Resolving Accuracy-Fairness Trade-offs in Clinical Decision Support: A Heterogeneous Neural Energy-Based Framework for Diabetes Management

Background Clinical AI models frequently exhibit degraded performance and systemic bias when applied to highly imbalanced healthcare datasets. Navigating the inherent trade-off between predictive fidelity and demographic equity remains a critical challenge for translational medicine. To address this, we propose the Heterogeneous Neural Energy-Based Classifier ( ), a novel dual-paradigm framework. Methods The proposed architecture integrates a Dynamic Class Matrix ( ) for adaptive under-sampling with Simulated Annealing ( ) for robust hyperparameter optimization across non-convex loss landscapes. Additionally, an Adaptive Fairness Regulation ( ) module is incorporated to explicitly penalize demographic parity deviations during gradient-based optimization, ensuring equitable model convergence. Results Empirical evaluations on the Diabetes 130-US Hospitals dataset demonstrate that the deep gradient-based configuration ( ) achieves highly robust predictive performance in medication prescription, yielding an Area Under the Receiver Operating Characteristic ( ) of , an accuracy of , and a Matthews Correlation Coefficient ( ) of . Crucially, the AFR module curtails the Predicted Positive Rate ( ) disparity to , effectively mitigating algorithmic bias. Conversely, for 30-day readmission prediction, the gradient-free heuristic configuration ( ) prioritizes rapid optimization (completing evaluation in ) and strict demographic equity ( ), accepting an anticipated trade-off in global discriminative power ( , ). Conclusions The framework offers a mathematically rigorous and adaptable methodology for trustworthy clinical decision support. By balancing highly discriminative capabilities with demonstrable algorithmic fairness, it establishes a reliable paradigm for deploying equitable AI interventions in endocrinology and broader healthcare networks.