Abstract 94: Artificial Intelligence-Based Application to Detect and Quantify Stroke Symptoms

Abstract

Assessment of patients suspected to have an acute stroke at the prehospital level or non-stroke centers may be challenging. We aimed to validate a novel Artificial Intelligence (AI) based application (CVAid) for smartphones able to score neurological deficit in acute stroke patients. Methods: Acute stroke patients, admitted in two different stroke units, and healthy volunteers were studied. A certified stroke neurologist determined NIHSS at bedside. A different user recorded a short neurological exam, using CVAid application. The software processed the video file presenting comprehensive neurological exam per NIHSS test to a remote neurologist on a dedicated tablet. Two different remote neurologists (remote1 and remote2) determined NIHSS reviewing neurological exam in the tablet. CVAid performed also a completely automated AI analysis of facial features, to detect stroke symptoms. Results: A total of 75 patients were included in the study, 64 (85.3%) stroke patients and 11 (14.7%) healthy volunteers. Correlation between bedside and remote1 NIHSS was r=0. 861 (p<0.001). Correlation between remote1 and remote2 NIHSS was r=0.865 (p=0.001). The CVAId facial recognition system showed an accuracy of 87% in stroke symptom detection, sensitivity 95%, Specificity 80%, false negative 5%, false positive 20%. Conclusions: An AI based application efficiently allowed remote neurological evaluation of stroke patients. The automated algorithm was able to accurately triage healthy volunteers from patients suffering a stroke. Feeding the machine learning system with additional patients and development of additional automated analysis beyond facial recognition will increase the system accuracy.