JMIR Publications

ABSTRACT

Background:

Stroke is a globally prevalent disease that imposes a significant burden on healthcare systems and national economies. Accurate and rapid stroke diagnosis can substantially increase reperfusion rates, mitigate disability, and reduce mortality. However, there are considerable discrepancies in the diagnosis and treatment of acute stroke.

Objective:

The aim of this study is to develop a stroke diagnosis and prediction tool based on ChatGLM3-6B, which utilizes free-text information from electronic health records (EHR) in conjunction with non-contrast computed tomography (NCCT) to enhance stroke detection and treatment.

Methods:

We utilized the free-text information from electronic health records (EHR) in conjunction with non-contrast computed tomography (NCCT) to enhance the detection and treatment of strokes. A total of 1,885 subjects, both stroke and non-stroke patients, were randomly selected from the neurology emergency room at a comprehensive stroke center to serve as our training set. We developed a large language model (LLM) based on ChatGLM3-6B by identifying optimal input combinations, employing external tools, and applying Instruction Tuning and Low-Rank Adaptation (LoRA) techniques. These strategies were implemented to improve the performance of critical procedures in the stroke diagnosis flowchart, and the results were subsequently validated using both internal and external datasets.

Results:

The multimodal LLM, which is based on clinical notes and NCCT, demonstrates exceptionally high accuracy in stroke diagnosis, achieving 99.0% in the internal validation dataset and 95.5% and 79.1% in two external test cohorts. It effectively distinguishes between ischemia and hemorrhage, with an accuracy of 100.0% in the validation dataset and 99.1% and 97.1% in the other test cohorts. Additionally, it identifies large vessel occlusions (LVO) with an accuracy of 80.0% in the validation dataset and 88.6% and 83.3% in the other test cohorts. Furthermore, it screens patients eligible for intravenous thrombolysis (IVT) with an accuracy of 89.4% in the validation dataset and 60.0% and 80.0% in the other test cohorts.

Conclusions:

We developed a large language model (LLM) that leverages clinical text and non-contrast computed tomography (NCCT) to identify strokes and guide recanalization therapy. While our results necessitate validation through widespread deployment, they hold the potential to enhance stroke identification and reduce reperfusion time.