Investigating machine learning-based detection of depression and anxiety in young people using voice biomarkers

Abstract

This thesis explores voice biomarkers for mental health disorders through a three-year collaboration between Ulster University and Canary Speech LLC, focusing on two research questions: 1) Can voice biomarker detection models accurately classify individuals with or without a mental health disorder based on depression and anxiety measures? 2) What demographic and mental health factors affect model performance?

The research investigates the potential of voice biomarkers as an accessible, app-based tool for mental health monitoring, aiming to support remote assessment via smartphones. The thesis begins with chapter one providing the background context of traditional mental health assessments and the challenges of under-provisioned mental health services. It then introduces voice biomarker technology and details the proposed ways it can help meet gaps in mental health services through accessible, scalable mental health assessments put directly in the hands of users. Following this, the literature reviews in chapters two and three evaluate existing detection models for depression and anxiety, respectively. The reviews contribute to knowledge by assessing the performance of these models and highlighting methodological issues in previous studies. A key finding is the challenge of comparing models due to inconsistent performance metrics, particularly in anxiety detection.

Chapters four and five detail a 12-week study involving 381 young people, employing both voice biomarker assessments and traditional mental health scales. Chapter six presents exploratory data analysis, including tests for patterns, correlations, and statistical assumptions. The exploratory data analysis ensured dataset integrity by identifying non-normal distribution patterns, allowing appropriate statistical tests to be employed. It also confirmed stable participant scale scores, providing a robust foundation for evaluating detection model performance in the following chapter. Chapter seven provides the overall results analysis, showing that detection models perform better for individuals with formal diagnoses compared to those who self-identify as having poor mental health. However, model accuracy decreases with the severity of mental health struggles.

Finally, Chapter eight discusses the findings of this thesis, the conclusions drawn from them and proposes directions for future research. The findings reveal that while voice biomarkers show promise for identifying mental health disorders, substantial improvements are needed for practical application. Depression and anxiety detection models currently underperform compared to traditional assessment methods, such as the PHQ-9 for depression and the GAD-7 for anxiety. The reviews in chapters 2 and 3 highlight key limitations such as a lack of standardised performance metrics compounded by methodological inconsistencies. Both our reviews of the wider field and the 12-week study data show it is necessary to improve accuracy and reporting for practical use. Future research should explore the impact of demographic and mental health factors, particularly why increased disorder severity hampers detection performance. Addressing dataset diversity by including participants with severe mental health conditions and investigating the effects of comorbidity may enhance model reliability and offer new opportunities for refinement.

Overall, this thesis provides a comprehensive evaluation of voice biomarker technology's potential and limitations for mental health monitoring, contributing valuable insights to this emerging field.

Date of Award	May 2025
Original language	English
Supervisor	Edel Ennis (Supervisor), Maurice Mulvenna (Supervisor) & Niamh Kennedy (Supervisor)