Artificial Intelligence Assisted Stroke Diagnosis Using Commercial Tools for Emergency Settings

Background Timely and accurate diagnosis of stroke is critical for optimizing patient outcomes, particularly in emergency settings. Artificial intelligence (AI)-based neuroimaging tools have emerged as potential adjuncts to assist clinicians in stroke detection. However, real-world validation of AI-assisted stroke diagnosis remains limited. This study aims to evaluate the clinical performance of JBS-01K (for ischemic stroke) and AVIEW NeuroCAD (for hemorrhagic stroke) in an emergency department setting. Methods A prospective clinical validation study was conducted using stroke registry data from Keimyung University Dongsan Hospital and Chilgok Kyungpook National University Hospital. A total of 900 patients were enrolled, including 200 ischemic stroke cases, 400 hemorrhagic stroke cases, and 300 non-stroke controls. JBS-01K was used for infarct detection in diffusion-weighted imaging (DWI) MRI, while AVIEW NeuroCAD was applied to non-contrast CT scans for hemorrhagic stroke detection. The primary endpoint was F1-score, while secondary endpoints included sensitivity, specificity, precision, and accuracy. Statistical analyses were performed using R (version 4.2.1). Results AVIEW NeuroCAD demonstrated a sensitivity of 0.97, specificity of 0.935, precision of 0.788, and an overall accuracy of 0.942 for hemorrhagic stroke detection. The performance improved slightly after clinician corrections. JBS-01K achieved a sensitivity of 0.98, specificity of 0.795, precision of 0.833, and an accuracy of 0.89 for ischemic stroke detection. False positive analysis revealed that AVIEW NeuroCAD occasionally misclassified physiological calcifications as hemorrhages, while JBS-01K exhibited errors in artifact-prone regions such as the pons. Despite these limitations, real-time AI-assisted decision support enhanced workflow efficiency by providing rapid alerts for potential stroke cases. Conclusion This study provides real-world clinical validation of AI-assisted stroke detection using commercially available software. While both JBS-01K and AVIEW NeuroCAD demonstrated high diagnostic performance, model limitations, particularly in small lesion detection and false positive cases, highlight the need for continued refinement. AI-based neuroimaging tools can enhance radiological workflows and expedite stroke diagnosis in emergency settings, but human oversight remains essential. Further research should focus on improving AI specificity, reducing false positives, and integrating AI into routine clinical practice to optimize stroke management outcomes.