大型语言模型预测急性缺血性中风后的功能结果

Accurate prediction of functional outcomes after acute ischemic stroke is crucial for clinical decision-making and resource allocation. Previous studies on modified Rankin Scale (mRS) prediction have primarily relied on structured variables such as age and NIHSS, along with conventional machine learning methods. The potential of large language models (LLMs) to directly infer future mRS scores from routine admission notes remains largely unexplored. This study evaluated encoder models (BERT, NYUTron) and generative models (Llama-3.1-8B, MedGemma-4B) in both frozen and fine-tuned settings for predicting discharge and 90-day mRS outcomes using a large, real-world stroke registry. The discharge outcome dataset comprised 9,485 History and Physical notes, while the 90-day outcome dataset included 1,898 notes.