Personalized vs. population-based speech models for multi-dimensional mental health prediction

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Speech signal processing

Background:

• Mental disorders like depression, anxiety, and stress are rising, especially in young adults.
• Traditional mental health assessments are resource-intensive and lack scalability.
• Existing speech-based models struggle to differentiate disorder signals from individual voice traits.

Purpose of the Study:

• To develop a hybrid framework combining population-level and individual-specific adaptation for enhanced personalized mental health prediction.
• To evaluate the framework's performance in predicting depression, anxiety, and stress severity using speech data.
• To compare the hybrid approach against population-only and individual-only models.

Main Methods:

• Utilized the longitudinal YouthDASS dataset with over 1,000 speech samples from individuals aged 18-30.
• Employed a hybrid machine learning framework integrating population-level modeling with incremental individual adaptation.
• Assessed various models, with a 1D Convolutional Neural Network (1D CNN) showing superior performance.

Main Results:

• The hybrid framework significantly outperformed population-level models in predicting depression, anxiety, and stress.
• Achieved lower Root Mean Square Error (RMSE) values: 6.95 for depression, 7.15 for anxiety, and 4.95 for stress.
• Individual-only models showed variable performance across different mental health conditions.

Conclusions:

• Integrating population-level insights with individual adaptation offers a superior balance of generalization and personalization.
• The proposed framework facilitates scalable, personalized speech-based mental health monitoring.
• Adaptive machine learning holds significant promise for longitudinal mental health assessment.