Developing aids to assist acute stroke diagnosis

Abstract

The only known therapy for stroke, a major leading cause of death and disability, has to be administered within 3 hours of the onset of symptoms for it to be effective. Accurately diagnosing a stroke as soon as possible after it occurs is difficult as it requires a subjective evaluation by a clinician in a hospital. With the narrow time window required for diagnosis, stroke evaluation would benefit from being aided by computational approaches that identify and quantify stroke symptoms in an efficient way. Here, we propose the design of a novel interface that provides clinicians with visualizations of the results of a machine learning-based technological aid for stroke diagnosis. To effectively support clinicians in determining stroke type, the proposed approach allows them to compare their own manual stroke evaluation with the results of the diagnostic system. By developing and evaluating our prototypes with neurologists, we explore how to best integrate technological aids into busy hospital workflows without burdening clinicians or biasing their decision making processes. We found that properly balancing the predictions of humans with that of technology is key to promoting the adoption of the latter in hospitals.