Infantile epilepsy detection: A breakthrough video-based approach for accurate identification of infantile spasms

Shenyang Institute of Computing Technology, CAS and Chinese PLA General Hospital Joint Team conduct series of investigations on Infantile Spasms Syndrome (IESS), also known as West syndrome, discovering a video-based epileptic seizure detection method that effectively enhances the accuracy of infantile spasm identification.

1. Background Introduction and Problem Bottleneck

IESS is an epileptic encephalopathy that manifests during infancy, characterized by unique epileptic seizures, including repeated muscle contractions, extensions, or alternating flexion-extension spasms. These seizures are accompanied by high-amplitude electroencephalogram (EEG) waveforms, known as hypsarrhythmia. IESS has adverse prognostic implications for intellectual development. In clinical practice, precise monitoring of bedridden patients' movements is crucial for effective disease management and epileptic seizure diagnosis. However, even experienced EEG technicians face challenges when analyzing relevant data.

2. Research Opportunity and Discovery

Given the massive generation of EEG data, the susceptibility of signal interpretation to interference, and the potential comfort issues for infants and young children when wearing EEG devices, we explored a video-based epileptic seizure detection method utilizing feature recognition. This method aims to simplify the assessment process, reduce non-medical expenditures, and ensure continuous evaluation of the patient's condition.

3. Brief Summary of Research Content

This study initially integrated target detection technology into the video data processing stage to accurately locate patients in clinical monitoring videos, thereby extracting video clips that solely contain the patients. Subsequently, an enhanced 3D-ResNet was employed for video-based IESS detection. This method utilizes an optimized 3DResNet-50 architecture, which deeply extracts local key features from the video through asymmetric convolution and CBR modules, and introduces a 3D Convolutional Block Attention Module (CBAM) to enhance the spatial correlation between channels in video frames.

4. Current Challenges and Future Directions

Currently, the main challenges faced by the research include issues such as occlusion, lighting variations, and similar human body interferences during the identification process. Future research directions will focus on further enhancing the network's generalization capability, optimizing algorithms to address various challenges in practical applications, and exploring more AI-based solutions to alleviate the workload of doctors when screening VEEG data.