Bias Mitigation in Healthcare-Based Machine Learning Systems

Abstract

This study investigates the impact of oversampling techniques, specifically SVMSMOTE and BorderlineSMOTE, on machine learning models for heart disease and diabetes risk prediction. Using Gradient Boosting Machine (GBM) and K-Nearest Neighbors (KNN) algorithms, we assess changes in accuracy, precision, recall, F1 score, Positive Predictive Value (PPV), Equal Opportunity Difference (EOD), Disparate Impact (DI), and Impact Ratio (IR) across diverse biomedical datasets. In both heart disease and diabetes risk prediction tasks, SVM-SMOTE and BorderlineSMOTE proved effective in enhancing machine learning model performance. For heart disease prediction, SVM-SMOTE and BorderlineSMOTE improved GBM model accuracy to 0.85 and 0.85 from an initial 0.74, precision to 0.79 and 0.77 from 0.69, and recall to 0.88 and 0.92 from 0.75, respectively. KNN models also showed enhancements in accuracy (0.71 from 0.68), precision (0.70 from 0.59), and recall (0.72 from 0.62). In diabetes risk prediction, both techniques consistently boosted accuracy, precision, and F1 score metrics across GBM and KNN models. Notably, DI values improved significantly to 1.11 with both SVM-SMOTE and BorderlineSMOTE from an initial 0.43, indicating improved fairness in model predictions across demographic groups. Overall, the strategic application of SVM-SMOTE and BorderlineSMOTE effectively addresses class imbalance challenges in biomedical datasets, enhancing both predictive accuracy and fairness in machine learning models. These results underscore the importance of tailored oversampling techniques in achieving robust and equitable healthcare predictions across diverse demographic groups