Abstract
As health systems move to make electronic health records (EHRs) accessible to patients, there is a need to examine if, and the extent to which, bias toward patients may be evident in these records. This scoping review aimed to summarize the scientific literature on methods used to detect biased language about patients in the EHR and the nature and object of the biases detected. A comprehensive literature search was conducted of PubMed, CINAHL, Web of Science, APA PsycInfo, and SOCIndex for peer-reviewed English language studies conducted in the United States published on or before December 22, 2022. Seven studies were included in this review. Four methods were identified: natural language processing methods including machine learning-based (n = 3), Linguistic Inquiry and Word Count (n = 2), and exploratory vocabulary analysis (n = 1), and manual content analysis (n = 2). In four studies, the EHR of Black patients contained significantly greater bias relative to the EHR of White patients. Bias about health conditions (i.e., diabetes, substance use disorder, and chronic pain), women, and preexposure prophylaxis-a medication that prevents HIV infection-were identified. Machine-based learning methods may be best to (a) analyze robust data sampling frames, (b) detect a rare outcome like bias, (c) facilitate inferential analysis, and (d) transcend limitations of manual content analysis. Findings provide an overview of methods that can be used by investigators to analyze EHR records for bias to inform clinical interventions, health policies, and procedures to reduce bias among health care providers.