Build fair machine learning models to predict adverse outcomes for heart failure patients with preserved ejection...

Develop and validate subtype-specific, fairness-aware machine learning (ML) models that integrate clinical and social determinants of health (SDoH) information to predict 6-month readmission or mortality after hospitalization among patients with HFpEF or HFrEF, and to evaluate the incremental contribution of SDoH, model explainability, and subgroup error disparities across demographic groups.

We used University of Florida Health electronic health record (EHR) data (2016-2022) to identify adult heart failure (HF) hospitalizations and followed patients for 6 months for a composite outcome of readmission or mortality. Features included clinical characteristics, contextual SDoH (eg, neighborhood deprivation), and individual SDoH extracted from clinical notes via natural language processing (NLP). Logistic regression and XGBoost models were trained with random oversampling. Performance metrics included the C statistic, F1-score, and recall. Fairness was evaluated using false negative rate (FNR) parity across sex, race/ethnicity, and age band, and mitigation methods were applied (eg, Disparate Impact Remover, Adversarial Debiasing, and Calibrated Equalized Odds).

Adding SDoH improved the C statistic for logistic regression in HFpEF (0.603 vs 0.586) and HFrEF (0.641 vs 0.637). SHapley Additive exPlanations (SHAP) highlighted sodium, financial constraint level, and emergency department visit count in HFpEF, and utilization measures and financial constraint level in HFrEF. FNR ratios indicated race/ethnicity disparities; HFpEF FNRBlack/FNRWhite was 0.7834 (0.8728 after Disparate Impact Remover), and HFrEF FNRHispanic/FNRWhite was 1.2217 (0.9880 after Adversarial Debiasing).

SDoH integration and mitigation can modestly improve performance while reducing subgroup error disparities.

Subtype-specific, fairness-aware ML models for HFpEF and HFrEF provided interpretable 6-month risk stratification and enabled subgroup fairness assessment. Integrating clinical and SDoH information added modest discrimination gains while strengthening interpretation and fairness assessment. These findings support further validation of explainable, equity-aware HF prediction models.