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

Area of Science:

• Psychiatry
• Computational Linguistics
• Biomedical Engineering

Background:

• Depressive and manic states represent a significant global health burden.
• Objective diagnostic tools for mood disorders are currently lacking.
• Voice analysis offers a potential non-invasive biomarker for mood state detection.

Purpose of the Study:

• To investigate the feasibility of using voice recordings as a biomarker for detecting depressive and manic states.
• To evaluate the performance of machine learning models in classifying mood states based on voice features.

Main Methods:

• Extracted 22 voice features from real-world emotional journal recordings.
• Applied a leave-one-subject-out cross-validation strategy.
• Trained and validated four classification models: Chinese-speech-pretrain-GRU, Gate Recurrent Unit (GRU), Bi-directional Long Short-Term Memory (BiLSTM), and Linear Discriminant Analysis (LDA).

Main Results:

• The Chinese-speech-pretrain-GRU model achieved the highest performance.
• Achieved 77.5% sensitivity and 86.1% specificity for depressive states.
• Achieved 54.8% sensitivity and 90.3% specificity for manic states, with an overall accuracy of 80.2%.

Conclusions:

• Machine learning models can reliably differentiate between depressive and manic mood states using voice analysis.
• Voice analysis provides a potential objective and precise method for assessing mood disorders.
• This approach may enhance the clinical evaluation and management of mood disorders.