Proposed Machine Learning Framework for Ethical Clinical Note-Taking

Abstract

Ethnic and gender bias can creep into medical notes by physicians within a clinical setting. Physician bias as seen through verbiage of clinical notes and electronic health records (EHR) for certain demographics (e.g. African-Americans, Hispanic/Latinx, and women) can affect the quality of the health care these patients receive. This is important for future disease diagnosis and solutions for medical issues toward certain demographics, especially when human bias can get in the way. This study first reviews the literature on the extent and types of biases and impacts on certain demographics in the clinical setting. By looking into these issues, we propose a ML question framework to mitigate these biases when a clinician is taking notes in their pre-processing phase of data. Then, we assess the relevance of this framework within the context of the MIMIC-III dataset, where we evaluate a comparison of negative descriptors in the post-processing phase. Finally, we provide a set of conditions with a patient's background to guide a physician to record and evaluate medical needs in a holistic fashion to treat all patients fairly.