LLMs Predict Stroke Recovery with 76% Accuracy Using Just Doctor's Notes

A new study shows fine-tuned LLMs like Llama-3.1 can predict functional outcomes for stroke patients directly from routine admission notes, bypassing manual data entry. On a dataset of nearly 9,500 notes, the model achieved 76.3% accuracy for 90-day outcomes and 75% for discharge predictions. This performance is comparable to traditional models that require structured data like age and NIHSS scores, but works from raw clinical text.