Interpretable AI-driven multi-objective risk prediction in heart failure patients with thyroid dysfunction

Introduction: Heart Failure (HF) complicated by thyroid dysfunction presents a complex clinical challenge, demanding more advanced risk stratification tools. In this study, we propose an AI-driven machine learning (ML) approach to predict mortality and hospitalization risk in HF patients with coexisting thyroid disorders. Methods: Using a retrospective cohort of 762 HF patients (including euthyroid, hypothyroid, hyperthyroid, and low T3 syndrome cases), we developed and optimized several ML models-including Random Forest, Gradient Boosting, Support Vector Machines, and others-to identify high-risk individuals. Results: The best-performing model, a Random Forest classifier, achieved robust predictive accuracy for both 1-year mortality and HF-related hospitalization (area under the ROC curve ∼0.80 for each). We further employed model interpretability techniques (Local Interpretable Model-agnostic Explanations, LIME) to elucidate key predictors of risk at the individual level. This interpretability revealed that factors such as atrial fibrillation, absence of cardiac resynchronization therapy, amiodarone use, and abnormal thyroid-stimulating hormone (TSH) levels strongly influenced model predictions, providing clinicians with transparent insights into each prediction. Additionally, a multi-objective risk stratification analysis across thyroid status subgroups highlighted that patients with hypothyroidism and low T3 syndrome are particularly vulnerable under high-risk conditions, indicating a need for closer monitoring and tailored interventions in these groups. Discussion: In summary, our study demonstrates an innovative AI methodology for medical risk prediction: interpretable ML models can accurately stratify mortality and hospitalization risk in HF patients with thyroid dysfunction, offering a novel tool for personalized medicine. These findings suggest that integrating explainable AI into clinical workflows can improve prognostic precision and inform targeted management, though prospective validation is warranted to confirm realworld applicability.