Explainability Versus Predictive Accuracy in Machine-Learning Models for 30-Day Hospital Readmission Among Patients With Heart Failure: A Systematic Evidence Synthesis

Abstract

Background: Thirty-day hospital readmission after heart-failure hospitalization remains difficult to predict because readmission reflects heterogeneous clinical, social, behavioral, and healthcare-utilization factors. Machine-learning models can capture nonlinear relationships and high-dimensional electronic health-record data, but increasingly complex models may sacrifice interpretability for only modest gains in discrimination.
Objective: To synthesize high-quality evidence available through 2022 concerning the predictive performance and explainability of machine-learning models for 30-day hospital readmission among patients with heart failure and to determine whether increases in model complexity translate into clinically meaningful predictive benefit.
Methods: A structured evidence synthesis was conducted using peer-reviewed Q1 literature published no later than 01 December 2022. Studies developing or comparing models for 30-day heart-failure readmission, readmission-or-death, or closely related early post-discharge outcomes were prioritized. Logistic regression, tree-based models, ensemble methods, neural networks, and deep-learning approaches were compared with respect to discrimination, class imbalance, calibration, interpretability, validation, and potential clinical utility. Methodological literature concerning explainable artificial intelligence and clinical prediction was incorporated.
Results: Prediction of 30-day readmission remained only moderately discriminative. Mortazavi et al. reported improvement of machine-learning approaches over logistic regression, but absolute discrimination remained limited. Golas et al. reported c-statistics of 0.664 for logistic regression, 0.650 for gradient boosting, 0.695 for maxout networks, and 0.705 for a deep unified network. The deep model therefore gained only 0.041 in c-statistic over logistic regression while losing direct feature-level interpretability. In a 2022 independent validation study, XGBoost achieved AUROC 0.65 compared with 0.58 for a neural network and 0.57 for the modified LaCE score. Feature-importance approaches improved global interpretation, but patient-level post-hoc explanations introduce additional questions concerning fidelity and stability.
Conclusion: More complex machine-learning models can modestly improve prediction of 30-day heart-failure readmission, but the gain in accuracy is frequently smaller than expected and may not justify reduced transparency. Model selection should consider calibration, precision-recall performance, external validity, explanation fidelity, and clinical actionability in addition to AUROC. For clinical deployment, the preferred model is the simplest approach that preserves clinically meaningful predictive performance while providing reliable individual-level explanations.