ACUTE ischemic stroke outcome prediction may be improved by a reasoning enhanced large language model that can extract prognostic value from routine clinical notes, according to findings presented at the AAN 2026 Annual Meeting in Chicago, Illinois, USA.
Stroke Outcome Prediction from Clinical Notes
Predicting recovery after acute ischemic stroke remains central to treatment planning, follow up, and counseling. Yet much of the clinical detail that shapes prognosis is buried in unstructured discharge summaries rather than neatly coded variables. This study evaluated whether a Chain of thought Outcome Prediction Engine, called COPE, could use that narrative information to predict 90 day modified Rankin Scale (mRS) outcomes more effectively than traditional approaches.
The analysis included 464 patients with acute ischemic stroke treated at a single center between 2010 and 2023, all of whom had discharge summaries and 90-day mRS scores available. COPE used a dual large language model framework in two stages. The first model generated clinical reasoning, and the second used that reasoning to predict functional outcome on the mRS scale.
How COPE Compared with Other Models
COPE achieved a mean absolute error of 1.00, with 75% of predictions falling within 1 mRS point of the observed outcome and exact accuracy of 33%. Those results matched GPT 4.1 across all primary measures. COPE also outperformed both Clinical BERT and a variable based support vector machine model, which each showed a mean absolute error of 1.28 and lower accuracy overall.
Importantly, the reasoning-based design appeared to matter. When investigators removed the reasoning component, model performance worsened, with exact accuracy falling to 23%. That finding suggests the intermediate reasoning step added clinically meaningful value rather than simply increasing model complexity.
Why the Findings Matter
The most informative parts of the discharge summaries were the Medications section and the Discharge and Follow up Summary. Removing either led to the largest performance drop in text ablation testing, highlighting where outcome related signals may be concentrated in routine documentation.
For clinicians, the appeal of this approach lies in its ability to work with the text already generated during care. The investigators described COPE as accurate, interpretable, and privacy preserving, offering a potential path toward stroke outcome prediction that does not rely solely on structured fields. While the findings remain early and were derived from a single center cohort, they point to a future in which narrative documentation could support 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.
Featured Image: amnaj on Adobe Stock.
