Hierarchical Clustering was then performed using Ward's minimum variance method to identify distinct clinical phenotypes (Supplementary Figure 3).

To validate cluster robustness against imputation or cohort artifacts, two sensitivity analyses were performed: a complete‐case validation excluding the missing variables and a bootstrap resampling framework evaluated via the Adjusted Rand Index (ARI). Furthermore, to objectively quantify the clinical drivers behind the identified cluster partitions, Random Forest (RF) and LASSO regression algorithms were utilized for feature selection to compute variable importance metrics (Supplementary Figure S4 and S5). Differences between the identified clusters were evaluated using the Independent Samples T‐test or Mann‐Whitney U test for continuous variables, and the Chi‐square or Fisher's Exact test for categorical variables. Survival trajectories for each cluster were estimated using the Kaplan‐Meier method and compared via the Log‐rank test. To evaluate clinical and hemodynamic trajectories (Baseline, 3, 6, 9, 12 months, and Final Follow‐up), Linear Mixed‐Effects (LME) models were constructed for key parameter, including a Cluster × Time interaction term to assess differences in treatment response rates. The monthly slope (β) represented the average rate of change in the reference group (Cluster 1), while the interaction p‐value assessed the statistical difference in slopes between the two clusters. To determine the prognostic value of parameters, Cox Proportional Hazards Regression models were constructed. All statistical analyses were performed using R Statistical Software (version 4.3.2) with the following packages: missForest (imputation), clustMixType and factoextra (clustering), lme4 (LME modeling), survival and survminer (survival analysis), and ggplot2 (visualizations).

3. Results

Mean age was 49.9 ± 20.5 years, and 75.4% of patients were female. Hierarchical clustering identified two distinctive clinical patterns across the cohort spectrum. Cluster 1 (n = 198, “Pulmonary Vascular Phenotype”) was characterized by younger patients (median age: 40 years) with a low comorbidity burden despite more severe baseline hemodynamic impairment (higher PVR and mPAP) (Figure 1). Cluster 2 (n = 99, “Cardiometabolic Phenotype”) consisted of older patients (median age 65) with a high prevalence of obesity (78%), diabetes (73%), and hypertension (57%). Longitudinally, Cluster 1 showed a superior functional response to PAH therapy, with a mean 6MWD increase of 75 m compared to 42 m in Cluster 2 (p < 0.001). Cluster 1 also achieved significantly lower NT‐proBNP levels at final follow‐up (55 vs. 145 pg/mL, p = 0.017) (Figure 2). Despite these phenotypic differences, 5‐year survival was similar between groups (Log‐rank p = 0.452), and cluster membership was not associated with mortality in either unadjusted or adjusted Cox regression analyses (Figure 3).

Figure 1.

Clinical Phenotypes of IPAH: Patient Characteristics and Final Treatment: Comprehensive phenotypic characterization of the two identified IPAH clusters. Red bars/boxes represent Cluster 1 (Pulmonary Vascular Phenotype), and blue bars/boxes represent Cluster 2 (Cardiometabolic Phenotype). (Upper Panel) Box plots illustrating significant differences in continuous variables at baseline, including Age, Body Mass Index (BMI), Pulmonary Vascular Resistance (PVR), and 6‐min Walk Distance (6‐MWD). (Middle Panel) Bar plots demonstrating the prevalence of key comorbidities across clusters, highlighting the markedly higher burden of obesity, diabetes mellitus, hypertension, and coronary artery disease in Cluster 2. (Lower Panel) Distribution of the total number of comorbidities per patient and the final treatment status (monotherapy, dual, or triple combination therapy) in each cluster.

Figure 2.

Longitudinal Functional, Hemodynamic, and Biochemical Trajectories: Comparative longitudinal assessment of treatment response between Cluster 1 (red) and Cluster 2 (blue). (A) Longitudinal change in 6‐min walk distance (6‐MWD) from baseline to the final follow‐up. (B) Box‐and‐whisker plots comparing 6‐MWD at baseline and final visit, illustrating the absolute functional gain in both groups. (C) Distribution of NYHA functional class over time (Baseline, 3, 6, 9, 12 months, and Final Follow‐up), demonstrating a more pronounced functional shift in Cluster 1. (D) Longitudinal evolution of the TAPSE/PASP ratio as a surrogate for right ventricular‐arterial coupling. (E) Longitudinal reduction in echocardiographic pulmonary artery systolic pressure (PASP). (F) Longitudinal change in NT‐proBNP levels.

