Cognitive priming in <scp>AI</scp> : Bias identification and analysis in electronic health records

Abstract Electronic healthcare records (EHRs) contain healthcare inequities that impede equitable patient care, frequently due to subtle and unconscious biases ingrained in medical language. These biases, when documented in EHRs, can significantly impact subsequent medical decisions and the quality of care. Leveraging priming theory in psychology, we introduce a novel natural language processing application that simulates cognitive activation to enhance the detection of subtle EHR biases. This approach transitions our focus from mere technical integration to comprehending the psychological underpinnings of bias perception within a machine learning framework. Moreover, employing a data‐driven strategy, we investigate potential correlations between patient demographics—like age, gender, and race—and the prevalence of subtle biases through analyzing model performance, attention shifts, and word contributions. Conducted on the MIMIC database, our research demonstrates that simulating cognitive memory activation improves the model's capability to detect and classify biased language. Additionally, our findings indicate a potential association between specific biased expressions and patient demographic traits. Our work deepens the understanding of discrimination mechanisms in healthcare and underscores the value of cognitive psychological theories in crafting AI systems aimed at social welfare.