Medical records can influence how healthcare providers evaluate and care for patients, yet hospital discharge notes written about gender-expansive patients (including transgender and non-binary people) often contain stigmatizing language that can affect clinical decision-making, propagate bias across care encounters, and undermine patient trust, according to a Rutgers study.
Evaluating 754 clinical notes from the MIMIC-IV (Medical Information Mart for Intensive Care IV) database, researchers Liyang Xue Ph.D.’25, Distinguished Teaching Professor Mary Chayko, and Professor of Library and Information Science Vivek Singh at the Rutgers School of Communication and Information found that 62% of anonymized hospital discharge notes about gender-expansive patients included stigmatizing language, compared with 26% of comparison notes. This disparity remained even when misgendering was not counted as stigmatizing language. Misgendering refers to pronouns or other gendered terms that conflict with the patient's gender identity documented in the note.
To support future research, the team also created what they believe is the first annotated dataset designed specifically to study stigmatizing language in records of gender-expansive patients.
The study also revealed that higher occurrences of stigmatizing language also appear in the medical records of Black patients, indicating a need, the authors said, to examine documentation practices across multiple patient groups, and highlighting the importance of respectful documentation in delivering high-quality care for all patients.
Their findings also indicated that AI might offer a way to flag stigmatizing language. “The words in a medical record can shape how the next doctor sees a patient,” Xue said. “We found that gender-expansive patients were more often described in ways that could make them feel blamed, doubted, or disrespected. AI could help hospitals spot this language in records that would be difficult to review entirely by hand. Our study also shows why we need to check how well AI works for different patient groups and keep people involved in deciding what needs to change.”
The study, “Stigmatizing Language in Gender-Expansive Patient Records: Corpus Development, Disparity Analysis, and Natural Language Processing–Based Detection Study,” was published in the Journal of Medical Internet Research.
As part of the study, the researchers created an annotated dataset that includes gender-expansive patients and tested AI tools for detecting stigmatizing language across patient groups.
“The goal,” Xue said, “would be to preserve necessary medical information while avoiding language that unnecessarily blames or dismisses patients. This is a potential application of our findings. Future research could build on this study by testing the specific impact of these tools on patient care, and extending the investigation to other medical facilities, which could have different methods of documentation.”
The authors also showed that adjusting how an AI model flags notes could reduce differences in error rates between patient groups while maintaining similar overall accuracy. They also found that AI’s overall accuracy did not tell the whole story. “Several models were less accurate for gender-expansive patients’ notes, and previously trained models missed more stigmatizing language in this group,” Xue said. “These systems learn language patterns from examples, so testing them on diverse patient records matters.”
The researchers also found a promising way to reduce these differences in the models. “In our strongest-performing model, adjusting the rules for flagging notes reduced the gap in false-alarm rates between groups by more than half, while overall accuracy stayed around 83%,” the authors said.
This suggests, he added, that “AI can help identify documentation problems, provided its errors are checked carefully and people remain involved in reviewing its decisions.”
Learn more about the Library and Information Science Department, the Digital Communication, Information, and Media (DCIM) Minor, and the SC&I Behavioral Informatics Lab at the Rutgers School of Communication and Information.
