Adolescent Depression Detection Model Based on Multimodal Data of Interview Audio and Text

Area of Science:

• Psychiatry
• Computer Science
• Machine Learning

Background:

• Depression is a prevalent mental health disorder often emerging in youth.
• Early detection and intervention are crucial for preventing youth suicide.
• Current automated depression detection methods primarily use unimodal data, limiting accuracy.

Purpose of the Study:

• To develop an advanced multimodal machine learning model for accurate adolescent depression detection.
• To address limitations in existing models that consider only global or local features during fusion.
• To improve the accuracy of automated depression detection systems.

Main Methods:

• Extracted four types of features (audio and text, global and local).
• Constructed coarse-grained and fine-grained fusion models.
• Integrated both fusion models for a comprehensive detection approach.

Main Results:

• The proposed multimodal model demonstrated improved accuracy in detecting adolescent depression.
• Feature fusion incorporating both global and local information enhanced model performance.
• The method shows promise for reliable, automated depression screening.

Conclusions:

• Multimodal machine learning offers a powerful approach for adolescent depression detection.
• Advanced feature fusion strategies are key to improving diagnostic accuracy.
• This model can support early intervention efforts and potentially reduce youth suicide rates.