Interpretable Probabilistic Identification of Depression in Speech.

Area of Science:

• Artificial Intelligence in Healthcare
• Computational Psychiatry
• Speech Signal Processing

Background:

• Current mental health assessments rely on time-consuming consultations, struggling to meet rising demand.
• National health systems face challenges in providing timely mental well-being evaluations.
• There's a need for efficient, accurate tools to aid in mental illness diagnosis.

Purpose of the Study:

• To develop an AI-based tool for accelerated and interpretable mental health assessment using speech signals.
• To enhance the diagnostic process by providing transparent AI predictions.
• To bridge the usability gap between AI systems and medical personnel.

Main Methods:

• Utilized an explainability-by-design approach for AI model development.
• Extracted audio descriptors including Mel-scaled spectrum summarization, Teager operator, and periodicity.
• Employed Hidden Markov Models adapted from an ergodic universal model with a specific data selection strategy.

Main Results:

• Achieved significantly higher performance compared to the current state-of-the-art on a public dataset.
• An ablation study confirmed the relevance and contribution of individual system components.
• Demonstrated superior accuracy in mental health assessment through speech analysis.

Conclusions:

• The proposed AI tool offers a transparent and interpretable solution for mental health assessment.
• The system's high performance and explainability pave the way for practical clinical adoption.
• This technology can significantly assist healthcare professionals in diagnosing mental illnesses more efficiently.