Adapting Action Recognition Neural Networks for Automated Infantile Spasm Detection

Abstract

Infantile spasms are a severe epileptic syndrome characterized by short muscular contractions lasting from 0.5 to 2 seconds. They are often misdiagnosed due to their atypical presentation, and treatment is frequently delayed, leading to stagnation or regression in psychomotor development and significant cognitive and motor sequelae. One promising approach to addressing this issue is the use of markerless computer vision techniques. In this paper, we introduce a novel approach for recognizing infantile spasms based exclusively on video data. We utilize an expanded 3D neural network pre-trained on an extensive human action recognition dataset called Kinetics. By employing this model, we extract features from short segments of varying sizes sampled from seizure videos, which allows us to effectively capture the spatio-temporal characteristics of infantile spasms. We then apply multiple classifiers to perform binary classification on these extracted features. The best system achieved an average area under the ROC curve of $0.813\pm 0.058$ for a 3-second window.

Keywords

• Infantile spasm
• video-based monitoring
• X3D-M
• dimension reduction
• action recognition