Towards Fairness in Classifying Medical Conversations into SOAP Sections

As machine learning algorithms are more widely deployed in healthcare, the question of algorithmic fairness becomes more critical to examine. Our work seeks to identify and understand disparities in a deployed model that classifies doctor-patient conversations into sections of a medical SOAP note. We employ several metrics to measure disparities in the classifier performance, and find small differences in a portion of the disadvantaged groups. A deeper analysis of the language in these conversations and further stratifying the groups suggests these differences are related to and often attributable to the type of medical appointment (e.g., psychiatric vs. internist). Our findings stress the importance of understanding the disparities that may exist in the data itself and how that affects a model's ability to equally distribute benefits.