Interpretation of Frequency Channel-Based CNN on Depression Identification

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• High-performance electroencephalogram (EEG) classification is crucial for assessing Major Depressive Disorder (MDD) and monitoring patient status.
• Challenges include significant noise, non-stationarity in brain states, and difficulty decoupling neural network complexity from brain disease dynamics.

Purpose of the Study:

• To develop an accurate and efficient method for identifying MDD using EEG data.
• To improve the classification performance by integrating brain rhythms with an attention mechanism.
• To quantify the complexity of the classification model using information entropy.

Main Methods:

• Design of a Frequency Channel-based convolutional neural network (FCCNN) integrating brain rhythms and an attention mechanism.
• Development of an information entropy calculation method based on affinity propagation (AP) clustering for model complexity analysis.
• Experimental validation on a dataset comparing healthy individuals and MDD patients.

Main Results:

• The FCCNN achieved high accuracy (99±0.08%), sensitivity (99.07±0.05%), and specificity (98.90±0.14%) in identifying MDD.
• Quantitative interpretation revealed significant differences in brain activity between MDD patients and controls across specific brain regions (frontal, left/right temporal lobes).

Conclusions:

• The FCCNN offers a promising, high-performance solution for automated MDD detection via EEG.
• The information entropy method provides valuable insights into classifier complexity and brain region-specific alterations in MDD.
• This approach facilitates objective assessment and monitoring of depression, potentially reducing risks associated with the condition.