Bias Mitigation In Large-Scale AI Systems For Healthcare Decision Support: Techniques, Data Processing, And Evaluation

Abstract

The increasing adoption of large-scale Artificial Intelligence (AI) systems in healthcare has raised critical concerns regarding algorithmic bias and fairness in clinical decision-making. Biased AI models can lead to unequal healthcare outcomes, particularly across diverse demographic groups. This research investigates effective bias mitigation techniques within large-scale AI systems, specifically for healthcare decision support. Healthcare data is collected from Electronic Health Records (EHRs) that encompass Patient Demographics, Clinical History, and Diagnostic Reports. Min-Max Normalization is utilized to standardize feature scales and mitigate bias, thereby enhancing model stability. Additionally, Principal Component Analysis (PCA) facilitates feature extraction by reducing dimensionality while maintaining data variance, thus improving computational efficiency and interpretability. The proposed model integrates Nesterov Accelerated Gradient with a Bi-directional Gated Recurrent Unit (NAG-Bi-GRU) to enhance learning efficiency and capture complex temporal dependencies in healthcare data. The NAG optimizer improves convergence speed and reduces training bias, while Bi-GRU effectively models sequential patterns from both forward and backward directions. Bias mitigation is further reinforced through fairness-aware training strategies and output calibration. Model performance metrics indicate high accuracy (94.8%), precision (93.9%), recall (93.5%), F1-Score (93.7%), and AUC-ROC (0.97), while fairness metrics show demographic parity and equal opportunity. The model demonstrates reduced bias and maintains high predictive performance, while promoting fair and scalable AI systems for healthcare decision support.