MD-DCNN: Multi-Scale Dilation-Based Deep Convolution Neural Network for epilepsy detection using electroencephalogram signals

Abstract

Approximately 65 million individuals experience epilepsy globally. Surgery or medication cannot cure more than 30% of epilepsy patients.However, through therapeutic intervention, anticipating a seizure can help us avoid it. According to previous studies, aberrant activity inside the brain begins a few minutes before the onset of a seizure, known as a pre-ictal state. Many researchers have attempted to anticipate the pre-ictal condition of a seizure; however, achieving high sensitivity and specificity remains challenging. Therefore, deep learning-based early diagnostic tools for epilepsy therapies using electroencephalogram (EEG) signals are urgently needed. Traditional methods perform well in binary epilepsy scenarios, such as normal vs. ictal, but poorly in ternary situations, such as ictal vs. normal vs. inter-ictal. This study proposes a multi-scale dilated convolution-based network (MD-DCNN) to predict seizures or epilepsy. Traditional DCNNs for epilepsy classification overfit due to insufficient training data (fewer subjects). Windowing 2-sec EEG recordings and extracting the frequency sub-band from each window prevents overfitting in deep networks, which lack training data. We convert each segmented window and its sub-bands into scalogram images and input them into MD-DCNN. The proposed MD-DCNN combines data from several scales without narrowing the acquisition domain. Integrating detailed information into high-level semantic features improves network interpretation and classification. The proposed MD-DCNN is evaluated for two-class, three-class, and cross-database strategy problems using three publicly accessible databases. Experiments show that the MD-DCNN statistically performs better than 13 other current approaches. This demonstrates its potential for developing equipment capable of measuring, monitoring, and recording EEG signals to diagnose epilepsy.