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

Area of Science:

• Computational linguistics
• Medical informatics
• Sociology of health

Background:

• Natural language processing (NLP) aids in detecting stigmatizing language in electronic health records (EHRs).
• Contextual information (situational and temporal) is crucial for accurate NLP model performance.
• Previous work highlights the need for context-aware approaches in EHR analysis.

Purpose of the Study:

• To emphasize the importance of context in NLP for identifying stigmatizing language in EHRs.
• To advocate for researchers to articulate their paradigms and positionality.
• To explore community differences in defining destigmatizing language and promote inclusive healthcare.

Main Methods:

• Building upon existing research on efficient detection of stigmatizing language.
• Discussing the necessity of situational and temporal context in annotation and modeling.
• Exploring participatory and trust-centered approaches for model development.

Main Results:

• NLP models require careful consideration of situational and temporal contexts for accurate detection of stigmatizing language.
• Explicit articulation of researcher paradigms and positionality is essential, especially when working with vulnerable populations.
• Community preferences regarding destigmatizing language can vary, impacting model development.

Conclusions:

• Incorporating context and researcher positionality is vital for developing unbiased NLP tools for EHRs.
• Participatory and trust-centered approaches can foster inclusive healthcare by mitigating the impact of stigmatizing language.
• Raising awareness and promoting inclusive practices are key outcomes of such research strategies.