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

Mental disorders such as depression, anxiety, and stress are increasingly prevalent, particularly among young adults. Traditional assessment methods rely on self-reports and resource-intensive clinician interviews, limiting scalability and accessibility. Speech-based machine learning models offer a scalable and non-invasive alternative; however, population-level models often struggle to distinguish disorder-related signals from speaker-specific traits, reducing individual prediction accuracy.

We propose a hybrid framework that combines population-level modelling with incremental individual-specific adaptation to improve personalized mental health prediction. The approach was evaluated using our longitudinal YouthDASS dataset, which contains over 1,000 speech samples collected up to two months from individuals aged 18-30 years, labelled with severity scores for depression, anxiety, and stress based on the DASS-21 scale. Multiple machine learning models were explored and assessed under population-only, individual-only, and hybrid modelling settings, among which a one-dimensional convolutional neural network (1D CNN) demonstrated the best performance.

The hybrid approach outperformed population-level models across all three mental health conditions, achieving lower individual-level root mean square error (RMSE) values of 6.95 for depression, 7.15 for anxiety, and 4.95 for stress on the DASS-21 scale. In comparison, individual-only models demonstrated mixed performance across disorders.

These findings suggest that integrating population-level knowledge with individual-specific adaptation provides a stronger balance between generalization and personalization than either approach alone. The proposed framework supports the development of scalable, personalized speech-based mental health monitoring systems and highlights the potential of adaptive machine learning methods for longitudinal mental health assessment.