Advanced Deep Learning Frameworks for Classification of Epileptic Seizure Using Preprocessed EEG Signals

Detection of epileptic seizures has achieved significant performance by using various methods. Some traditional methods like Machine Learning (ML) exhibit satisfactory results but it has drawbacks of handcrafted features. In recent years, raw EEG data have been applied to various deep neural networks which give the advantage of automatic feature selection. In this paper, some advanced deep-learning architectures have been proposed with preprocessed EEG data. Experiments are conducted by the UCI-Bonn Dataset, employing five-class classifications. Raw EEG data has been transformed to FFT (Fast Fourier Transform) and it is input to three different models (1) Convolution Neural Networks fed with attention layer, (2) A cascaded of CNN (Convolutional Neural Network), attention and LSTM (Long Short Term Memory) layers are employed. Subsequently, the output of the attention layer has concatenated with fully connected layers, and (3) Three branches of CNN (Convolutional Neural Network) incorporating attention mechanisms are utilized, followed by the concatenation of their outputs. Later, layer normalization and dropout mechanisms have been applied at each layer. Finally, the results of these three different architectures have been compared and analyzed with previous related works.