Figure 3.

Long‐term Survival Analysis by Phenotypic Clusters: Kaplan‐Meier survival curves and Cox proportional hazards regression analysis for the identified IPAH phenotypes. Five‐year survival probabilities compared between Cluster 1 and Cluster 2 using the log‐rank test. Despite significant baseline clinical and hemodynamic differences, overall survival was statistically similar between the two groups (p = 0.452). Inset results of Cox proportional hazards regression models. Cluster membership was not associated with a significantly higher risk of mortality in either univariate analysis.

Cluster 2 versus Cluster 1 was characterized by a female predominance (p < 0.012), an older age, higher body mass index (BMI) and body surface area (BSA), a lower 6MWD (p < 0.001, for all), and a lower rate of triple targeted combination therapies (p = 0.048) (Table 1). However, FC, serum NT‐proBNP level were comparable between clusters (p > 0.05).

Table 1.

Baseline Clinical, Hemodynamic, and Functional Characteristics of the IPAH Cohort Stratified by Phenotypic Clusters.

Variable	Overall (N = 297)	Cluster 1 (n = 198)	Cluster 2 (n = 99)	p‐value
Demographics & clinical variables
Age (years)	53.6 [33.6–65.9]	40.1 [29.8–58.1]	64.6 [56.6–72.4]	< 0.001
Female sex, n (%)	224 (75.4%)	140 (70.7%)	84 (84.8%)	0.012
BMI (kg/m2)	27.2 [24.8–31.4]	25.6 [24.4–27.6]	32.5 [30.6–34.7]	< 0.001
BSA (m2)	1.8 [1.7–1.9]	1.8 [1.7–1.8]	1.9 [1.9–2.0]	< 0.001
Final treatment status				0.048
Monotherapy	75 (25.1%)	45 (22.8%)	29 (29.6%)
Dual therapy	115 (38.7%)	70 (37.5%)	44 (44.9%)
Triple therapy	108 (36.3%)	82 (41.6%)	25 (25.5%)
Functional & risk parameters
6‐MWD (m)	250.0 [120–348]	275.0 [121–360]	180.0 [100–300]	< 0.001
NYHA Class				0.093
II	19 (6.5%)	17 (8.6%)	2 (2.1%)
III	184 (62.1%)	119 (59.9%)	66 (66.7%)
IV	93 (31.4%)	62 (31.5%)	31 (31.2%)
NT‐proBNP (pg/mL)	436.0 [191–1225]	372.7 [142–1280]	633.0 [283–1172]	0.199
REVEAL 2.0 Lite	8.5 [7.0–10.0]	8.5 [7.0–10.0]	8.5 [7.0–10.0]	0.625
REVEAL 2.0	9.0 [7.0–11.0]	9.0 [7.0–11.2]	9.0 [7.0–11.0]	0.365
FPHN				0.299
0	173 (58.2%)	117 (59.1%)	56 (56.6%)
1	109 (36.7%)	68 (34.3%)	41 (41.4%)
