Classification of Meditation Expertise from EEG Signals Using Shallow Neural Networks

Abstract

Recently machine-learning and deep-learning approaches have been widely adopted for automatic inter-group and inter-state classification of the brain states. Traditional machine learning-based classification requires a complicated pipeline of signal processing, artifact removal, feature extraction, and selection, which is time-consuming and requires human intervention. Whereas deep learning-based classifiers attempt to overcome this manual overhead by allowing end-to-end processing however at the cost of training time and model complexity. In this work, we have explored the use of different Shallow Convolutional Neural Networks (SCNN) based classifiers for inter-group (expert and non-expert) classification of EEG signals during Himalayan Yoga meditation. Several experiments were carried out to record the effects on the classification performance under varied conditions. We experimented with input representation of the signal, window size for signal segmentation, and various model parameters. It was observed that a very simple time domain representation of the raw signal, segmented by a window of 5s, when combined with a shallow 1D-CNN showed the best performance, with an accuracy score of 99.48%. The performance of all the shallow networks was found to be at par with the pre-trained state-of-the-art deep-CNN model, VGG-16, fine-tuned on the same data. Such lightweight models can be particularly useful for the on-the-go wearable and personal EEG devices which are latency sensitive, and highly susceptible to artifactual noise.