Machine Learning Predicts Functional Decline and Risk Phenotypes in Older Patients With Heart Failure

BACKGROUND: Preserving functional independence is critical in older patients with heart failure (HF), yet tools predicting long-term functional trajectories are scarce. AIMS: To develop and validate a machine learning (ML) model predicting functional decline, accounting for the competing risk of mortality. METHODS: We analyzed 6382 older patients (median age, 82 years) from a nationwide prospective cohort (J-Proof HF registry). Patients aged ≥65 years, hospitalized for HF, prescribed rehabilitation, and independent preadmission (Barthel Index [BI] ≥85) were included. An eXtreme Gradient Boosting (XGBoost) model was developed to predict a 1-year 3-class outcome: functional maintenance, functional decline, or death. Functional decline was defined as transitioning to a housebound/bedridden state (equivalent to BI <85) based on a national long-term care scale. Performance was evaluated via nested leave-one-site-out validation and benchmarked against the Kihon Checklist (KCL) frailty screening tool. RESULTS: At 1 year, 32.5% of patients experienced functional decline and 14.5% died. A parsimonious Top-10 XGBoost model (including maximum gait speed, discharge BI, preadmission frailty score, and age) demonstrated good discrimination (area under the receiver operating characteristic curve [AUC], 0.75; 95% confidence interval [CI], 0.74-0.76), significantly outperforming KCL-based scores (AUC, 0.66-0.69; P < .001). The model identified 4 risk phenotypes, including a "low-mortality/high-decline" group (n = 1732) with a 57.0% functional decline rate despite an 89.0% survival probability. CONCLUSION: This ML model accurately identifies older HF patients at high risk for functional decline despite favorable survival, enabling targeted rehabilitation to preserve quality of life.