Interpretable machine learning for the identification and prognostic phenotyping of concurrent HFpEF in hypercapnic COPD

Background Unrecognized heart failure with preserved ejection fraction (HFpEF) frequently complicates hypercapnic chronic obstructive pulmonary disease (COPD). The overlapping symptomatology poses significant diagnostic challenges, potentially leading to delayed interventions. Methods Clinical and echocardiographic parameters from 441 patients with hypercapnic COPD were retrospectively analyzed. Five machine learning algorithms were trained and evaluated to identify concurrent HFpEF. SHapley Additive exPlanations (SHAP) analysis was applied to assess mathematical feature associations rather than causal relationships. Additionally, unsupervised consensus clustering was utilized for clinical phenotypic stratification. Results The Random Forest model demonstrated the highest discrimination capability [(test AUC=0.940, 95% CI: 0.898–0.982; 10-times repeated 5-fold cross-validation mean AUC = 0.928 ± 0.019)]. SHAP analysis suggested complex non-linear associations, indicating a potential synergistic interaction between elevated PaCO2 and body mass index in being associated with an elevated predicted HFpEF risk. Furthermore, consensus clustering identified three distinct clinical sub-phenotypes. Subsequent survival analyses indicated that the severe cardiac dysfunction phenotype, characterized by markedly elevated NT-proBNP and substantial hypercapnia, was independently associated with a higher risk of adverse clinical trajectories (Hazard Ratio=3.20, P < 0.001). Conclusion In this exploratory single-center study, interpretable machine learning algorithms demonstrated potential for assisting in the screening of concurrent HFpEF among hypercapnic COPD patients. While integrating SHAP-derived statistical associations and unsupervised phenotyping offers hypothesis-generating insights into cardiopulmonary risk stratification, rigorous external validation in multicenter prospective cohorts is required before clinical implementation.