Detection of epileptic spasms using foundational AI and smartphone videos

Abstract

Infantile epileptic spasm syndrome (IESS) is a severe neurological disorder characterized by epileptic spasms (ES). Timely diagnosis is crucial, but it is often delayed due to symptom misidentification. Smartphone videos can aid in diagnosis, but the availability of specialist review is limited. We fine-tuned a foundational video model for ES detection using social media videos, thus addressing this clinical need and the challenge of data scarcity in rare disorders. Our model, trained on 141 children with 991 ES and 127 children without seizures, achieved high performance (area under the receiver–operating-characteristic curve (AUC) 0.96, 82% sensitivity, 90% specificity) including validation on external datasets from social media derived smartphone videos (93 children, 70 seizures, AUC 0.98, false alarm rate (FAR) 0.75%) and gold-standard video-EEG (22 children, 45 seizures, AUC 0.98, FAR 3.4%). We demonstrate the potential of smartphone videos for AI-powered analysis as the basis for accelerated IESS diagnosis and a novel strategy for the diagnosis of rare disorders.