Identifying Bias at Scale in Clinical Notes Using Large Language Models

To evaluate whether generative pretrained transformer (GPT)-4 can detect and revise biased language in emergency department (ED) notes, against human-adjudicated gold-standard labels, and to identify modifiable factors associated with biased documentation.

We randomly sampled 50,000 ED medical and nursing notes from the Mount Sinai Health System (January 1, 2023, to December 31, 2023). We also randomly sampled 500 discharge notes from the Medical Information Mart for Intensive Care IV database. The GPT-4 flagged 4 types of bias: discrediting, stigmatizing/labeling, judgmental, and stereotyping. Two human reviewers verified model detections. We used multivariable logistic regression to examine associations between bias and health care utilization, presenting problems (eg, substance use), shift timing, and provider type. We then asked physicians to rate GPT-4's proposed language revisions on a 10-point scale.

The GPT-4 showed 97.6% sensitivity and 85.7% specificity compared with the human review. Biased language appeared in 6.5% (3229 of 50,000) of Mount Sinai notes and 7.4% (37 of 500) of Medical Information Mart for Intensive Care IV notes. In adjusted models, frequent health care utilization (adjusted odds ratio [aOR], 2.85; 95% CI, 1.95-4.17), substance use presentations (aOR, 3.09; 95% CI, 2.51-3.80), and overnight shifts (aOR, 1.37; 95% CI, 1.23-1.52) showed elevated odds of biased documentation. Physicians were more likely to include bias than nurses (aOR, 2.26; 95% CI, 2.07-2.46); GPT-4's recommended revisions received mean physician ratings above 9 of 10.

The study showed that GPT-4 accurately detects biased language in clinical notes, identifies modifiable contributors to that bias, and delivers physician-endorsed revisions. This approach may help mitigate documentation bias and reduce disparities in care.