A reasoning-enhanced large language model called COPE may improve prediction of 90-day functional outcomes after acute ischemic stroke by extracting prognostic signals from routine clinical discharge notes, researchers reported at the AAN 2026 Annual Meeting in Chicago (18–22 April 2026). The study found COPE matched a commercial model and outperformed other text- and variable-based approaches on key accuracy measures while remaining interpretable and privacy-preserving.
Predicting recovery after acute ischemic stroke is essential for treatment planning, follow-up and counseling, but much prognostic detail is locked in unstructured narrative text rather than in coded fields. The investigators tested whether COPE, a Chain-of-thought Outcome Prediction Engine, could use those narratives to predict the modified Rankin Scale (mRS) score at 90 days more effectively than standard methods.
The analysis included 464 patients treated at a single center between 2010 and 2023 who had both discharge summaries and 90-day mRS scores. COPE uses a two-stage, dual large language model framework: the first model generates explicit clinical reasoning from the note, and the second model uses that reasoning to predict functional outcome on the mRS scale.
COPE achieved a mean absolute error (MAE) of 1.00, with 75% of predictions falling within one mRS point of the observed outcome and exact-match accuracy of 33%. Those results matched GPT‑4.1 across the primary measures. By contrast, Clinical BERT and a variable-based support vector machine each showed an MAE of 1.28 and lower overall accuracy.
The study also tested whether the intermediate reasoning step mattered. When investigators removed COPE’s reasoning component, exact accuracy fell to 23%, suggesting the chain-of-thought stage provided clinically meaningful information rather than merely increasing model complexity.
Text ablation experiments identified the Medications section and the Discharge and Follow‑up Summary as the most informative parts of the notes; removing either produced the largest drops in performance. That finding points to where outcome-related signals may concentrate in routine documentation.
Investigators described COPE as accurate, interpretable and privacy-preserving, emphasizing its ability to work with text already produced during care. They cautioned the results are preliminary and derived from a single-center cohort, but suggested narrative documentation could become a practical adjunct for more personalized prognostication in acute ischemic stroke.
Reference: Liu Y et al., “COPE: Chain-of-thought Prediction Engine for Open-source Large Language Model Based Stroke Outcome Prediction from Clinical Notes.” Abstract 001, AAN Annual Meeting, 18–22 April 2026.
Original Source: https://www.emjreviews.com/neurology/news/aan-2026-stroke-outcome-prediction-improved-by-clinical-notes/
Category:
Tags:
Publish Date: 2026-04-20 04:03:00
