Discovery of Shared Latent Nonlinear Effective Connectivity for EEG-Based Depression Detection

Area of Science:

• Neuroscience
• Machine Learning
• Computational Psychiatry

Background:

• Granger causality (GC) effective connectivity (EC) from electroencephalogram (EEG) is used for mental disorder detection.
• Existing methods often overlook nonlinear dynamics shared across subjects within the same class.

Purpose of the Study:

• To propose a novel model combining graph neural networks (GNNs) and variational autoencoders (VAEs) for detecting depression.
• To construct shared latent nonlinear EC from raw EEG signals, capturing inter-subject dynamics.

Main Methods:

• Utilized GNNs with convolution modules and fully connected layers for graph encoding of EEG channel connectivity.
• Incorporated a class-specific Gaussian mixture model (GMM) within VAEs for modeling shared nonlinear dynamics.
• Employed a node-to-edge encoding and edge-to-node decoding process to learn shared latent nonlinear EC.

Main Results:

• The proposed method successfully learned generalized nonlinear EC representations from EEG signals.
• Discovery of shared latent dynamics significantly improved depression identification accuracy.
• Validated performance on multiple open-accessed datasets.

Conclusions:

• The GNN-VAE model effectively captures complex, shared nonlinear dynamics in EEG for depression detection.
• This approach offers a more robust and generalized method for analyzing effective connectivity in mental health research.
• The findings highlight the potential of leveraging inter-subject dynamics for improved diagnostic tools.