Predicting 30-day readmission after heart failure hospitalization using interpretable machine learning: evidence from a 20-year population-based rural registry.

Abstract

Readmissions after heart failure hospitalization are common, costly, and potentially preventable, but discharge risk stratification remains challenging, particularly in rural health systems with constrained resources. We aimed to develop an interpretable artificial intelligence approach to estimate individual 30-day all-cause readmission risk using routinely collected variables. We analyzed nearly 5,000 heart failure admissions from a non-public regional Spanish registry representing a rural healthcare setting (2000-2019; 8.2% readmitted within 30 days) and not previously used for artificial intelligence modeling. Three machine learning models -random forest, extreme gradient boosting, and support vector machine- were trained, validated, and compared with binary logistic regression. SHAP quantified predictor contributions and assessed the direction and consistency of their effects. Random forest showed the best performance (AUC 0.812, 95% CI 0.744-0.867), outperforming binary logistic regression (AUC 0.686, 95% CI 0.617-0.755). The most influential predictors were admission period, renal dysfunction markers, prior heart failure, age, and length of stay. Interpretable machine learning improved 30-day readmission risk stratification using routine data and provided transparent explanations that could support targeted post-discharge interventions. External validation and prospective impact evaluation are required before clinical implementation.