The Impact of Stigmatizing Language in EHR Notes on AI Performance and Fairness

Yizhi Liu, Weiguang Wang, Gordon Gao, Ritu Agarwal

Abstract: Description Today, there is significant interest in using electronic health record data to generate new clinical insights for diagnosis and treatment decisions. However, there are concerns that such data may be biased and result in accentuating racial disparities. We study how clinician biases reflected in EHR notes affect the performance and fairness of artificial intelligence models in the context of mortality prediction for intensive care unit patients. We apply a Transformer-based deep learning model and explainable AI techniques to quantify negative impacts on performance and fairness. Our findings demonstrate that stigmatizing language written by clinicians adversely affects AI performance, particularly so for black patients, highlighting SL as a source of racial disparity in AI model development. As an effective mitigation approach, removing SL from EHR notes can significantly improve AI performance and fairness. This study provides actionable insights for responsible AI development and contributes to understanding clinician EHR note writing.