Towards predictive, preventive, and personalized medicine in HFpEF with obstructive sleep apnea: a multi-horizon explainable machine learning framework for readmission risk stratification

Background Heart failure with preserved ejection fraction (HFpEF) complicated by obstructive sleep apnea syndrome (OSAS) represents a high-risk multimorbid phenotype characterized by frequent hospital readmissions and substantial clinical heterogeneity. Current prediction approaches largely focus on static risk estimation and provide limited support for predictive, preventive, and personalized medicine (PPPM). Understanding how readmission risk evolves across different clinical time points may facilitate targeted prevention and individualized management strategies. Objective To develop and externally validate a multicenter, explainable machine learning (ML) framework for multi-horizon readmission risk stratification in patients with HFpEF and OSAS, and to explore time-horizon-specific predictive contributions of risk determinants to support PPPM-oriented clinical decision-making. Methods In this multicenter retrospective cohort study, 2,211 patients with HFpEF and OSAS from two tertiary medical centers in Xinjiang, China, were enrolled between January 2018 and December 2021 and followed for a mean of 33.5 months. Patients were allocated to a development cohort (n = 1,338) and an external validation cohort (n = 873). A dual-stage feature selection strategy integrating Boruta and least absolute shrinkage and selection operator regression was applied to identify key predictors. Twelve ML algorithms were trained and evaluated for predicting first unplanned heart failure readmission at 1-, 2-, and 3-year clinical time points. Model performance was assessed using discrimination, calibration, and decision curve analyses. Explainable ML techniques were used to characterize horizon-specific patterns of model attribution. Results Eight predictors were retained in the final models, including apnea-hypopnea index (AHI), body mass index, diabetes mellitus, total cholesterol, fasting plasma glucose, left ventricular end-diastolic dimension, longest duration of apnea, and the atherogenic index of plasma–body mass index (AIP-BMI). The best-performing models achieved areas under the receiver operating characteristic curve of 0.820 (95% CI 0.729–0.884) at 1 year, 0.783 (95% CI 0.727–0.842) at 2 years, and 0.830 (95% CI 0.792–0.863) at 3 years in the external validation cohort. Calibration and decision curve analyses demonstrated acceptable agreement and potential decision-analytic net benefit under the evaluated threshold assumptions. Explainability analyses revealed horizon-specific differences in risk architecture: sleep-disordered breathing severity, represented by AHI, predominated in short-term readmission risk, whereas metabolic burden, represented by AIP-BMI, became increasingly influential during longer follow-up. A web-based calculator was developed to facilitate individualized risk assessment. Conclusions This study developed and center-based externally validated multi-horizon, explainable machine-learning models for individualized estimation of first unplanned heart failure readmission risk at 1, 2, and 3 years in patients with HFpEF and OSAS. These models may support future research on phenotype-specific predictive risk stratification within a PPPM framework; however, prospective validation, implementation studies, and evidence of clinical benefit are required before clinical use. Chinese clinical trial registry The trial was registered in the Chinese Clinical Trial Registry (ChiCTR2300075727; 2023-09-14).