2	8 (2.7%)	7 (3.5%)	1 (1.0%)
3	7 (2.4%)	6 (3.0%)	1 (1.0%)
Echocardiographic data
TR Vmax (m/s)	4.1 [3.7–4.7]	4.5 [3.9–4.9]	3.8 [3.5–4.0]	< 0.001
TR Grade				0.09
None/trace	1 (0.3%)	1 (0.5%)	0 (0.0%)
Mild	69 (23.4%)	47 (23.7%)	22 (22.7%)
Moderate	91 (30.6%)	56 (28.4%)	35 (35.1%)
Severe	78 (26.1%)	47 (23.7%)	31 (30.9%)
IVC (cm)	1.9 [1.7–2.3]	1.9 [1.6–2.2]	1.8 [1.7–2.3]	0.968
PASP (mmHg)	78.0 [63.0–97.0]	90.0 [70.0–101.0]	65.0 [55.0–75.0]	< 0.001
PAMP (mmHg)	51.0 [44.0–63.5]	55.0 [46.2–66.8]	46.0 [39.8–52.0]	< 0.001
RA Pressure (mmHg)	10.0 [5.0–10.0]	10.0 [5.0–10.0]	7.0 [5.0–10.0]	0.036
RA Area (cm2)	22.0 [17.5–26.0]	22.0 [17.2–26.4]	22.0 [17.6–26.0]	0.987
TAPSE (cm)	1.8 [1.4–2.2]	1.6 [1.4–2.1]	2.0 [1.7–2.3]	< 0.001
RV S' (cm/s)	11.2 [9.3–13.5]	11.0 [9.0–13.0]	12.0 [10.0–14.8]	0.004
TAPSE/PASP ratio	0.2 [0.2–0.3]	0.2 [0.1–0.3]	0.3 [0.2–0.4]	< 0.001
Pericardial Effusion, n (%)	53 (17.7%)	45 (22.6%)	8 (8.2%)	0.004
Right heart catheterization
PAMP (mmHg)	50.0 [38.0–63.0]	57.5 [46.8–69.0]	38.0 [33.0–43.8]	< 0.001
PASP (mmHg)	82.0 [61.0–101.0]	93.0 [74.5–110.0]	62.0 [53.0–74.5]	< 0.001
PADP (mmHg)	30.5 [22.0–41.2]	36.0 [27.5–45.0]	22.0 [19.0–26.0]	< 0.001
RAP (mmHg)	8.0 [5.0–12.0]	8.0 [5.0–12.0]	8.0 [5.0–11.0]	0.284
PVR (WU)	8.7 [5.5–14.0]	11.4 [8.1–17.0]	5.5 [4.0–6.9]	< 0.001
PVR/SVR ratio	0.4 [0.3–0.6]	0.5 [0.4–0.7]	0.3 [0.2–0.3]	< 0.001
MvO2 (%)	61.5 [55.0–68.0]	61.8 [53.0–68.0]	61.0 [57.0–67.8]	0.355
Cardiac Index (L/min/m2)	2.3 [2.0–2.7]	2.2 [1.8–2.6]	2.4 [2.1–2.8]	0.006
LVEDP (mmHg)	12.0 [10.0–14.0]	12.0 [9.0–14.0]	12.0 [10.0–14.0]	0.032
Comorbidities, n (%)
Obesity	92 (31.0%)	15 (7.6%)	77 (77.8%)	< 0.001
Diabetes mellitus	75 (25.3%)	3 (1.5%)	72 (72.7%)	< 0.001
Hypertension	64 (21.5%)	8 (4.0%)	56 (56.6%)	< 0.001
Coronary artery disease	28 (9.4%)	3 (1.5%)	25 (25.3%)	< 0.001
Atrial fibrillation	9 (3.0%)	3 (1.5%)	6 (6.1%)	0.073
Chronic lung disease	15 (5.1%)	7 (3.5%)	8 (8.1%)	0.160
Number of comorbidities				< 0.001
0	163 (54.9%)	163 (82.3%)	0 (0.0%)
1	47 (15.8%)	31 (15.7%)	16 (16.2%)
2	41 (13.8%)	4 (2.0%)	37 (37.4%)
3	30 (10.1%)	0 (0.0%)	30 (30.3%)
4	16 (5.4%)	0 (0.0%)	16 (16.2%)

