A machine learning model for identifying heart failure in patients with chronic obstructive pulmonary disease based on routine laboratory indicators

Abstract Background Heart failure (HF) is a frequent and prognostically important comorbidity of chronic obstructive pulmonary disease (COPD), yet it is often documented late. We developed and temporally validated an interpretable machine learning model that identifies clinically documented comorbid HF in patients with COPD using routinely collected laboratory indicators. Methods In this retrospective study, 1,087 patients with COPD (292 with clinically documented HF) hospitalised between January 2023 and December 2025 formed the development cohort, and 268 patients hospitalised between January and September 2026 (63 with HF) formed a temporal external validation cohort. Twenty-five candidate predictors were screened by least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm; nine algorithms were trained without class-imbalance correction, and the final model was selected on cross-validated discrimination and out-of-fold calibration. Internal validity was quantified by bootstrap optimism correction with the entire pipeline nested (1,000 resamples). The model was benchmarked against NT-proBNP. Results The elastic net model, built on 12 routine predictors, achieved an apparent development AUC of 0.817, an optimism-corrected AUC of 0.799 and an external validation AUC of 0.759 (95% CI 0.683–0.835), with near-ideal external calibration (intercept − 0.133, slope 0.958). At a development-derived threshold targeting 90% sensitivity, 98 of the 268 patients in the external cohort (36.6%) were classified as negative and could potentially have been spared echocardiography, at the cost of missing 11 of the 63 patients with HF (17.5%). NT-proBNP alone achieved an external AUC of 0.983; the model adds no discrimination where natriuretic peptide testing is already available. SHAP variable rankings were preserved across cohorts. Conclusion An elastic net model built exclusively on routine laboratory indicators identified clinically documented comorbid HF in patients with COPD, with discrimination and calibration preserved across a temporal external cohort. Because NT-proBNP outperformed the model, its intended role is a zero-marginal-cost pre-screening layer applied to blood tests that already exist, to prioritise which patients with COPD warrant natriuretic peptide testing and echocardiography. External validation in independent centres is required before clinical use.