Intelligent Patient Risk Prediction and Bias-Aware Diagnosis Framework

Abstract: The proliferation of electronic health records (EHRs) and multimodal clinical data has created unprecedented opportunities for machine learning-driven patient risk assessment. However, existing clinical decision-support systems frequently suffer from systematic algorithmic bias arising from imbalanced training corpora, demographic skews, and feature selection disparities, leading to inequitable diagnostic outcomes across protected population subgroups. This paper presents the Intelligent Patient Risk Prediction and Bias-Aware Diagnosis (IPRB-AD) framework—a novel hybrid architecture that integrates a Graph Attention Network (GAT) for patient similarity modeling, a Transformer- based temporal encoder for longitudinal EHR sequences, and an adversarial debiasing module grounded in fairness constraints. The proposed system jointly optimizes predictive accuracy and demographic parity through a multi- objective loss formulation, incorporating Counterfactual Fairness Regularization (CFR) to mitigate bias without sacrificing clinical utility. Experiments conducted on three publicly available benchmarks—MIMIC-IV, eICU Collaborative Research Database, and the PhysioNet Sepsis Challenge dataset—demonstrate that IPRB-AD achieves an AUROC of 0.934, F1-score of 0.891, and reduces disparity gap by 41.7% compared to state-of-the-art baselines. The framework provides interpretable risk scores via SHAP-based attribution maps, enabling clinicians to audit model decisions and identify latent bias sources. These results underscore the potential of fairness-constrained deep learning pipelines in realizing trustworthy, equitable clinical AI systems.

Keywords: Patient Risk Prediction; Algorithmic Bias; Graph Attention Networks; Fairness-Aware Machine Learning; Electronic Health Records; Adversarial Debiasing; Clinical Decision Support