Application of Conv-1D and Bi-LSTM to Classify and Detect Epilepsy in EEG Data

Abstract

EEG is used to study the electrical changes in the brain and can derive a conclusion as epileptic or not, using an automated method for accurate detection of seizures. Deep learning, a technique ahead of machine learning tools, can self-discover related data for the detection and classification of EEG analysis. Our work focuses on deep neural network architecture to visualize the temporal dependencies in EEG signals. Algorithms and models based on Deep Learning techniques like Conv1D, Conv1D + LSTM, and Conv1D + Bi-LSTM for binary and multiclass classification. Convolution Neural Networks can spontaneously extract and learn features independently in the multichannel time-series EEG signals. Long Short-Term Memory (LSTM) network, with its selective memory retaining capability, Fully Connected (FC) layer, and softmax activation, discover hidden sparse features from EEG signals and predicts labels as output. Two independent LSTM networks combine to form Bi-LSTM in opposite directions and appreciate added visibility to upcoming information to provide efficient work contrary to previous methods. Long-term EEG recordings on the Bonn EEG database, Hauz Khas epileptic database, and Epileptic EEG signals from Spandana Hospital, Bangalore, assess performance. Metrics like precision, recall, f1-score, and support exhibit an improvement over traditional ML algorithms evaluated in the literature.