COIOS
Emergent technology Item new preprint

LLM-extracted autopsy brain features classify neurodegenerative disease with 75% accuracy

A fine-tuned large language model scored 39 features from narrative gross descriptions in 5,613 Mayo Clinic brain autopsies, and classification reached 0.75 accuracy, from 0.93 sensitivity for progressive supranuclear palsy to 0.03 for mixed Alzheimer–Lewy body disease.

Researchers fine-tuned a large language model to extract 39 semi-quantitative features from the narrative gross descriptions in 5,613 autopsy reports held by the Mayo Clinic Brain Bank between 1998 and 2023, then classified each case into one of seven neurodegenerative diagnoses. In a held-out set of 562 cases, accuracy was 0.75 and Cohen's kappa 0.68, with macro-average sensitivity 0.66. Sensitivity was 0.93 for progressive supranuclear palsy and 0.92 for multiple system atrophy, but 0.03 for combined Alzheimer and Lewy body disease.

Why it is interesting: Shows how much diagnostic signal sits unstructured in existing autopsy text, and where the mixed pathologies that dominate dementia deaths defeat it.

Source
medRxiv (preprint), 2026-09-14
DOI
10.64898/2026.09.12.26362866
Type
Preprint
Design
Retrospective classification study of 5,613 Mayo Clinic Brain Bank autopsies, 1998–2023, held-out test set of 562
Verdict
New finding
Driver
AI in medicine