Electroencephalogram for epileptic seizure detection using stacked bidirectional LSTM_GAP neural network

• Published in: Evolutionary intelligence Vol. 14; no. 2; pp. 823 - 833
• Main Authors: Thara, D. K., Premasudha, B. G., Nayak, Ramesh Sunder, Murthy, T. V., Ananth Prabhu, G., Hanoon, Naeem
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.06.2021; Springer Nature B.V
• Subjects: Accuracy; Applications of Mathematics; Artificial Intelligence; Bioinformatics; Control; Convulsions & seizures; Electroencephalography; Engineering; Epilepsy; Mathematical and Computational Engineering; Mechatronics; Model accuracy; Muscles; Neural networks; Noise; Performance enhancement; Performance measurement; Robotics; Seizures; Special Issue; Statistical Physics and Dynamical Systems; Deep learning; LSTM; EEG; Epilepsy; GAP; Machine learning; Artefacts; Seizure
• ISSN: 1864-5909; 1864-5917
• DOI: 10.1007/s12065-020-00459-9

Epilepsy seizure classification has been an ongoing research from decades. Auto seizure detection methods are developed by many researchers. Year by year new methods and new approaches are introduced by researchers to improve the performance of these automatic detection methods. Electroencephalogram (EEG) is the most widely used tool for detecting seizure as it has high resolution which makes it more feasible for analysis. In this paper we are introducing an approach based on stacked bidirectional long short term memory with global average pooling (LSTM_GAP) neural network for detecting epileptic seizure events. We use open access EEG dataset available in Bonn University Germany for the experiment. The model is evaluated based on three performance metrics, specificity, sensitivity and accuracy. It resulted with outstanding performance with 100% accuracy, sensitivity and specificity in detecting epileptic seizure events which is the highest accuracy among all the techniques available in state-of-the art literature. Another advantage of LSTM_GAP model lies in its robustness in noise. In real, the EEG recorded in hospital laboratories consists of artefacts generated from eye blink and muscle activities of the patients. When our model will be put to practical use by the neurologists, the EEG recordings consisting of artefacts will be given as input to the model for detecting seizure events. Hence we need to develop a model which is robust against noise and which can provide good results even when the EEG data consisting artefacts is given as input. Hence we extended our experiment to check the performance of the LSTM_GAP model by adding eye blink artefacts and muscle artefacts to the input EEG data. The model was able to give superior results even after adding artefacts. We were able to achieve 97.67% sensitivity, 98.83% specificity and 97.65% accuracy. Thus LSTM_GAP model gives superior results for the EEG with and without noise compared to other methods in the literature.