A new AI model named COPE, presented at the AAN 2026 Annual Meeting, shows promise in improving acute ischemic stroke outcome prediction by extracting prognostic value from routine clinical notes. Predicting recovery after stroke is crucial for treatment planning, but much of this vital information is embedded in unstructured discharge summaries.
The study analyzed 464 patients with acute ischemic stroke, utilizing COPE's dual large language model framework. The first model generated clinical reasoning, while the second used this reasoning to predict 90-day modified Rankin Scale (mRS) outcomes.
COPE achieved a mean absolute error of 1.00, with 75% of predictions within 1 mRS point of the observed outcome. This performance matched GPT 4.1 and surpassed other models like Clinical BERT and a variable-based support vector machine, which had higher error rates.
Investigators found that removing the reasoning component of COPE significantly worsened performance, indicating the intermediate reasoning step adds meaningful clinical value. The most informative sections of discharge summaries for prediction were identified as the Medications section and the Discharge and Follow up Summary.
This approach is appealing to clinicians as it utilizes existing text generated during patient care. COPE is described as accurate, interpretable, and privacy-preserving, offering a path towards more personalized prognostication in acute ischemic stroke. While early, these findings suggest narrative documentation could become a key tool in stroke outcome prediction.