Eko Arisetijono Marhaendraputro, Shelby Amrus Ernanda, Intan Dwi Karlina, Raehani Ajeng Prabaswaratna Nindyasti, Rivan Tegar Syahriar
Accurate and timely stroke type diagnosis (hemorrhagic vs. ischemic infarction) is crucial for treatment decisions. Limited access to CT scans in resource-constrained settings can hinder diagnosis. This study investigates the feasibility of machine learning (ML) for stroke type prediction using readily available clinical data. We evaluated Brain.js, a JavaScript library, for stroke type prediction. Anonymized data (n=138) from neurology study program morning reports (2021-2024) was used. Inclusion criteria ensured stroke onset within 24 hours and no prior hospital referral. Data included demographics, clinical presentation, and medical history. Head CT scan results served as the gold standard for stroke type. Data was split 98/40 for training and testing a Brain.js ML model. Model performance was evaluated using accuracy, sensitivity, and specificity. The Brain.js model achieved an accuracy of 75%, sensitivity of 71%, and specificity of 79% in predicting stroke type on unseen test data with hemorrhagic stroke as target test. This study demonstrates the potential of Brain.js for accurate stroke type prediction using readily available clinical data. This approach may be particularly valuable in settings with limited CT scan access, potentially aiding in early stroke diagnosis and treatment decisions. © 2025 The Authors, published by EDP Sciences.
Neurology Department, Faculty of Medicine, Universitas Brawijaya, Malang, Indonesia; Resident of Neurology Department, Faculty of Medicine, Universitas Brawijaya, Malang, Indonesia; Medical Doctor Profession Program, Faculty of Medicine Universitas Islam Malang, Malang, Indonesia; Brawijaya University Clinic, Universitas Brawijaya, Malang, Indonesia; Independent Software Engineer, Malang, Indonesia
