The TAMJISTROKE expert system for stroke localization from clinical free text

Abstract TAMJISTROKE is a transparent hybrid expert system that localizes stroke lesions from real-world clinical notes using LLM-assisted text processing and rule-based reasoning. We conducted an observational study using deidentified MIMIC-IV discharge summaries, applying TAMJISTROKE, a three-layer modular pipeline that integrates large language model-based structured symptom extraction from chief complaints and neurologic examinations with rule-score-based regional and vascular territory localization. A total of 367 filtered ischemic stroke cases were evaluated across three trials, with performance measured with accuracy, sensitivity, specificity, F1 score, and structured error analyses. TAMJISTROKE showed high functional output. Brain region localization was more challenging (accuracy 0.71, sensitivity 0.73, specificity 0.70, F1 0.51), performing better on hemispheric than posterior fossa lesions. Vascular territory prediction was strong (accuracy 0.97, sensitivity 0.93, specificity 0.98, micro-averaged F1 0.93, macro-averaged F1 0.61, versus naïve baseline micro/macro F1 of 0.87 and 0.19), though posterior fossa strokes again lagged. Case-based analysis aligned with the trial-based results. Error analysis identified ambiguous or broadly localizable clinical documentation as the main source of errors; excluding these cases in post-hoc analysis improved performance across all vascular territories. By combining LLM-based symptom structuring with rule-based localization, this multimodular system achieved reproducible stroke localization from real-world free-text clinical notes, highlighting its potential utility for education, retrospective clinical review, and research.