Bias Mitigation in Clinical AI: Auditing Race/Gender Disparities in Sepsis Prediction Models

Abstract

This paper exposes persistent race and gender biases in AI-based sepsis prediction models, arguing that these inequities undermine patient outcomes and demanding prioritization of fairness as a core clinical metric. An audit of multiple AI tools in urban hospitals revealed consistent accuracy gaps, notably more false negatives for Black, Hispanic, female, and non-binary patients, which delayed care and worsened clinical results. These disparities stem from imbalanced training data, distance miscalibration, and structural inequities embedded in clinical practice. The manuscript surveys algorithmic bias types, presents audit frameworks (Fairlearn, Aequitas), and evaluates mitigation strategies such as data rebalancing, fair regularization, threshold adjustment, and explainable tools like SHAP and LIME. It further argues that implementing AI in healthcare must be grounded in the ethical imperatives of beneficence, non-maleficence, and justice. Future research will focus on intersectional bias analysis and prospective audits integrated into electronic health records. The findings attribute an immediate need for institutional responsibility towards facilitating clinical AI systems that promote health equity among all populations.