DisNet : Learning interpretable depression representations in speech.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Speech-based depression detection (SDD) is a promising tool for mental health screening.
• Current deep learning models for SDD lack interpretability, hindering clinical application.
• Objective and quantitative depression assessment remains a challenge.

Purpose of the Study:

• To develop an interpretable deep learning framework for speech-based depression detection.
• To enhance the accuracy and provide insights into depression-related speech features.
• To offer an alternative to traditional depression assessment scales.

Main Methods:

• Proposing an interpretable Depression screening Network (DisNet) with a Learnable frequency-domain FilterBank (LFB) and Hierarchical speech Representation Extraction (HRE) modules.
• Implementing a self-supervised learning strategy (SLRD) to improve feature interpretability.
• Utilizing cross-sectional and longitudinal analysis on the AMHS-corpus and evaluating on public datasets (DAIC-woz, CMDC, EATD-corpus).

Main Results:

• DisNet achieved average F1 score improvements of 15.9% (DAIC-woz), 2.8% (CMDC), and 13% (EATD-corpus).
• The model identified specific pronunciation variations in phonemes /i/, /a/, /e/, and /u/ within the depression group.
• DisNet demonstrated end-to-end interpretability, offering a visualization method for depression assessment.

Conclusions:

• DisNet provides an interpretable and accurate approach to speech-based depression detection.
• The model can identify specific acoustic correlates of depression, aiding in quantitative assessment.
• Interpretable AI holds significant potential for advancing mental health diagnostics.