Performance Measurements of Deep Learning-Based EEG Signal Classification Using Multi-class Seizure Type with CNN and RNN – LSTM Algorithms

Abstract

The objective of this research is to characterize and categorize EEG patterns to distinguish between epileptic and healthy brain functions. By applying deep learning. To apply CNN and RNN-LSTM models to multi-classify different kinds of seizures. The EEG signal is converted to images to feed into the CNN. This study mainly examined two approaches. First, we examine transfer learning with ten pretrained networks, namely AlexNet, VGG-16, VGG-19, SqueezeNet, GoogLeNet, Inceptionv3, Densenet201, ResNet-18, ResNet-50, and ResNet-101. This research seeks to find out the best network for the job. Images are classified using an SVM after passing through the pretrained model for feature extraction. The study finds that Densenet201 with RNN-LSTM is the best model with an accuracy of 97% for classifying seizure types. CNNs capture spatial aspects in images by means of local receptive fields, hierarchical features, pooling, and non-linear activations. The extracted features comprise mean, skewness, variance, and power spectral features, and so on. Thus, the feature vector consists of 20588 features. RNNs and their variants, namely LSTM and Gated Recurrent Unit (GRU), are good at processing sequential data. The RNN feature extraction is stored as a 1,920-dimensional vector. Combining CNNs and RNNs in CRNNs enables effective extraction and analysis of key data features. The combined implementation of the LSTM Network and the CNN Network will be to classify the variants of the seizures effectively.