Abbreviations: 6‐MWD, 6‐min walk distance; BMI, body mass index; BSA, body surface area; CAD, coronary artery disease; DM, diabetes mellitus; FPHN, French Pulmonary Hypertension Network; HT, systemic hypertension; IPAH, idiopathic pulmonary arterial hypertension; IVC, inferior vena cava; LVEDP, left ventricular end‐diastolic pressure; MvO2, mixed venous oxygen saturation; NT‐proBNP, N‐terminal pro‐brain natriuretic peptide; NYHA, New York Heart Association; PADP, pulmonary artery diastolic pressure; PAMP, pulmonary artery mean pressure; PASP, pulmonary artery systolic pressure; PVR, pulmonary vascular resistance; RA, right atrium; RAP, right atrial pressure; REVEAL, Registry to Evaluate Early and Long‐term PAH Disease Management; RV S', right ventricular tissue Doppler systolic velocity; SVR, systemic vascular resistance; TAPSE, tricuspid annular plane systolic excursion; TR Vmax, tricuspid regurgitant velocity maximum; WU, Wood units.

Apart from obesity, the prevalence of DM, HT, and CAD, as well as the total comorbidity count, were significantly higher in Cluster 2 (p < 0.001), while AF and chronic lung disease rates were comparable (Table 1 and Figure 1). Moreover, Cluster 1 versus Cluster 2 related to significantly higher tricuspid regurgitant velocity, PASP and PAMP (p < 0.001, for all) and RA pressure estimates (p = 0.036), lower TAPSE, (p < 0.001), S' (p = 0.004) and TAPSE/PASP ratio (p < 0.001), and a higher rate of pericardial effusion (p = 0.004) on echocardiographic assessment.

Comparison of hemodynamic measures revealed that Cluster 1 versus Cluster 2 related to significantly PASP, PAMP, PVR and PVR/SVR ratio (p < 0.001, for all) and a lower cardiac index (p = 0.006). PAWP was significantly higher in Cluster 2 (p = 0.032) whereas RA pressure and mixed venous oxygen saturation were not different between clusters (p > 0.05) (Table 1). Baseline REVEAL 2.0, REVEAL 2.0 lite and FPHN scores were similar between Clusters (p > 0.05) (Figure 4).

Figure 4.

Principal Component Analysis of IPAH Phenotypes: Principal Component Analysis (PCA) biplot visualizing the clinical and hemodynamic distribution of patients and the contribution of individual variables to the phenotypic separation. Individual patient data are represented by dots, color‐coded by cluster membership (Red: Cluster 1; Blue: Cluster 2). The spatial separation along the first two principal components (PC1 and PC2) confirms the distinct clinical signatures of the identified phenotypes. Vector arrows (eigenvectors) represent the underlying clinical variables. The length of each vector is proportional to its contribution (eigenvalue/importance) to the variance explained by the principal components, while the direction and angle between vectors indicate the strength and nature of correlations between variables. Vectors pointing in the same direction represent positively correlated parameters, whereas those in opposite directions represent inverse relationships.

To validate the stability and clinical drivers of this two‐cluster topology, sensitivity analyses and ML feature selection were executed. First, a complete‐case analysis (n = 186) re‐clustered 89.8% of patients into their exact original assignments, demonstrating an ARI of 0.6572 (Supplementary Figure S7). Second, a bootstrap resampling framework confirmed this partition stability, yielding a mean bootstrap ARI score of 0.6572 (95% CI: [0.4370, 0.8054]; Supplementary Figure S6). Finally, RF and LASSO feature selection algorithms identified BMI, DM, HT, PVR, and mPAP as the most powerful clinical variables driving the phenotypic classification (Supplementary Figure S4 and S5).

Longitudinal assessment revealed distinct baseline profiles and functional trajectories between the two clusters (Table 2). At baseline, Cluster 1 demonstrated significantly higher functional capacity (6MWD: 261.2 vs. 190.7 m; p < 0.001) but presented with more severe hemodynamic overload, characterized by higher PASP (88.0 ± 25.8 vs. 65.9 ± 16.9 mmHg; p < 0.001) and a more impaired echocardiographic RV‐PA coupling (TAPSE/PASP: 0.22 ± 0.11 vs. 0.32 ± 0.11; p < 0.001) compared to Cluster 2. Following treatment initiation, LME modeling demonstrated significant longitudinal improvements in Cluster 1 across all key parameters: 6MWD (+ 7.09 m/month, p < 0.001), PASP (−0.76 mmHg/month, p < 0.001), TAPSE/PASP ratio (+ 0.008 units/month, p < 0.001), and NT‐proBNP (βlog10 = −0.017, p = 0.024). Notably, the rate of clinical improvement (slope) was statistically similar between the two phenotypes for PASP, TAPSE/PASP, and NT‐proBNP (all p_interaction > 0.05), indicating a uniform longitudinal treatment response across clusters despite their diverging baseline characteristics (Supplementary Tables).

Table 2.

Longitudinal Changes in Clinical and Hemodynamic Parameters and Comparative Treatment Response Rates Between Phenotypic Clusters.

