Cardiomonostack: A postoperative heart failure risk prediction model for total hip arthroplasty driven by ensemble learning

Background Accurate perioperative risk prediction remains challenging because of class imbalance, cohort heterogeneity, unstable probability estimates and limited interpretability. Objective To develop and externally evaluate CardioMonoStack, an ensemble architecture integrating heterogeneous base learners, branch-level monotonicity priors, stacked meta-learning and feature-dependent residual logit correction. Methods We conducted a multicentre retrospective cohort study of 1911 patients aged ≥65 years undergoing total hip arthroplasty across five healthcare databases. CardioMonoStack was developed using MIMIC-III and MIMIC-IV, validated in INSPIRE, and externally tested in eICU and MOVER. Performance was assessed using AUROC, AUPRC, Brier score, threshold-dependent classification measures, calibration plots and decision curve analysis. Interpretability was examined using SHAP, interaction and individual conditional expectation analyses. Results CardioMonoStack achieved the highest AUROC point estimates among evaluated models in eICU (0.790; 95% CI 0.711–0.865) and MOVER (0.866; 95% CI 0.761–0.949), with AUPRC values of 0.569 (95% CI 0.415–0.717) and 0.571 (95% CI 0.328–0.792), respectively. Calibration and net benefit varied across cohorts and comparators. Omission of L3 increased Brier-score and log-loss point estimates in both external cohorts; paired confidence intervals supported these differences in MOVER but were inconclusive in eICU. SHAP analyses prioritised blood urea nitrogen, creatinine, age, and metabolic and demographic factors. Conclusion CardioMonoStack is a candidate framework for estimating in-hospital heart-failure risk after total hip arthroplasty. Despite promising external discrimination, prospective validation, local recalibration, and clinical impact evaluation are required before implementation.