Depression level prediction via textual and acoustic analysis.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Automatic depression diagnosis often relies on video data, posing privacy challenges.
• Voice data offer a less intrusive alternative, analyzing tone, emotion, and self-focus.
• Existing multimodal approaches often neglect temporal alignment between speech and text.

Purpose of the Study:

• To introduce timestamp-integrated multimodal encoding for depression (TIMEX-D).
• To synchronize acoustic speech features with text data for improved depression prediction.
• To address limitations in analyzing speech-text interactions for mental health assessment.

Main Methods:

• Developed TIMEX-D with timestamp extraction, multimodal encoding, and depression analysis blocks.
• Extended transformer positional encoding to mimic human speech recognition.
• Reduced model complexity compared to existing transformer models.

Main Results:

• TIMEX-D achieved high accuracies: 99.17% on DAIC-WOZ and 99.81% on EDAIC.
• Outperformed previous methods by approximately 13%.
• Demonstrated effective prediction of depression levels through synchronized multimodal analysis.

Conclusions:

• TIMEX-D offers a highly accurate and efficient method for depression level prediction.
• Synchronizing speech and text data is crucial for understanding their interplay in mental health.
• This approach can significantly enhance mental health diagnostics and monitoring.