Parameter	Baseline (C1 vs C2)	6 Months (C1 vs C2)	Final Visit (C1 vs C2)	Monthly Slope (β ± SE)	P‐interaction (Slope Diff)
6MWD (m)				+7.09 ± 1.01*	0.085
Cluster 1	246.5 ± 146.0	350.5 ± 127.1	335.0 ± 156.5
Cluster 2	189.8 ± 118.3	218.5 ± 124.0	235.0 ± 125.6
PASP (mmHg)				−0.76 ± 0.19*	0.819
Cluster 1	88.0 ± 25.8	86.0 ± 37.5	69.9 ± 31.7
Cluster 2	65.9 ± 16.9	62.6 ± 16.2	57.3 ± 21.2
TAPSE/PASP				+0.008 ± 0.001*	0.778
Cluster 1	0.218 ± 0.119	0.276 ± 0.154	0.356 ± 0.230
Cluster 2	0.324 ± 0.117	0.371 ± 0.141	0.429 ± 214
NT‐proBNP†				−0.017 ± 0.007*	0.781
Cluster 1	373 [142–1280]	296 [122–608]	55 [29–135]
Cluster 2	633 [283–1172]	290 [53–425]	145 [55–222]

Note: The monthly slope (β) reflects the average rate of change per month in Cluster 1, derived from linear mixed‐effects models. p‐interaction values indicate the statistical comparison of these slopes between Cluster 1 and Cluster 2.

Abbreviations: 6MWD, 6‐min walk distance; C1, Cluster 1; C2, Cluster 2; IQR, interquartile range; NT‐proBNP, N‐terminal pro‐brain natriuretic peptide; PASP, pulmonary artery systolic pressure; SD, standard deviation; SE, standard error; TAPSE, tricuspid annular plane systolic excursion.

*p < 0.05 for the longitudinal within‐cluster change over time (monthly slope) derived from linear mixed‐effects modeling.

By the final follow‐up, Cluster 1 reached a more profound state of neurohormonal stabilization with significantly lower NT‐proBNP levels (55 vs. 145 pg/mL, p = 0.017) and superior functional gains (mean increase: 75.3 vs. 42.3 m) compared to Cluster 2. However, despite these diverging longitudinal trajectories and Cluster 1's “high‐responder” profile, long‐term survival did not significantly differ between the groups in both univariate (HR: 0.868; 95% CI: 0.60–1.26; p = 0.452) and adjusted analyses (Adjusted HR: 0.711; 95% CI: 0.47–1.07; p = 0.099).

4. Discussion

Cluster analysis in our single‐center IPAH cohort identified two different yet overlapping phenotypic patterns rather than sharply distinct disease subtypes, reflecting a clinical continuum within this heterogeneous population. The first pattern, represented by Cluster 1 (“Pulmonary Vascular Phenotype”), captured a younger population with a low comorbidity burden but more compromised pulmonary hemodynamics and echocardiographic RV/PA uncoupling, and a higher rate of pericardial effusion, whereas Cluster 2 (“Cardiometabolic Phenotype”) aligned with older patients showing a female predominance and a higher prevalence of obesity, DM, HT, and CAD and higher left ventricular filling pressures. Further feature selection confirmed this binary classification, identifying BMI, DM, HT, PVR, and mPAP as the most powerful discriminators. These data verify that our clustering topography is organically shaped by a combination of host metabolic burden and underlying pulmonary hemodynamics. Although WHO FC and MRSs were comparable at baseline, triple combinations were more frequently utilized in Cluster 1. Along this clinical spectrum, Cluster 1 showed a significantly higher increase in 6MWD and significantly lower NT‐proBNP levels at the final assessment, whereas changes in other measures were comparable. Despite these phenotypic differences, the similar 5‐year survival between the groups suggests that mortality is driven by hemodynamic severity in the former and cardiometabolic frailty in the latter.

PAH is increasingly reported in older patients with multiple cardiometabolic and pulmonary comorbidities. These phenotypes are sometimes referred to as “PAH with cardiometabolic phenotype” and “PAH with lung phenotype.” [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25] In real clinical practice, comorbidities can influence the onset, progression, treatment response, and prognosis of PAH. REVEAL 2.0 risk score also incorporate age and comorbidity burden [30]. Several studies and registries including REVEAL, COMPERA Registry and French PAH Registry, GoDeep Metaregistry, and GRIPHON and AMBITION randomized clinical trial have demonstrated that patients with ≥ 3 cardiometabolic comorbidities tend to have lower exercise capacity and worse RV function at baseline, and often show an attenuated clinical and hemodynamic response to PAH combination therapies as assessed by FC, 6MWD and MRSs, echocardiographic and hemodynamic measures of RV‐PA coupling [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 31, 32, 33, 34, 35, 36, 37, 38, 39]. These differential characteristics might be attributed to contribution of post‐capillary PH component, impaired RV reserve and drug intolerance, and multidisciplinary and tailored treatments strategies rather than strict guideline algorithms should be taken into the consideration in these settings [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 31, 32, 33, 34, 35, 36, 37, 38, 39].

