Efficiency of convolutional neural networks of different architecture for the task of depression diagnosis from EEG data

Abstract

The purpose of this paper is to comparatively analyse the efficiency of using artificial neural networks with different convolutional and recurrent architectures in the task of depression diagnosis based on electroencephalogram (EEG) data. Open datasets were chosen as objects of the study and own EEG data of real patients with depression were collected. Methods. To solve the problem of identifying biomarkers of depressive disorder from EEG data, we used convolutional neural networks using two-dimensional or one-dimensional convolution operation, as well as hybrid models of convolutional and recurrent neural networks. To test the developed models of artificial neural networks, we selected open data sets, performed an experiment to collect our own data from real depressed patients, and merged the prepared data sets. The result of this work is analysis and comparison of the performance of different classifiers based on convolutional and recurrent neural network models. Conclusion. We show that the average accuracy of classification of depressive disorder in a test sample using cross-validation was 0.68. The results are consistent with the known results from the literature for small patient-disaggregated datasets. Although the accuracy obtained in this task is insufficient for practical application of the model, it can be argued that further research to improve the efficiency of the model is promising, as well as the need to create a sufficiently large representative dataset of depressed patients, which is an important scientific task for further construction of biophysical models of depressive disorders.

Keywords

• non-invasive electroencephalogram
• depression diagnosis
• deep learning
• convolutional neural networks (cnn)
• recurrent neural networks (rnn)