2025 American Academy of Neurology Abstract Website

Objective:

Based on our previous study to localize stroke lesions, we tested GPT-4 to predict stroke etiology using several key elements: brain imaging, brain vessel imaging, neck vessel imaging, transthoracic echocardiogram (TTE), and history of atrial flutter/fibrillation (AFL/AF).

Background:

Determining stroke etiology requires several factors to be weighed and taken into consideration.

Design/Methods:

Key elements used to predict stroke etiology included brain and vessel imaging, echocardiography, a history of atrial flutter/fibrillation (AFL/AF), patient history, and laboratory data. These factors were obtained from 14 case reports and processed across five trials using GPT-4. Prompt engineering involved text classification, applying the TOAST classification system. Performance metrics, including specificity, sensitivity, precision, and F1 score, were calculated by comparing GPT-4’s predictions with the etiologies described in the case reports.

Results:

The overall performance metrics demonstrated a specificity of 0.99, sensitivity of 0.97, precision of 0.96, and an F1 score of 0.97. Cardioembolic and small vessel occlusion etiologies achieved 100% sensitivity, specificity, precision, and F1 scores. Although sensitivity remained high at 0.91 for large vessel atherosclerosis, precision decreased to 0.83, resulting in an F1 score of 0.87. The "other determined" category also performed consistently well, with a specificity of 0.97, sensitivity of 0.97, precision of 0.97, and an F1 score of 0.97. This result demonstrates GPT-4's accuracy in predicting stroke etiology, with variability performance observed only in large vessel atherosclerosis cases.

Conclusions:

GPT-4 demonstrated high accuracy in predicting stroke etiology which suggests GPT-4’s potential utility in assisting to determine stroke etiology. However, further research with larger datasets and actual clinical data is required to test its validity.