Objective:

We developed a multimodular expert system for stroke vascular localization based on verbatim clinical text using contemporary LLM-driven approaches.

Background:

Being able to neurologically localize to a vascular territory is essential in distinguishing true strokes from mimics—a challenge in settings with limited neurologist availability.

Design/Methods:

Chief complaints, neurological examinations, and neuroradiologic imaging were extracted from 369 cases in the MIMIC-IV-Note database and processed through a three-layered multimodular system across three trials. The first layer utilized gemini-2.5-flash-lite to structure clinical free-text into symptom surveys. The second layer generated regional brain predictions through a scoring system, while the third mapped these predictions to major vascular territories using rule-based logic.

Results:

The stroke symptom survey, brain-region, and vascular-territory predictions achieved average Jaccard indices of 0.81, 0.88, and 0.93, respectively across three trials. Lesion-side were highly reliable (accuracy 0.91, F1 0.91), while brain region-level localization was more challenging (accuracy 0.71, F1 0.51). The vascular-territory classifier showed strong overall performance (accuracy 0.93, F1 0.92), maintaining high accuracy even when combined with side predictions (accuracy 0.93, F1 0.86). Across vascular territories, middle cerebral artery predictions achieved an accuracy of 0.93 and F1 score of 0.96. The posterior cerebral artery achieved the highest accuracy (0.99) and F1 of 0.85. The basilar artery showed an accuracy of 0.96 and F1 of 0.56. Vertebral predictions showed accuracy of 0.98 and F1 of 0.71.

Conclusions:

The new multimodular system demonstrates that integrating large language models for structured data extraction with transparent rule-based reasoning provides both flexibility and stability in stroke localization from free-text clinical documentation, enabling its use as a powerful clinical, educational, and research tool.