Optimization of model hyperparameters.

Depression is a prevalent mental health condition that significantly impacts individuals’ daily lives, work productivity, relationships, and overall well-being. The lack of reliable biomarkers complicates screening, contributing to underdiagnosis. Depression’s impact on voice and acoustic parameters enables differentiation between adaptive and non-adaptive mood profiles, offering potential classifiers for screening. This study evaluates the capability of seven distinct machine learning models to identify depression in speech samples. WhatsApp™ audio messages (WA), clinical, and sociodemographic data were collected from 160 individuals divided into two groups: one for algorithm development and the other for testing. Each group included patients with Major Depressive Disorder and healthy controls. In the test group, participants were interviewed using the Mini-International Neuropsychiatric Interview (MINI), and their WhatsApp™ audio recordings included both structured and semi-structured formats. After pre-processing the audio, 68 acoustic features were used to train the machine learning models. Results shows that: i) The algorithms evaluated WhatsApp™ audio recordings from the test group, achieving peak accuracies of 91.67% for women and 80% for men, with an AUC of 91.9% for women and 78.33% for men. ii) The accuracy of Machine Learning (ML) classification varies depending on the type of audio instruction provided. ML can classify, with reasonable accuracy, whether a WhatsApp™ audio message represents a depressive patient or a healthy individual. Future studies should further explore the relationship between voice characteristics, different mood profiles, and emotional states.