From Context to Care: Rethinking Stigma Detection in Clinical Language Models

Abstract Natural language processing techniques are useful for identifying stigmatizing language in electronic health records but require careful consideration. This commentary article builds on “ Efficient Detection of Stigmatizing Language in Electronic Health Records via In-Context Learning ” by Chen et al, which highlights the importance of incorporating situational and temporal contexts in annotation and modeling efforts. We emphasize the need for researchers to explicitly articulate their paradigms and positionality, particularly when working with populations disproportionately affected by stigmatizing language. We also explore the differences arising from conflicting preferences across communities about what constitutes destigmatizing language. We discuss participatory and trust-centered approaches for model development to work toward unbiased impact. Such strategies have a crucial role in raising awareness and fostering inclusive health care.