Deep Learning Approaches for Electroencephalography (EEG)-Based User Response Prediction

Abstract

In deep learning, finding the best algorithms for time series data can be challenging due to its stochastic and nonlinear nature. This study endeavours to address the challenges posed by a 10-class classification and binary classification problem employing deep learning algorithms. We collected Electroencephalography (EEG) data from participants engaged in the Corsi Block Tapping Task, utilizing various combinations of Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM) models across multiple layers to achieve the highest accuracy across different frequency bands - namely, beta band (14-30 Hz), alpha band (8-13 Hz), theta band (4-8 Hz), and delta band (0.5-4.35 Hz). Our findings in the context of the 10-class classification problem highlight the superior performance of the 1 CNN + 4 linear layers model, boasting an accuracy of 64.47%. In the realm of binary classification, the 1 LSTM + 4 linear layers model emerged as the top performer, achieving an impressive accuracy of 93.30%. Notably, the beta wave exhibited enhanced predictive capabilities. These results hold promising implications for the design of brain-computer interface experiments, where specific brain regions can predict responses with heightened accuracy. Furthermore, future applications may encompass the development of cognitive systems where both time and accuracy play pivotal roles.