Automated classification of EEG into meditation and non-meditation epochs using common spatial pattern, linear discriminant analysis, and LSTM

EEG-based investigation of effects of mindfulness meditation training on state and trait by deep learning and traditional machine learning

“…The inter-subject classification accuracy using traditional machine learning technique FBCSP + SVM in our study is 68.50%, which is comparable to 74.0% reported in Panachakel et al (2021b) using also traditional machine learning method, but as discussed above our task is much more difficult. The inter-subject classification accuracy for Raja yoga experts is greatly improved from 74.0% in Panachakel et al (2021b) to 79.1, 86.5, 91.0, and 94.1% in Panachakel et al (2021a) for using alpha, beta, low gamma, high gamma features, respectively, followed by the CSP + LDA + LSTM deep learning framework. However, the meditation state classification accuracy for our MBSR novices using deep and shallow ConvNets does not improve over the traditional FBCSP + SVM method but instead drops to around chance levels of 48.34 and 50.60%, respectively.…”

“…However, the meditation state classification accuracy for our MBSR novices using deep and shallow ConvNets does not improve over the traditional FBCSP + SVM method but instead drops to around chance levels of 48.34 and 50.60%, respectively. Though in this study we use a CNN architecture ConvNet as the deep learning architecture, which is different from the RNN architecture LSTM in Panachakel et al (2021a) , the main reason accounting for the failure of ConvNet in the inter-subject classification scenario should be that ConvNet uses an end-to-end training architecture whereas in Panachakel et al (2021a) the LSTM architecture does not work directly on raw EEG data but instead on EEG features extracted by CSP + LDA.…”

“…To overcome this disadvantage of end-to-end ConvNets on inter-subject classification, we may either enlarge the dataset by collecting data from a large number of subjects for better covering broad individual features so as to reduce the problem of individual difference, or try not to use end-to-end learning for the case where only a small number of subjects are available, as demonstrated in Panachakel et al (2021a) .…”

“…The classification techniques used for meditation state recognition include linear discriminant analysis (LDA) ( Panachakel et al, 2021b ), support vector machine (SVM) ( Shaw and Routray, 2016 ; Han et al, 2020 ), random forest (RF) ( Huang et al, 2021 ), and artificial neural network (ANN). Recently, deep learning techniques including long short-term memory (LSTM) framework ( Panachakel et al, 2021a ) as well as various deep convolutional neural networks (CNN) such as VGG16, ResNet50, and MobileNet ( Pandey and Miyapuram, 2021 ), have also been exploited for meditation state recognition.…”

“…There are only a small number of publications reporting meditation state classification, mostly on yoga meditation ( Han et al, 2020 ), including Raja yoga ( Panachakel et al, 2021a , b ), Kriya yoga ( Shaw and Routray, 2016 ), and Himalayan yoga ( Pandey and Miyapuram, 2021 ), with only one on non-specified meditation ( Goshvarpour and Goshvarpour, 2012 ).…”