To develop and evaluate an artificial intelligence-based clinical decision support system to improve the accuracy and clinical utility of stroke diagnosis in pre-hospital settings and to ultimately optimize time to treatment decisions, allowing more patients to benefit from  thrombolytics.

Stroke remains one of the leading causes of mortality and disability worldwide. Timely and accurate differentiation between stroke and non-stroke patients in pre-hospital environments is critical to improve outcomes and avoid unnecessary transfers. Yet current clinical stroke scales and predictive models often lack sufficient accuracy and ease of use under real-world conditions, resulting in patients sometimes being routed to a hospital that cannot meet their clinical needs.

We retrospectively analyzed the Geisinger FABS dataset (n=780) creating two independent cohorts:  a training set comprised of 633 patients, and a test set comprised of 113 stroke patients and 46 non-stroke patients.    The trained models examined a number of features including demographics, patient history and facial droop.

Our approach improved stoke prediction accuracy over EMS stroke scoring models by increasing sensitivity to 93-96% while maintaining specificity levels comparable to the FAST scale (sensitivity 77%, specificity 60%).  By correcting data, we increased sensitivity to 96% with little impact on specificity (64% versus 65%).  We have developed a proof of concept based on only one NIHSS clinical covariate suggesting the potential for new models with other clinical covariates, such as limb weakness and slurred speech, to more accurately predict stroke and stroke mimics in a pre-hospital setting.

Our analysis suggests the ability to develop a clinical decision support tool that is a clinically interpretable, scalable approach to improve pre-hospital stroke triage and enable optimal ambulance routing. The next step is to develop a device for first responders to collect patient data and then to train a model on additional co-variates to further increase the predictive ability.