Abstract WMP81: A ChatGLM-based stroke diagnosis and prediction tool

• Published in: Stroke (1970) Vol. 56; no. Suppl_1; p. AWMP81
• Main Authors: Song, Xiaowei, Wang, Jiayi, Ma, Weizhi, Wu, Jian, Wang, Yueming, Gao, Ceshu, Wei, Chenming, Pi, Jingtao
• Format: Journal Article
• Language: English
• Published: Hagerstown, MD Lippincott Williams & Wilkins 01.02.2025
• Subjects: Access to care; Stroke; Artificial Intelligence
• ISSN: 0039-2499; 1524-4628
• DOI: 10.1161/str.56.suppl_1.WMP81

Background and purpose: Stroke as a world-wide prevalent disease, has brought a huge burden to health care and the national economy, accurate and fast stroke diagnosis can significantly increase reperfusion rate, mitigate disability, and reduce deaths. However, there exists a great discrepancy in acute stroke diagnosis and treatment due to the diverse medical information for making decisions. This study aims to develop a stroke diagnosis and prediction tool based on Large Language Models (LLM) to combine heterogeneous information for reasoning. Methods: By taking the electronic health record's (EHR) free-text information combined with non-contrast computed tomography (NCCT) to improve stroke discovery and treatment, We randomly included 1885 stroke and non-stroke subjects admitted at neurology ER in a comprehensive stroke center as a training set. We developed an LLM based on ChatGLM3-6B by selecting optimal entry combinations, using external tools, Instruction Tuning, and Low-Rank Adaptation (LoRA) techniques to enhance the performance of key procedures in stroke diagnosis flow-chart, and finally validating the results at both internal and external datasets. Results: The multimodal LLM based on clinical notes and NCCT has very high accuracy in stroke diagnosis (99.0% in the internal validation dataset, 95.5% and 79.1% in other 2 external test cohorts), distinguish ischemia and hemorrhage (100.0% in validation dataset, 99.1% and 97.1% in other test cohorts), LVO identification (80.0% in validation dataset, 88.6% and 83.3% in other test cohorts), and screening patients eligible for IVT (89.4% in validation dataset, 60.0% and 80.0% in other test cohorts). Conclusion: We derived an LLM that utilizes clinical text and NCCT to identify stroke and guide recanalization therapy. Our results require wide-scale deployment validation but can potentially improve stroke identification and narrow reperfusion time.