Early Stroke Detection Through Machine Learning in the Prehospital Setting

María Ríos^1*Gemma Reig²Nicolas Riera³Jose A. Vivancos²José Ayala¹

• ¹Complutense University of Madrid, Madrid, Spain
• ²Servicio de Neurología. Hospital Universitario de La Princesa, IIS-Princesa - Instituto de Investigacion Sanitaria Hospital Universitario de La Princesa, Madrid, Spain
• ³Stroke Commission. Madrid Emergency Medical Service (SUMMA 112), Madrid, Spain

Stroke is a leading cause of death and disability globally, with rising prevalence driven by modern lifestyle factors. Despite the critical nature of stroke as a time-sensitive condition requiring prompt diagnosis, current pre-diagnostic practices are often limited by reliance on specific patient symptoms, which can delay appropriate treatment, especially for large vessel occlusions (LVO). This study introduces a novel approach utilizing machine learning (ML) to accurately identify stroke type and severity. By enhancing the pre-hospital diagnosis process, the research aims to optimize hospital selection and improve emergency stroke care, ensuring timely treatment at specialized centers. The methodology consists on two phases. The first step involves developing two specialized models to predict the type of stroke -ischemic or hemorrhagic- along with a Bayesian rule to determine the final classification. The second step, applied only in cases of ischemic stroke, identifies whether the episode is a LVO or not. The results for ischemic episodes showed that the LVO model achieved 91.67% recall and 64.71% precision, outperforming the prehospital scale used as a reference in all performance metrics except specificity. This model utilized only 20 out of the 271 original variables, with the most representative variables including blood pressure, heart rate, oxygen saturation, and arm movement. The integration of the LVO model for the complete sample with a Bayesian pipeline resulted in a precision of 59% and a recall of 74%, while applying the LVO model to the entire population yielded a precision of 60.60% and a recall of 80.19%. The study concluded that the implementation of ML techniques can significantly improve the diagnostic accuracy of stroke in the context of Emergency Medical Services (EMS). The LVO model demonstrated promising results, with an improvement in approximately 10-13% compared to the baseline paradigm. The use of objective variables, such as blood pressure and heart rate, was a key factor in this enhancement. The study highlights the potential benefits of leveraging ML techniques in Emergency Medicine, particularly in the diagnosis and management of stroke. The results suggest that the LVO model can potentially augment the precision of stroke diagnosis, facilitating more efficacious and timely interventions.