Infantile Epileptic Spasms Accurately Detected Through Home Video Analysis
Automated analysis of home videos using a computer model was successful in detecting seizures in infants with epileptic spasms (ES), according to results of a phase 2 retrospective study presented at the American Epilepsy Society (AES) 2024 Annual Meeting. The computer model had an 81% accuracy rate in detecting ES, with a false alarm rate (FAR) under 10%, demonstrating high performance in an in-field cohort with a high degree of heterogeneity.
Researchers developed a computer model trained using 5-second video segments, each segment depicting a clearly visible infant with semiology recognizable by a specialist. Video segments from 152 infants with ES were annotated based on the presence of seizure, with 991 seizure and 597 nonseizure video segments. The model was further trained using 1385 video segments from 127 healthy infants.
Researchers used the model to detect ES in the video segments, demonstrating:
- An area under the operating-receiving-characteristic curve (AUC) of 0.94; an AUC of 1.0 represents a perfect model
- A sensitivity of 78%; measuring the percentage of correctly identified videos positive for ES
- A specificity of 85%; measuring the percentage of correctly identified videos negative for ES
- An accuracy of 81%
FAR was calculated based on 666 nonseizure video segments from an additional group of 67 healthy infants. The average FAR was 5.7%, meaning that 37 of these 666 additional segments were wrongly classified.
Researchers applied a 90% sensitivity threshold to the validation data, revealing:
- An AUC of 0.94
- A sensitivity of 88%
- A specificity of 66%
- An accuracy of 79%
- A FAR of 7.3% for the 666 out-of-sample video segments
The Berlin Institute of Health provided grant funding for the study.