A scoping review of algorithmic equity, data diversity, and inclusive design in the transformer era of clinical NLP

The rapid digitization of healthcare has positioned transformer-based natural language processing (NLP) models as powerful tools for managing clinical textual data. However, their integration into practice raises unresolved questions regarding equity and inclusivity.

This scoping review examines how equity is addressed in transformer-based clinical NLP, with a focus on algorithmic equity, data diversity and representativeness, and participatory design.

A scoping review was conducted, synthesizing 56 studies published between 2017 and 2024. Guided by an intersectionality approach and the Digital Health Equity framework, studies were analyzed to assess how equity-related considerations are operationalized in transformer-based clinical NLP research.

Most equity audits were post hoc and fragmented, with limited influence on model development. Persistent underrepresentation of linguistic, demographic, and clinical subgroups was identified, giving rise to what we define as Data Diversity Debt. Participatory design was observed in 11% of studies, indicating limited stakeholder inclusion beyond clinicians. Fairness metrics were inconsistently defined, limiting comparability and accountability across studies.

These findings highlight the need to move beyond descriptive equity audits toward equity-by-design approaches. We translate the synthesized evidence into an equity-by-design roadmap that embeds fairness, inclusivity, and accountability across the full lifecycle of healthcare NLP systems. We argue that equity must shift from reactive evaluation to proactive design, incorporating participatory governance, fairness-aware training objectives, and continuous monitoring to address Data Diversity Debt and reduce the risk of reproducing health disparities.