Fairness Audit and Bias Mitigation in an AI-Based Heart Disease Risk Prediction Model

Artificial intelligence (AI) is increasingly used for healthcare prediction, but machine learning models may produce different outcomes for different demographic groups.This study presents a fairness audit and bias mitigation analysis of an AI-based heart disease risk prediction model using a dataset of 918 patient records.A Logistic Regression model was developed using one-hot encoding, feature scaling, and an 80:20 stratified train-test split.The model achieved an accuracy of 88.59%, with a precision of 87.16%, recall of 93.14%, and F1-score of 90.05%.Fairness was evaluated for Sex and AgeGroup using Demographic Parity Difference (DPD), Equalized Odds Difference (EOD), and Disparate Impact Ratio (DIR) through the Fairlearn framework.The baseline model showed noticeable demographic differences, particularly across Sex, with a DPD of 0.5144 and a DIR of 0.2637.To reduce these disparities, two bias mitigation approaches, Exponentiated Gradient and Threshold Optimizer, were applied using Sex as the sensitive feature.Exponentiated Gradient reduced the Sex DPD to 0.0512 and increased the Sex DIR to 0.9188, while Threshold Optimizer reduced the Sex DPD to 0.0865 and increased the Sex DIR to 0.8749.However, these fairness improvements were accompanied by lower overall accuracy, with Exponentiated Gradient achieving 82.61% and Threshold Optimizer achieving 81.52%.The results show that fairness improvements can involve a trade-off with predictive performance and that evaluating multiple fairness measures is important when assessing machine learning models for healthcare applications.