This research investigates why advanced pre-trained models perform poorly in depression detection without extensive data augmentation.

• Speech self-supervised models show unexpectedly low performance in depression detection settings
• Research identifies feature entanglement as a potential root cause of poor performance
• Study explores the untapped potential of Large Language Models for multi-modal depression detection
• Findings have significant implications for developing reliable AI-based mental health screening tools

This work addresses a critical gap in medical AI applications, as improved depression detection systems could help address a pressing global mental health challenge through earlier, more accurate diagnosis.

Why Pre-trained Models Fail: Feature Entanglement in Multi-modal Depression Detection