Machine learning for mortality prediction in COPD-CVD comorbidity

Chronic Obstructive Pulmonary Disease (COPD) and Cardiovascular Disease (CVD) frequently co-occur, substantially elevating mortality risk, yet machine learning models for this comorbid population remain understudied. We study mortality prediction in 2,799 UK Biobank participants with COPD-CVD comorbidity using clinical, lifestyle, lung function, and biomarker variables collected at baseline with a 14.8-year follow-up. We compare nine supervised learning models including Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machines, and Neural Networks under a consistent preprocessing and evaluation pipeline and report AUC-ROC, accuracy, precision, recall, and F1-score with bootstrap confidence intervals. Gradient Boosting achieves the highest accuracy of 76.6% (95% CI: 72.9–80.0%) with an AUC-ROC of 0.724 (95% CI: 0.677–0.773), yet recall for mortality events remains low at 0.33, indicating that models predominantly predict survival rather than death. SHAP analysis reveals that demographic and socioeconomic factors are the dominant predictors, with CRP ranking lower than previously suggested by impurity based importance metrics. These results suggest that conventional Machine Learning (ML) approaches struggle to reliably identify at-risk individuals in comorbid COPD-CVD populations, and that the complexity arising from co-occurring diseases presents challenges that basic supervised learning methods are not well equipped to resolve.