AI can now extract stigma from HIV patient notes — but only catches 62% of cases

Researchers built the first machine learning tool that can read clinical notes and identify mentions of HIV-related stigma — the psychosocial stress that degrades treatment outcomes. The tool works well enough to automate what used to require human review, but it misses some cases and struggles with certain types of stigma, particularly when patients internalize shame rather than facing external judgment.

HIV clinicians have always known that stigma hammers treatment adherence and mental health outcomes, but they've had no systematic way to track it across patient populations or identify which patients need extra support. This tool makes that tracking possible at scale — a hospital system can now scan thousands of patient notes and flag cases automatically instead of waiting for doctors to notice and document it. The catch is real: the best model catches only 62% of stigma instances, and it fails completely on the type most connected to internalized shame. That means hospitals will see better data, but incomplete data — they'll spot some problems they'd otherwise miss while remaining blind to others.