A New Class of Hybrid LSTM-VSMN for Epileptic EEG Signal Generation and Classification

Epilepsy is a widespread neurological disorder affecting approximately 50 million people worldwide, significantly impacting quality of life and placing a heavy burden on healthcare systems. Early and reliable seizure detection remains a critical challenge, often hindered by limited availability of high-quality electroencephalogram (EEG) data and the suboptimal performance of existing classification methods. In this work, we propose a novel two-stage framework that addresses both data scarcity and classification accuracy. The first stage involves generating synthetic EEG signals that realistically mimic epileptic patterns using the Variable Structure Model Neuron (VSMN) with multidendrites, providing an effective means of data augmentation. In the second stage, we introduce a hybrid LSTM-VSMN model, where the VSMN activation function is integrated within the Long Short-Term Memory (LSTM) network gates, replacing conventional activations such as tanh. This integration improves the model’s ability to capture complex temporal dependencies in EEG sequences. To the best of our knowledge, this is the first study to leverage VSMN both for EEG signal synthesis and as an activation function within a deep recurrent neural network for seizure detection. The proposed model is rigorously evaluated against conventional activation functions, achieving an accuracy of 98.16% in single-fold validation and 97.59% under 3-fold cross-validation. Furthermore, it achieves a Mean Absolute Error (MAE) as low as 0.0241 and a Mean Absolute Percentage Error (MAPE) of 2.41%, substantially outperforming baseline approaches. These results demonstrate the effectiveness of the hybrid LSTM-VSMN architecture in enhancing automated seizure detection, offering a promising tool for clinical decision support and real-time monitoring applications.