Efficient Detection of Stigmatizing Language in Electronic Health Records via In-Context Learning: A Comparative Analysis and Validation Study

This study investigates the efficacy of in-context learning (ICL) in detecting stigmatizing language in Electronic Health Records (EHRs) under data-scarce conditions. The performance of ICL, using four prompting strategies (Generic, Chain of Thought, Clue and Reasoning Prompting, and a novel Stigma Detection Heuristic Prompt), was compared to zero-shot (textual entailment), few-shot (SetFit), and supervised fine-tuning approaches. ICL significantly outperformed zero-shot and few-shot methods, achieving an F1 score comparable to the supervised model despite using significantly less data. Fairness evaluation revealed that supervised fine-tuning models exhibited greater bias than ICL models.

Hongbo Chen, Myrtede Alfred, Eldan Cohen