Related Experiment Video
Updated: Jul 10, 2026
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Symptom-Only Localization of Brainstem Ischemia Using Large Language Models Versus Neurologists in Diffusion-Weighted
Nedim Beste1, Thomas Dratsch1, Jonathan Kottlors1
1University Hospital Cologne, Kerpener Str 62, Cologne, 50937, Germany, 49 2214780.
Background:
Symptom-based localization of brainstem ischemia is challenging because of the anatomical complexity of the brainstem and the nonspecific overlap of clinical syndromes. Whether large language models (LLMs) can meaningfully assist in this task remains uncertain.
Objective:
This study aimed to compare the performance of several OpenAI LLMs and neurologists in localizing diffusion-weighted imaging (DWI)-confirmed brainstem ischemic lesions based on symptom descriptions alone.
Methods:
In this retrospective single-center study, 109 patients with DWI-confirmed acute brainstem ischemia were included. Three neurologists and 6 LLMs (GPT-5, GPT-4, GPT-4.1, GPT-4o, o3, and o3-pro) predicted lesion localization using a combined anatomical-lateral end point (left or right midbrain, pons, and medulla) based on symptom descriptions alone. Overall and regional accuracy, the Cohen κ, 6-class confusion matrices, and point-biserial correlations between symptom count and correct prediction were assessed. Because all raters evaluated the same cases, paired McNemar tests with Benjamini-Hochberg correction were used for pairwise performance comparisons.
Results:
GPT-4 and GPT-4o achieved the highest overall accuracy (61/109, 56%; 95% CI 46.1%-65.5%). Agreement with the DWI reference standard remained limited across all raters, with the Cohen κ reaching a maximum of 0.291 for GPT-4o. Confusion matrices showed that higher performance was driven mainly by pontine cases, whereas misclassification remained frequent in mesencephalic and medullary lesions. Regional analyses outside the pons were imprecise because mesencephalic and medullary subgroups each contained only 16 cases. A higher number of documented symptoms was associated with correct prediction for GPT-4, GPT-5, GPT-o3, and 1 neurologist.
Conclusions:
Although some LLMs showed higher relative accuracy than the participating neurologists, absolute performance remained limited and clinically insufficient. These findings are best interpreted as an exploratory benchmark under constrained conditions: absolute performance remained modest, agreement beyond chance was limited, and performance outside pontine lesions was inconclusive.
