Convolutional neural networks for classifying healthy individuals practicing or not practicing meditation according to the EEG data.
Vavilovskii zhurnal genetiki i selektsii – December 01, 2023
Source: PubMed
Summary
A novel convolutional neural network achieved 82% accuracy in classifying individuals as meditators or non-meditators based on EEG data. Analyzing event-related brain potentials during a stop-signal paradigm, the study involved 100 participants (51 meditators and 49 non-meditators) and later validated findings with an additional 25 individuals. Meditation practices significantly enhance self-control over mental states, potentially reducing anxiety and stress levels. The advanced model shows promise for objectively assessing stress and predicting susceptibility to anxiety and depression disorders across diverse populations.
Abstract
The development of objective methods for assessing stress levels is an important task of applied neuroscience. Analysis of EEG recorded as part of a behavioral self-control program can serve as the basis for the development of test methods that allow classifying people by stress level. It is well known that participation in meditation practices leads to the development of skills of voluntary self-control over the individual's mental state due to an increased concentration of attention to themselves. As a consequence of meditation practices, participants can reduce overall anxiety and stress levels. The aim of our study was to develop, train and test a convolutional neural network capable of classifying individuals into groups of practitioners and non-practitioners of meditation by analysis of eventrelated brain potentials recorded during stop-signal paradigm. Four non-deep convolutional network architectures were developed, trained and tested on samples of 100 people (51 meditators and 49 non-meditators). Subsequently, all structures were additionally tested on an independent sample of 25 people. It was found that a structure using a one-dimensional convolutional layer combining the layer and a two-layer fully connected network showed the best performance in simulation tests. However, this model was often subject to overfitting due to the limitation of the display size of the data set. The phenomenon of overfitting was mitigated by changing the structure and scale of the model, initialization network parameters, regularization, random deactivation (dropout) and hyperparameters of cross-validation screening. The resulting model showed 82 % accuracy in classifying people into subgroups. The use of such models can be expected to be effective in assessing stress levels and inclination to anxiety and depression disorders in other groups of subjects.