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

Background The presence of stigmatizing language within electronic health records (EHRs) poses significant risks to patient care by perpetuating biases. While numerous studies have explored the use of supervised machine learning models to detect stigmatizing language automatically, these models require large, annotated datasets, which may not always be readily available. In-context learning (ICL) has emerged as a data-efficient alternative, allowing large language models to adapt to tasks using only instructions and examples. Objective We aimed to investigate the efficacy of ICL in detecting stigmatizing language within EHRs under data-scarce conditions. Methods We analyzed 5043 sentences from the Medical Information Mart for Intensive Care–IV dataset, which contains EHRs from patients admitted to the emergency department at the Beth Israel Deaconess Medical Center. We compared ICL with zero-shot (textual entailment), few-shot (SetFit), and supervised fine-tuning approaches. The ICL approach used 4 prompting strategies: generic, chain of thought, clue and reasoning prompting, and a newly introduced stigma detection guided prompt. Model fairness was evaluated using the equal performance criterion, measuring true positive rate, false positive rate, and F1-score disparities across protected attributes, including sex, age, and race. Results In the zero-shot setting, the best-performing ICL model, GEMMA-2, achieved a mean F1-score of 0.858 (95% CI 0.854-0.862), showing an 18.7% improvement over the best textual entailment model, DEBERTA-M (mean F1-score 0.723, 95% CI 0.718-0.728; P<.001). In the few-shot setting, the top ICL model, LLAMA-3, outperformed the leading SetFit models by 21.2%, 21.4%, and 12.3% with 4, 8, and 16 annotations per class, respectively (P<.001). Using 32 labeled instances, the best ICL model achieved a mean F1-score of 0.901 (95% CI 0.895-0.907), only 3.2% lower than the best supervised fine-tuning model, ROBERTA (mean F1-score 0.931, 95% CI 0.924-0.938), which was trained on 3543 labeled instances. Under the conditions tested, fairness evaluation revealed that supervised fine-tuning models exhibited greater bias compared with ICL models in the zero-shot, 4-shot, 8-shot, and 16-shot settings, as measured by true positive rate, false positive rate, and F1-score disparities. Conclusions ICL offers a robust and flexible solution for detecting stigmatizing language in EHRs, offering a more data-efficient and equitable alternative to conventional machine learning methods. These findings suggest that ICL could enhance bias detection in clinical documentation while reducing the reliance on extensive labeled datasets.