Article: Depression diagnosis using a hybrid residual neural network Journal: International Journal of Biomedical Engineering and Technology (IJBET) 2023 Vol.42 No.3 pp.244 - 261 Abstract: Depression is one of the most widespread psychiatric disorders. EEG signals can be utilised as a tool to diagnose depression objectively. This paper employs a hybrid method to classify healthy and depressed signals, which uses a pre-trained ResNet101 to extract features automatically. Thereby, the problem of designing and training deep networks for automatic feature extraction is solved. The hypothesis in the present study is that feature-extraction layers in ResNet101 also perform desirably in detecting depressed signals. In hybrid structures, SVM, KNN, and DT classifiers are used for final classification purposes. ResNet101-SVM, ResNet101-KNN, and ResNet101-DT structures have reached accuracy of 93.8%, 90.1%, and 82.1%, respectively. Moreover, for the ResNet101-SVM structure, which has shown the best performance among all structures, the accuracy, sensitivity, and specificity are 94.7%, 94.0%, and 95.2% after applying the ten-fold cross-validation method. The results indicate the proper performance of all structures, especially the ResNet101-SVM structure, in diagnosing depression. Inderscience Publishers - linking academia, business and industry through research

Title: Depression diagnosis using a hybrid residual neural network

Authors: Mahsa Ofoghi Rezaei; Somayeh Makouei; Sebelan Danishvar

Addresses: Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, 51664, Iran ' Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, 51664, Iran ' Department of Computer Science, College of Engineering, Design and Physical Sciences, Brunel University, UK

Abstract: Depression is one of the most widespread psychiatric disorders. EEG signals can be utilised as a tool to diagnose depression objectively. This paper employs a hybrid method to classify healthy and depressed signals, which uses a pre-trained ResNet101 to extract features automatically. Thereby, the problem of designing and training deep networks for automatic feature extraction is solved. The hypothesis in the present study is that feature-extraction layers in ResNet101 also perform desirably in detecting depressed signals. In hybrid structures, SVM, KNN, and DT classifiers are used for final classification purposes. ResNet101-SVM, ResNet101-KNN, and ResNet101-DT structures have reached accuracy of 93.8%, 90.1%, and 82.1%, respectively. Moreover, for the ResNet101-SVM structure, which has shown the best performance among all structures, the accuracy, sensitivity, and specificity are 94.7%, 94.0%, and 95.2% after applying the ten-fold cross-validation method. The results indicate the proper performance of all structures, especially the ResNet101-SVM structure, in diagnosing depression.

Keywords: depression; diagnosis; classification; electroencephalogram; EEG; deep learning; residual network; hybrid model; support vector machine; SVM; K-nearest neighbour; KNN; decision tree; DT.

DOI: 10.1504/IJBET.2023.132545

International Journal of Biomedical Engineering and Technology, 2023 Vol.42 No.3, pp.244 - 261

Received: 28 Jan 2021
Accepted: 19 Sep 2021
Published online: 28 Jul 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Depression diagnosis using a hybrid residual neural network

Keep up-to-date