Depressive and mania mood state detection through voice as a biomarker using machine learning.

INTRODUCTION

Depressive and manic states contribute significantly to the global social burden, but objective detection tools are still lacking. This study investigates the feasibility of utilizing voice as a biomarker to detect these mood states.

Methods

From real-world emotional journal voice recordings, 22 features were retrieved in this study, 21 of which showed significant differences among mood states. Additionally, we applied leave-one-subject-out strategy to train and validate four classification models: Chinese-speech-pretrain-GRU, Gate Recurrent Unit (GRU), Bi-directional Long Short-Term Memory (BiLSTM), and Linear Discriminant Analysis (LDA).

RESULTS

Our results indicated that the Chinese-speech-pretrain-GRU model performed the best, achieving sensitivities of 77.5% and 54.8% and specificities of 86.1% and 90.3% for detecting depressive and manic states, respectively, with an overall accuracy of 80.2%.

DISCUSSION: These findings show that machine learning can reliably differentiate between depressive and manic mood states via voice analysis, allowing for a more objective and precise approach to mood disorder assessment.