Automated detection of stigmatizing language in Electronic Health Records (EHRs) using a multi-stage transfer learning approach.

Abstract

Stigmatizing language (SL) in Electronic Health Records (EHRs) can perpetuate biases and negatively impact patient care. This study introduces a novel method for automatically detecting such language to improve healthcare documentation practices. We developed a multi-stage transfer learning framework integrating semantic, syntactic, and task adaptation using three datasets: hate speech, clinical phenotypes, and stigmatizing language. Experiments were conducted on stigmatizing language dataset which consists of 4,129 de-identified EHR notes (72.7% stigmatizing, 27.3% non-stigmatizing), split 80/20 for training and testing. Longformer, BERT, and ClinicalBERT models were evaluated, and model performance was assessed on 35 randomized subsets of the test set (each comprising 70% of test data). The Wilcoxon-Mann-Whitney test was used to evaluate statistical significance, with Bonferroni correction applied to control for multiple hypothesis testing. Baseline models included zero-shot and few-shot GPT-4o, Support Vector Machine, Random Forest, Logistic Regression, and Multinomial Naive Bayes. The proposed framework achieved the highest accuracy, with fully adapted Longformer reaching 89.83%. Performance improvements remained statistically significant after Bonferroni correction compared to all baselines (p < .05). The framework demonstrated robust gains across different stigmatizing language types. This study underscores the value of domain-adaptive NLP for detecting stigmatizing language in EHRs. The multi-stage transfer learning framework effectively captures subtle biases often missed by conventional models, enabling more objective and respectful clinical documentation. This framework offers a statistically validated, high-performing framework for detecting stigmatizing language in EHRs, supporting responsible AI and promoting equity in clinical care.