Double Sparse Dictionary-Based Electroencephalography Channel Selection for Depression Analysis

To address channel redundancy and high computational complexity in high-density electroencephalography (EEG) for depression (DP) analysis, this study proposes an elastic net-based double sparse dictionary channel selection (EN-DSDCS) method, aimed at screening core EEG channels and analyzing abnormal topological changes in the brain functional network (BFN) of DP patients. An improved coarse-graining method is proposed to reconstruct EEG signals, calculate their multi-scale permutation entropy (MSPE), and construct an MSPE matrix that characterizes signal complexity. Based on this, a double sparse dictionary structure is designed, combining a fixed-base dictionary (constructed via the Kronecker product of two DCT(discrete cosine transform) matrices and a learning dictionary optimized through iterative sparse K-SVD. The sparse dictionary D is dervied by multiplying these two components. Subsequently, elastic net regularization jointly optimizes D and the sparse coefficient matrix X, enabling the selection of key channels based on their sparsity levels. The BFN is then constructed based on the phase lag index (PLI) using the selected channels to analyze abnormal changes in network topology and the distribution of Hub node in DP patients. Experimental results demonstrate that EN-DSDCS reduces signal reconstruction error by 3 × 10^-4, decreases channel sparsity by 3.93% compared with Lasso-based double sparse dictionary channel selection (L-DSDCS), and most of the selected channels are distributed in the frontal and temporal lobes. In addition, BFN analysis further reveals differential connectivity patterns in these regions among DP patients, with Hub node distribution exhibiting a left hemispheric bias.