Towards Fairness and Interpretability: Clinical Decision Support for Acute Coronary Syndrome

Machine learning based prediction of acute coronary syndrome positive patients is becoming increasingly common. However, there is little to no research in terms of assessing fairness and interpretability of these methods. This study discusses methodologies to reduce fairness related disparities in acute coronary syndrome predictions made by machine learning algorithms for different demographics sections of the population. Additionally, existing interpretability technique called LIME is used to understand predictions made using fairness methodologies. Our analysis show that discussed methods have: (1) potential to reduce treatment disparities; (2) generalize well to different healthcare setting in the United States; and (3) reduce overall false negative rates by 0.34 and improve AUROC score by 0.11 when compared to algorithms like XGBoost.