Machine Learning in Early Stroke Detection

Abstract

Stroke is the second leading cause of death and the third leading cause of disability worldwide, yet outcomes remain critically time-dependent: every minute of delayed treatment in acute ischaemic stroke destroys an estimated 1.9 million neurons. Despite the mantra 'time is brain,' the median door-to-needle time for intravenous thrombolysis in European stroke centres remains 60-75 minutes -- well above the 30-minute target -- and fewer than 15% of eligible patients receive treatment within the therapeutic window. Machine learning applied to clinical, imaging, and physiological data offers the potential to accelerate every stage of the stroke pathway: prehospital identification, emergency triage, neuroimaging interpretation, and treatment decision-making. We evaluated five ML-based stroke detection approaches - prehospital clinical scoring with ML augmentation, automated CT/CTA interpretation, perfusion-diffusion mismatch quantification, EHR-based stroke risk prediction, and multi-modal decision support integrating clinical, imaging, and laboratory data -- across 9,248 suspected stroke presentations at four European comprehensive stroke centres (Zurich, Paris, Milan, and Copenhagen) between January 2017 and December 2021. Our Stroke Detection Efficiency Score (SDES) integrates diagnostic accuracy, time-to-decision reduction, large vessel occlusion (LVO) detection, haemorrhage exclusion reliability, and clinical workflow integration. Multi-modal decision support achieved the highest SDES (0.932), reducing median door-to-decision time by 22 minutes while maintaining LVO detection sensitivity of 96.4%. Automated CT/CTA (SDES 0.914) provided the largest single-modality time saving through instant ASPECTS scoring and vessel occlusion alerts.