Current PH guidelines recommend initial monotherapy for non‐vasoreactive idiopathic, heritable or drug‐associated PAH with cardiopulmonary comorbidities whereas recommendation level for additional PAH medication has remained as Class II category on individual basis in these patients who present at intermediate or high risk of death while receiving PDE5i or ERA monotherapy [2, 4, 5]. However, real‐life data from COMPERA registry demonstrated a different landscape for combination therapies in this setting [40]. Initial PAH combination therapies were noted in 15.9% in patients with 1‐2 comorbidity and 11.1% in patients with 3‐4 comorbidities, respectively [40]. At 1 year after diagnosis, % of combination therapies increased to 37.3% in patients with 1‐2 comorbidity and 34.7% % in patients with 3‐4 comorbidities, respectively [40].

A retrospective analysis on PAH patients enrolled in the FPHN, at least one cardiovascular comorbidity was documented in 61% of patients, and 20% these had three or more comorbidities [9]. The majority of patients were female, with idiopathic/heritable/drug‐induced PAH at intermediate risk status. Mono and dual combination PAH therapies were noted in 30% and 29% of patients with 2 co‐morbidities, and in 22% and 19.5% of those with > 3 co‐morbidities, respectively [9]. Although initial dual therapy group had a worse clinical, functional, neurohumoral and hemodynamic status compared with monotherapy population, dual therapy led to larger improvements in symptoms, exercise capacity, neurohumoral and hemodynamic parameters, and risk status than initial monotherapy at early follow‐up [9]. Treatment discontinuation rates were not different, but cumulative incidence of treatment escalation over time was significantly higher in patients initiated on monotherapy as compared to in those initiated on dual therapy. Baseline PAWP 13 to 15 mmHg and diabetes independently predicted the increasing PAWP above 15 mmHg during follow‐up [9]. Absence of comorbidity versus at least 1 comorbidity was associated with a significantly higher survival, but initial treatment strategy and increase in PAWP above 15 mmHg did not relate to survival difference in the matched comorbid population [9].

The PAH with a left heart disease (LHD) was frequent in the heterogenous spectrum of Group 1 PH, and was associated with an older age, a higher BMI, a less severe pulmonary hemodynamics and a better RV adaptation [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 37, 40]. Although the PAH‐LHD patients were less commonly treated with dual PAH therapy, survival rate was not different from those in classical PAH [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 37, 40]. Only COMPERA and COMPERA 2.0 risk scores at follow‐up assessment, but not LHD phenotype, discriminated all‐cause mortality risk. Using the age, sex and diffusion capacity of the lung for carbon monoxide (DLCO), three clusters were identified in COMPERA Registry population, and Cluster 1, 2 and 3 were documented in 12.6%, 35.8% and 51.6% of them, respectively [8]. Cluster 1 characterized by median age of 45 years, 76% females, no comorbidities, mostly never smokers and DLCO ≥ 45. Cluster 2 consisted almost exclusively of elderly women with a median age of 75 years who had no smoking history but multiple other risk factors for left heart disease and DLCO mostly ≥ 45% [8]. This cluster presented with an HFpEF phenotype. The third cluster of patients was the most complex with a male predominance and frequent history of smoking, with an age comparable to those in Cluster 2 and abundant risk factors for left heart disease, and DLCO and arterial oxygen saturation % lower than in the others [8]. Cluster 1 related to a significantly higher 5‐year survival with PAH treatments, compared with others. Another matched‐pair analysis based on age, sex, WHO FC and 4‐strata risk in patients from COMPERA Registry also showed that initial combination versus monotherapy was associated with more pronounced improvements in WHO‐FC, BNP/NT‐proBNP and risk status, but not with any difference in the rates of PAH‐related hospitalizations, drug discontinuation and survival [19].

