Detection of Depression Using Weighted Spectral Graph Clustering With EEG Biomarkers

The alarming annual growth in the number of people affected by Major Depressive Disorder (MDD) is a problem on a global scale. In the primary scrutiny of depression, Electroencephalography (EEG) is one of the analytical tools available. Machine Learning (ML) and Deep Neural Networks (DNN) methods are the most common techniques for MDD diagnosis using EEG. However, these ML methods heavily rely on manually annotated EEG signals, which can only be generated by experts, for training. This also necessitates a large amount of memory and time constraints. The requirement of huge amounts of data to foresee emerging tendencies or undiscovered alignments is enforced. This article develops an unsupervised learning method for identifying MDD in light of these difficulties. The preprocessed EEG is used to extract three quantitative biomarkers (Band Power: Beta, Delta, and Theta), and three signal features (Detrended Fluctuation Analysis (DFA), Higuchi’s Fractal Dimension (HFD), and Lempel-Ziv Complexity (LZC)). Through the extracted features, an undirected graph is created using the features as a weight along the edges, with nodes as channels in EEG recording. The bifurcation of the subjects in either of the classes (MDD or N) is done by implementing spectral clustering. A 98% accuracy with a 2.5% of miss-classification error is achieved for the left hemisphere. In contrast, a 97% accuracy with a 3.3%
$CEP$
(or miss-classification error or Classification Error Percentage) is achieved for the right hemisphere. FP1 and F8 channels have achieved the highest possible level of classification accuracy.