NLM IRP Seminar Schedule

Abstract:

Although the powerful applications of machine learning (ML) are revolutionizing medicine, current algorithms are not resilient against bias. Fairness in ML can be defined as measuring the potential bias in algorithms with respect to characteristics such as race, gender, age, etc. In this paper, we perform a comparative study and systematic analysis to detect bias caused by imbalanced group representation in sample medical datasets. We investigate bias in major medical tasks for three datasets: UCI Heart Disease dataset (cardiac disease classification), Stanford Diverse Dermatology Image (DDI) dataset (skin cancer prediction), and chestX-ray dataset (CXR lung segmentation). Our results show differences in the performance of the state-of-the-arts across different groups. To mitigate this disparity, we explored three bias mitigation approaches and demonstrated that integrating these approaches into ML models can improve fairness without degrading the overall performance.