A review of studies using machine learning to detect voice biomarkers for depression

Voice biomarkers developed using machine learning are a promising potential biomarker for mental disorders, including depression. This paper presents a narrative review with a systematic search of the evidence surrounding the efficacy of voice biomarkers as indicators of depression. The review considers two research questions: (i) What is the efficacy of voice biomarkers as potential biomarkers for depression? (ii) What are the variations in the samples and design methodologies employed? Nineteen papers were identified as examining voice biomarkers for depression using machine learning methods between January 2019 and February 2022. A subset of guidelines recommended in a previous systematic review was selected and adapted to investigate aspects of the field since that review. Seventeen studies used classification methods, and two used regression methods. Within the papers that examined classification, sensitivity (recall) was used by 76% of papers, accuracy by 65%, AUC by 59%, and F1 score by 59%. From these papers, the average performance achieved for the following metrics was 0.78 for sensitivity (recall), 0.76 for F1 score, and 0.78 for AUC. This review found that the efficacy of vocal biomarkers as indicators for depression is below that of the PHQ-9 form, a tool commonly used in psychology. The PHQ-9 can serve as a benchmark against which to compare these models. Difficulties were observed in comparing these models due to the variety of performance metrics used. Recommendations are presented as to how the generalisability of these models may be strengthened, e.g., testing on unseen data after models are developed.