Audio multi-feature fusion detection for depression based on graph convolutional networks.

Area of Science:

• Psychiatry and Mental Health
• Computer Science and Engineering
• Speech Processing and Signal Analysis

Background:

• Depression is a widespread mental health condition requiring early detection for effective management.
• Speech-based depression detection offers a convenient, computer-aided diagnostic approach.
• Current methods face challenges in reliable feature extraction and classification of speech patterns.

Purpose of the Study:

• To introduce a novel audio feature set (SJTU-LWDLab DACD) for depression analysis.
• To propose a new method using summed graph convolutional networks for enhanced depression identification from speech.
• To address the loss of spatial features during audio feature fusion in depression detection.

Main Methods:

• Development of the SJTU-LWDLab DACD audio feature set.
• Application of summed graph convolutional networks for speech pattern classification.
• Mitigation of spatial feature loss through structured fusion of audio features.

Main Results:

• The proposed method achieved a high accuracy of 92.4% in recognizing depression from speech.
• The novel approach demonstrated effectiveness in distinguishing individuals with depression from healthy controls.
• The study provided objective indicators for auxiliary depression identification.

Conclusions:

• The novel speech-based method offers a promising tool for the auxiliary identification of depression.
• The SJTU-LWDLab DACD feature set and summed graph convolutional networks contribute to improved depression detection accuracy.
• This research lays a foundation for more objective and accessible mental health diagnostics.