In the AMBITION trial, patients without multiple risk factors for left ventricular diastolic dysfunction (primary analysis set, PAS) and patients excluded from the primary analysis set [ex‐PAS]) were defined [18, 35]. Former versus latter group were younger, and had a greater 6MWD and fewer comorbidities [18, 35]. Initial combination therapy with ambrisentan and tadalafil compared with pooled monotherapy reduced the risk of clinical failure by 50% in PAS patients and 30% in ex‐PAS patients, and patients in the PAS had lower rates of permanent study drug withdrawal due to AE than ex‐PAS patients [18, 35]. Analyses of OPUS (prospective, observational drug registry) and OrPHeUS (retrospective, medical chart review) multicenter real‐life US studies [12], and combined data from the TRITON and REPAIR clinical trials [13], demonstrated that initial double combination therapy with macitentan plus tadalafil resulted in consistent short and long‐term efficacy, safety and tolerability in patients with PAH irrespective of the co‐morbidity status [12, 13]. EXPOSURE, GRIPHON, TRIUMPH and FREEDOM‐EV trials also revealed similar trends with selexipag, inhaled or oral treprostinil in patients with cardiovascular comorbidities [15, 31, 41]. Moreover, post hoc analysis on cumulated data from the PATENT‐1, PATENT‐2, PATENT PLUS, and REPLACE trials showed that efficacy and safety outcomes, and rates and severity of AEs requiring treatment discontinuation were consistent regardless of the comorbidity status either in the main treatment phase and along the 2‐year open‐label extension period with riociguat [17]. In consistent with these studies, we demonstrated significant differences between “pulmonary vascular” and “cardiometabolic” phenotypes in terms of clinical, hemodynamic, and echocardiographic RV‐PA coupling status, and utilization of triple combination therapies without any influence of comorbidities on 5‐year survival. These results suggest that long‐term mortality is driven by hemodynamic severity in the former and cardiometabolic frailty in latter phenotypes.

In the context of the existing literature, our findings both validate established registry trends and provide novel insights into the longitudinal behavior of these phenotypes. While COMPERA and the French Registry demonstrate that older, highly comorbid PAH patients exhibit an attenuated clinical response to combination therapies, our single‐center longitudinal data confirm that despite their limited functional reserve, these cardiometabolic phenotypes still derive substantial clinical stabilization from targeted treatments. In contrast to these large registry cluster analyses that typically report a distinct survival advantage for younger, non‐comorbid cohorts, our findings revealed no significant difference in 5‐year all‐cause mortality between the two phenotypes. This survival difference likely arises from residual treatment heterogeneity inherent to observational cohorts, a limitation of purely data‐driven unsupervised ML approaches and can only be definitively resolved through prospectively designed, standardized clinical trials. Nonetheless, our study introduces a vital caveat to the traditional paradigm: although the Pulmonary Vascular Phenotype presented with significantly worse baseline hemodynamics and severe mechanical uncoupling, they manifested a “high‐responder” trajectory. By the final follow‐up, Cluster 1 achieved a more profound state of neurohormonal stabilization (NT‐proBNP: 55 vs. 145 pg/mL) and superior functional gains (75.3 vs. 42.3 m) compared to Cluster 2. This distinct longitudinal acceleration highlights that in classical IPAH, aggressive upfront targeting of the pulmonary vasculature can effectively rescue a heavily overloaded right ventricle, whereas mortality in the older, cardiometabolic cohort is inherently driven by age‐related systemic frailty rather than pure hemodynamic failure.

The divergence in cluster numbers across major studies reflects distinct data layers and phenotypic resolutions, where macro‐clinical data capture systemic axes while omics data uncover deep pathobiological fragmentation. While high‐dimensional whole‐blood transcriptomics expands IPAH into five or more molecular clusters due to its immense biological complexity [42], the multi‐center COMPERA registry identified three clinical phenotypes [8], and our single‐center database converged on a parsimonious two‐cluster model based on core cardiopulmonary hemodynamics. Crucially, the final selection of three clusters in COMPERA was driven by clinical pragmatism and interpretability rather than strict mathematical optimization, operating within a broad registry characterized by a substantial missing data burden of up to 63% [8]. In contrast, our cohort relied on a highly complete, meticulously standardized dataset where the primary variance naturally optimized into two robust and stable systemic axes (pulmonary vascular vs. cardiometabolic).

