Multimodal Multitask Learning for Predicting Depression Severity and Suicide Risk Using Pretrained Audio and Text...

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Deep learning for mental health assessment

Background:

• Depression severity and suicide risk require prompt assessment and treatment.
• Accurate identification of depression severity (DS) and suicide risk (SR) is crucial for effective management.
• Existing machine learning and deep learning research has limitations in simultaneously addressing DS and SR.

Purpose of the Study:

• To evaluate deep learning models integrating multitask learning (MTL), multimodal learning, and transfer learning.
• To enhance the efficacy of joint classification for depression severity and suicide risk.
• To assess the combined performance of audio and text data using pretrained embeddings.

Main Methods:

• A multitask framework using multimodal fusion of pretrained audio and text embeddings was proposed.
• Data included Chinese audio recordings and clinical questionnaire scores from 200 participants.
• Pretrained embeddings were integrated using concatenation and hard parameter sharing, compared with single-task learning (STL) models.

Main Results:

• Single-task learning models achieved high performance for DS (AUC=0.878) and SR (AUC=0.876) prediction.
• Multitask learning models significantly improved SR prediction over DS prediction.
• MTL models achieved the highest DS classification (AUC=0.887) and SR classification (AUC=0.883).

Conclusions:

• The proposed MTL models effectively enhance depression severity and suicide risk classification using specific audio and text embeddings.
• Caution is advised during MTL implementation to mitigate potential negative transfer effects.
• This research offers a promising, objective method for clinical decision support in parallel DS and SR diagnosis.