Mixture of Expert System for IESS Detection Based on EEG Signal

Childhood epilepsy presents significant challenges in terms of medical diagnosis and treatment, particularly when it comes to infantile epileptic spasm syndrome (IESS). This condition is characterized by electroencephalogram (EEG) signals with short seizure duration and insignificant waveform, which complicates manual efforts to detect IESS seizures. To address these challenges, we introduce an innovative model called the IESS mixture of experts (MOE) model to assist EEG technologists in their work. The model utilizes a group of subdata experts operating through a 3D ResNet network to enhance the detection of seizure signals. Specifically, we propose an unsupervised clustering method for EEG signal partitioning into some balanced sub-data domains, addressing the imbalance between IESS seizure and non-seizure data. The IESS MOE model consists of a navigation network that controls input and output, as well as some chosen sub-expert networks that make judgments on sub-expert data domains. We use the PLA IESS dataset for unsupervised learning-based data partitioning among sub-models and subsequent expert modeling. Additionally, we establish a comparison between the performance of the IESS MOE model and six human experts from different levels of medical centers using external validation on 19 cross-individual patient data sets. Finally, test results demonstrate that the sensitivity of the IESS MOE model is 0.539 3, surpassing the average level achieved by human experts at 0.514 3 while maintaining high specificity at 0.905 0 and achieving a high F1 score at 0.675 8, confirming its superior generalization ability. Our proposed model can assist human experts in monitoring episodes related to IESS.