Symmetric Source Separation and Improved 1-Dimenstional ResNet Model for Epileptic Seizure Detection

Abstract

Repeated seizures brought on by irregular electrical activity in the brain characterize epilepsy, a long-term neurological condition. In order to improve outcomes and avoid cognitive decline, early detection is essential. Electroencephalography (EEG) is a key diagnostic tool, but manual interpretation is prone to delays and variability, prompting the need for automated methods using Artificial Intelligence (AI). While Machine Learning (ML) relies on handcrafted features, leading to limitations in complex scenarios, Deep Learning (DL) methods improve performance by extracting features directly from raw data. DL methods often require large training datasets for effective model learning. In the case of epileptic EEG samples, collecting a substantial number of labeled signals is impractical. So, the detection can rely on a few-shot learning scenario, where there are limited labeled samples available. However, directly using raw data in such cases risks local optima, overfitting, and reduced test accuracy. To solve these issues, Empirical mode decomposition with Power Spectral density and One-dimensional Convolutional Neural Network with ResNet (EMD-PSD-CNN1D-5Res) model was introduced for epilepsy detection. However, it suffered from poor generalization issue. This research introduces a Phase-aware Symmetric-Percussive Source Separation (PSPSS) technique as an alternative to the DFT in the EMD-PSD method. PSPSS effectively isolates the harmonic components of EEG signals by eliminating percussive signals, which are treated as noise, thereby enhancing the quality and interpretability of the EEG harmonics. In the classification module, an improved CNN1D-5Res (ICNN1D-5Res) model is introduced where a channel attention block is integrated with the existing CNN1D-5Res model to enhance the accuracy of epilepsy seizure detection. To further improve the accuracy, a novel optimization approach called Adaptive Gradient Descent with Exponential Growth (AGDEG) is used for training the ICNN1D-5Res model. This algorithm offers stable and fast convergence while working with EEG data. The ICNN1D-5Res model is proven to be a strong and effective method for detecting epileptic seizures, as shown by comparative analysis utilizing the CHB-MIT, Bonn EEG, and TUEP datasets. The model achieves an accuracy of 94.51%, 94.62%, and 93.79%, respectively, outperforming existing methods.