Machine learning prediction of 1-year mortality in older patients with heart failure: a nationwide, multicenter, prospective cohort study.

Background: Risk prediction in older adults with heart failure (HF) often depends on conventional clinical variables and does not incorporate direct measures of physical function. This study aimed to develop and validate a machine learning model that integrates routinely collected functional assessments to predict one-year all-cause mortality. Methods: We analyzed data from the J-Proof HF Registry, a nationwide prospective cohort encompassing 96 institutions in Japan. Patients aged ≥65 years who were hospitalized for HF and received a prescription for rehabilitation between December 2020 and March 2022 were included. An eXtreme Gradient Boosting (XGBoost) model was developed using 77 candidate predictors, including demographic, clinical, laboratory, echocardiographic, and directly measured functional variables. Validation employed a leave-one-site-out (LOSO) internal-external framework. Discrimination, calibration, and clinical utility were compared with established risk scores. A 20-predictor model was derived using SHAP-based importance and clinical review. Findings: Among 9700 eligible patients, the median age was 83 years (interquartile range [IQR] 77-88), 4915 (50.7%) were male, and 1601 (16.5%) died within one year. The full XGBoost model achieved an aggregated area under the receiver operating characteristic curve (AUC) of 0.76 (95% confidence interval [CI] 0.75-0.77) across LOSO test sets. The 20-predictor XGBoost model demonstrated similar discrimination (AUC 0.76; 95% CI 0.74-0.77). Both models outperformed the AHEAD and BIOSTAT compact scores. Functional measures at discharge, including Barthel index, the Short Physical Performance Battery, gait speed, and handgrip strength, were among the strongest contributors to model predictions. Reclassification metrics and decision curve analysis indicated greater clinical utility compared with benchmark scores. Interpretation: A machine learning model incorporating functional assessments predicts one-year mortality in older patients with HF and improves risk stratification beyond established scores. Functional status at discharge is an important prognostic indicator and may inform post-discharge care planning. Funding: This work was supported by research funding of Japanese Society of Cardiovascular Physical Therapy and JSPS KAKENHI Grant Number JP25K02969.