Speech-based depression detection using higher-order spectral features and a multi-level transformer.

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Psychiatric diagnostics

Background:

• Limited access to mental health care globally due to psychiatrist shortages and high diagnostic costs.
• Approximately 60% of psychiatric patients lack access to necessary mental health services.
• Need for precise and prompt automated systems for mental health assessment.

Purpose of the Study:

• To develop an automated depression detection system using auditory signals (voice).
• To integrate hierarchical higher-order spectral distribution with deep learning for improved depression identification.
• To address the accessibility and cost barriers in current mental health diagnosis.

Main Methods:

• Utilized the Distress Analysis Interview Corpus, Wizard of Oz (DAIC-WOZ) dataset containing clinical interview audio recordings.
• Extracted statistical, handmade, bispectral, and deep features.
• Employed a Multi-level Convolutional Transformer with attention learning (ML-CoT-AL) model, integrating Convolutional Transformer (CoT), Channel and Element-wise Attention Module (CEAM), and Negotiator Modules (NM).
• Applied the RIME optimization technique for hyperparameter tuning of the ML-CoT-AL model.

Main Results:

• The ML-CoT-AL model, optimized with RIME, demonstrated improved accuracy in multi-level depression identification.
• Integration of advanced deep learning architectures and attention mechanisms facilitated robust feature extraction.
• The system effectively processed auditory signals for psychological distress detection.

Conclusions:

• The proposed voice-based depression detection system offers a promising, accessible, and cost-effective solution.
• Deep learning models, particularly the ML-CoT-AL, show significant potential in analyzing voice for mental health assessment.
• Automated systems can augment traditional diagnostic methods, expanding mental health care access.