Depression detection using deep learning and large language models from multimodalities.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Psychiatry

Background:

• Depression assessment is challenging due to its complex impact on neural functioning, cognition, emotion, and behavior.
• Traditional diagnostic methods rely on subjective interpretation, necessitating objective, data-driven approaches.
• Deep learning offers automated detection using physiological and behavioral data.

Purpose of the Study:

• To review recent advancements in AI-driven depression detection using electroencephalography (EEG) and multimodal data.
• To evaluate the impact of modality diversity and architectural sophistication on model performance.
• To explore the potential of Large Language Models (LLMs) in depression assessment.

Main Methods:

• Analysis of unimodal (EEG) and multimodal (speech, facial expression, EEG) deep learning architectures.
• Review of transformer-based fusion and attention mechanisms for cross-modal feature integration.
• Examination of Large Language Models (LLMs) for cross-modal alignment and few-shot learning.

Main Results:

• Multimodal systems integrating speech, facial expression, and EEG outperform unimodal approaches.
• Transformer and attention models achieve high accuracy (90-95%) on lab datasets and good F1-scores (0.80-0.90) on real-world datasets.
• LLMs offer data-efficient adaptation and improved cross-modal reasoning for depression detection.

Conclusions:

• Combining diverse modalities and sophisticated AI architectures significantly enhances depression detection.
• LLMs represent a methodological shift, enabling scalable, explainable, and clinically relevant AI systems.
• Further research is needed to address limitations in dataset diversity, standardization, interpretability, and clinical validation.