IMPROVED DEEP LEARNING ARCHITECTURE WITH BATCH NORMALIZATION FOR EEG SIGNAL PROCESSING

Abstract

Deep learning is commonly used to solve problems such as biomedical problems and many other problems. The most common architecture used to solve those problems is Convolutional Neural Network (CNN) architecture. However, CNN may be prone to overfitting, and the convergence may be slow. One of the methods to overcome the overfitting is batch normalization (BN). BN is commonly used after the convolutional layer. In this study, we proposed a further usage of BN in CNN architecture. BN is not only used after the convolutional layer but also used after the fully connected layer. The proposed architecture is tested to detect types of seizures based on EEG signals. The data used are several sessions of recording signals from many patients. Each recording session produces a recorded EEG signal. EEG signal in each session is first passed through a bandpass filter. Then 26 relevant channels are taken, cut every 2 seconds to be labeled the type of epileptic seizure. The truncated signal is concatenated with the truncated signal from other sessions, divided into two datasets, a large dataset, and a small dataset. Each dataset has four types of seizures. Each dataset is equalized using the undersampling technique. Each dataset is then divided into test and train data to be tested using the proposed architecture. The results show the proposed architecture achieves 46.54% accuracy for the large dataset and 93.33% accuracy for the small dataset. In future studies, the batch normalization parameter will be further investigated to reduce overfitting.