On the other hand, there is an emerging lung phenotype characterized by a low DLCO (< 45% of the predicted value) and a smoking history in patients with PAH without overt signs of parenchymal lung disease [8, 9, 20, 21, 22, 23, 24, 25]. Patients with IPAH and a lung phenotype and patients with group 3 PH in COMPERA and in ASPIRE registries were older than those with classical IPAH of these two registries [8, 9, 22, 23]. A significant female predominance in patients with classical IPAH as compared to those in IPAH with lung phenotype and group 3 PH were noted in both COMPERA and ASPIRE registries [8, 9, 22, 23]. Improvements in WHO FC and 6MWD, and reductions in NT‐proBNP levels were more pronounced in patients with classical IPAH as compared to group of IPAH with lung phenotype and group 3 PH, respectively. The longitudinal follow‐up data from COMPERA and ASPIRE registries showed that 5‐year survival rates were significantly higher in patients with classical IPAH as compared to those in IPAH with lung phenotype, and group 3 PH, respectively [8, 9, 19, 20, 21, 22, 23]. These results seem to be consistent with the presence of a transition phenotype sharing many features with those in group 3 PH regarding the sex and age distribution, functional impairment at diagnosis, response to PAH therapies, and survival rather than classical “vascular” IPAH [8, 9, 19, 20, 21, 22, 23, 24].

5. Limitations

Several limitations should be acknowledged. First, the retrospective, single‐center design may introduce selection bias and limit the generalizability of our findings. Second, the absence of routine DLCO measurements hindered the definitive differentiation between the “Cardiometabolic” phenotype and a potential “Lung” phenotype (Group 3 PH). Furthermore, the attrition of sample size over the longitudinal follow‐up period, although handled by mixed‐effects modeling, may have influenced the statistical power of late‐term comparisons. Additionally, the lack of an external validation cohort remains a drawback; the identified clusters represent a data‐driven exploration that requires confirmation in larger, independent IPAH populations. Finally, the potential for unmeasured confounders in treatment escalation strategies across the follow‐up period must be considered when interpreting the longitudinal response rates.

6. Conclusion

IPAH patients can be phenotype‐stratified into a “classic” pulmonary vascular group and a “cardiometabolic” group. While the vascular phenotype demonstrates a more robust functional and biochemical response to therapy, the similar survival rates suggest that mortality is driven by distinct risk factors, hemodynamic severity in the younger group versus metabolic frailty in the older group.

Author Contributions

Cihangir Kaymaz: Senior clinical supervision, study oversight, conceptual guidance, interpretation of data, manuscript drafting and critical revision of the manuscript for important intellectual content. Barkin Kultursay: Conceptualization, study design, machine learning methodology, statistical analysis, data curation, interpretation of results, manuscript drafting, and critical revision of the final manuscript. Ibrahim Halil Tanboga: Methodological and statistical guidance, advanced data interpretation, and critical revision of the manuscript for important intellectual and methodological content. Nihal Ozdemir: Study oversight, clinical validation of registry data, and critical revision of the final manuscript. Hacer Ceren Tokgoz, Seda Tanyeri, Cagdas Bulus, Metehan Kibar, Can Erdem, Aziz Vezir, Ipek Akkus, Muhammet Bulut: Clinical data collection, patient follow‐up, medical record review, and verification of longitudinal hemodynamic parameters. Berhan Keskin, Aykun Hakgor, Caglar Emre Cagliyan: Data validation and quality control, and providing critical feedback on the manuscript. All authors contributed to the work, reviewed the final version of the manuscript, approved it for submission, and agreed to be accountable for all aspects of the study.

Funding

The authors have nothing to report.

Ethics Statement

Ethical approval was obtained from the Kartal Kosuyolu High Specialization Training and Research Hospital Ethics Committee (Approval no: 2013.3/4, Date: 31.05.2013). The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to the privacy of research participants. The author(s) affirm that the methods used in the data analyses are suitably applied to their data within their study design and context, and the statistical findings have been implemented and interpreted correctly. The author(s) agrees to take responsibility for ensuring that the choice of statistical approach is appropriate and is conducted and interpreted correctly as a condition to submit to the Journal. There is a researcher Dr. Barkin Kultursay who has statistical knowledge of the author team. The author(s) agrees to take responsibility for ensuring that the choice of statistical approach is appropriate and is conducted and interpreted correctly as a condition to submit to the Journal.

Conflicts of Interest

The authors declare no conflicts of interest.

Guarantor

Prof. Dr. Cihangir Kaymaz accepts full responsibility for the integrity of the data and the accuracy of the data analysis, and serves as the guarantor for this manuscript.

Supporting information

Supporting File 1

Supporting File 2

Supporting File 3

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.