Rutgers Builds AI Tool to Remove Bias from Medical Records

Words in a medical chart can follow a patient for years, shaping how doctors treat them long after the ink dries.

Now researchers at Rutgers University have built an AI tool that catches biased language before it does harm. Their system flags stigmatizing words in patient records with 83% accuracy, giving hospitals a practical way to review and fix notes that could damage future care.

The three-person research team examined 754 hospital discharge notes and found that 62% of records for transgender and non-binary patients contained language that made them seem blameworthy, unreliable, or undeserving of care. Only 26% of comparison records showed similar problems. The gap persisted even when researchers excluded misgendering and counted only other forms of bias.

"Words in a medical record shape how the next clinician perceives a patient," explained Liyang Xue, who completed his doctorate at Rutgers in 2025 and led the study. Those perceptions travel across care encounters and erode trust over time.

The team also found elevated rates of stigmatizing language in records written about Black patients. That pattern suggests the problem extends well beyond any single patient population.

The Ripple Effect

To help other researchers tackle bias in medical documentation, the Rutgers team built the first public dataset specifically designed to study stigmatizing language in gender-expansive patient records. Other scientists can now use it to develop better detection tools.

The team tested several AI models and discovered that systems trained on general medical notes often missed bias in records for transgender and non-binary patients. The models only learned to recognize what they had seen before. Testing against diverse patient records became essential before any hospital could use them safely.

The strongest model showed a clear path forward. By adjusting when the system flags a note, researchers cut the difference in false-alarm rates between patient groups by more than half. Overall accuracy held steady at roughly 83%.

The team emphasized that AI only does part of the work. Hospitals can use the technology to surface candidates for review when examining large volumes of historical notes, but clinical staff and patients must make the final judgment calls about what language needs to change.

Distinguished Teaching Professor Mary Chayko and Professor Vivek Singh joined Xue on the research team. They published their findings on September 11, 2026, framing responsible documentation as a core component of high-quality care for all patients.

The tool gives hospitals a practical starting point for cleaning up records that shape patient outcomes for years to come.