Abstract
BACKGROUND: Heart failure (HF) is increasing in Japan's rapidly aging population, yet use of implantable cardioverter defibrillators and cardiac resynchronization therapy remains lower than in Western countries. Using data from HINODE, which prospectively evaluated Japanese patients with cardiac devices, we developed interpretable machine learning (ML) models to improve risk stratification and identify key predictors of adverse outcomes. METHODS AND RESULTS: Among 354 HINODE participants, 332 with adequate data were analyzed. Predictive models (XGBoost; 5-fold cross-validation) targeted HF hospitalization and all-cause mortality. Missingness was handled with multiple imputation; calibration was assessed by calibration plots and Hosmer-Lemeshow tests. Model discrimination was strong (area under the curve 0.83 and 0.85 for HF events and mortality). Shapley additive explanations (SHAP) highlighted QRS duration, QT interval, left ventricular (LV) volumes, and selected medications as major contributors. Using top SHAP features, K-means (k=2) identified low-risk (n=236) and high-risk (n=86) clusters. The high-risk cluster had larger LV volumes, wider QRS, and higher event rates. Kaplan-Meier curves showed significant differences between clusters for HF events (15.7% vs. 47.7%, log-rank P<0.001) and mortality (8.1% vs. 20.9%; hazard ratio 2.58, 95% confidence interval 1.45-4.60). Performance was temporally stable across enrollment periods. CONCLUSIONS: Interpretable ML provided accurate risk prediction and phenotype-based stratification in Japanese HF patients with cardiac devices, supporting personalized management.