Automatic feature learning model combining functional connectivity network and graph regularization for depression detection

Abstract

Depression has become a major health and economic burden worldwide. Electroencephalography (EEG) data has been used by a growing number of researchers to study depression. EEG-based functional connectivity (FC) features have emerged since they can account for the relationships between different brain regions. In this paper, the time–frequency analysis technique is introduced into the construction of the FC matrix. Specifically, instead of directly building the FC matrix from the EEG signals, the intrinsic time-scale decomposition (ITD) method is employed to mine the time–frequency information, and then the Pearson correlation is used to measure the FC between channels. The results show the significant differences in the FC networks between different groups. Furthermore, the graph-based adaptive least absolute shrinkage and selection operator model (GA-LASSO) is proposed in this paper to learn the discriminative features from the FC matrix, which is mainly achieved by adding both the adaptive L1 and graph regularized terms to the original least absolute shrinkage and selection operator (LASSO) model. The advantages of GA-LASSO come from the processing of discriminative weights of different features, and the connections between features by graph topology. In addition, the effectiveness of the proposed strategy of depression detection is validated on the open dataset MODMA, as well as the self-collected dataset called EDRA. The experimental results show that the current study sheds new light on the pathological mechanism of subclinical depression and suggests that EEG resting-state FC analysis may identify potentially effective biomarkers for its clinical diagnosis.