TF-MCL: Time-frequency fusion and multi-domain cross-loss for self-supervised depression detection.

Area of Science:

• Neuroscience and Computational Psychiatry
• Machine Learning in Healthcare

Background:

• Supervised learning methods for major depressive disorder (MDD) detection using electroencephalogram (EEG) signals face challenges due to the difficulty of labeling data.
• Existing self-supervised contrastive learning methods lack specific design for EEG time-frequency distributions and adequate low-semantic representation acquisition for MDD detection.

Purpose of the Study:

• To propose a novel time-frequency fusion and multi-domain cross-loss (TF-MCL) model to overcome the limitations of existing contrastive learning methods for MDD detection.
• To enhance the characterization of time-frequency distributions in EEG signals for more accurate MDD detection.

Main Methods:

• Developed the TF-MCL model incorporating a fusion mapping head (FMH) to generate time-frequency hybrid representations by remapping information to a fusion domain.
• Optimized a multi-domain cross-loss function to reconstruct representation distributions across time-frequency and fusion domains, improving fusion representation acquisition.
• Evaluated the model on the MODMA and PRED+CT public datasets.

Main Results:

• The TF-MCL model demonstrated significant improvements in accuracy on both MODMA and PRED+CT datasets.
• Achieved superior performance compared to existing state-of-the-art (SOTA) methods, with accuracy increases of 5.87% and 9.96%, respectively.
• The model effectively enhanced the synthesis of time-frequency information and the acquisition of fusion representations.

Conclusions:

• The proposed TF-MCL model offers a promising approach for improving MDD detection using EEG signals by effectively integrating time-frequency information.
• The method addresses key limitations in contrastive learning for neurophysiological signal analysis, paving the way for more robust diagnostic tools.