Deep learning architectures for EEG-based classification of Dravet syndrome: A comparative study of pre-trained and non-pretrained hybrid CNN-LSTM models

Abstract

Objective This study explores the potential of artificial intelligence (AI) using a hybrid deep learning Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) framework, for EEG-based classification and analysis of Dravet Syndrome (DS). Method The study cohort comprised nine pediatric patients with DS, confirmed through either a heterozygous pathogenic mutation in the SCN1A gene or a clinical diagnosis consistent with established diagnostic criteria. In addition, EEG recordings from age-matched healthy control subjects and pediatric patients with non-Dravet epilepsy (“abnormal” EEG) were included. Data on demographic information, seizure characteristics, developmental skills, cognitive functions, and genetic results were gathered from patient records. EEG recordings were analyzed using a subject-independent leave-one-subject-out validation strategy, spatial and temporal features by employing this model on preprocessed EEG data, effectively differentiating DS patients from Abnormal cases and healthy controls. Result Among the evaluated CNN-LSTM models, the pre-trained architecture achieved superior performance with improved stability across most subjects, with an overall accuracy of 85%, balanced accuracy of 85%, a macro-averaged F1-score of 0.85, and a macro-averaged ROC–AUC of 0.87, demonstrating stable performance for multi-class EEG classification of DS, Abnormal, and control subjects. The non-pretrained model showed reduced sensitivity and increased inter-class confusion, particularly for DS and Abnormal classes. Conclusion This study demonstrates that a pre-trained CNN-LSTM framework can support automated EEG-based classification of DS-related patterns as a proof-of-concept methodological approach, even in the context of limited subject availability. EEG-specific pretraining improves classification consistency and feature separability compared with training from scratch, highlighting the value of representation learning for rare epilepsy syndromes. Larger multi-center datasets and prospective validation will be required to assess robustness, generalizability, and clinical utility.