CEMTNet: a cognitive emotion modulated network for multimodal depression detection.

Area of Science:

• Computational linguistics
• Affective computing
• Machine learning for healthcare

Background:

• Early depression detection is crucial for public health.
• Existing multimodal methods struggle with emotional inconsistency and dynamic modeling.
• Automatic depression identification requires robust and accurate techniques.

Purpose of the Study:

• To propose a novel Cognitive-Emotion-Modulated Trend Network (CEMTNet) for robust multimodal depression recognition.
• To address challenges in cross-modal emotional inconsistency, subtle cue extraction, and dynamic emotional modeling.
• To enhance the accuracy and reliability of automatic depression identification systems.

Main Methods:

• Developed CEMTNet integrating Cognitive Consistency Inference Mechanism (CCIM), Emotion-Modulated Structured Attention (EMSA), and Contrastive Emotion Trend Modeling (CE-TM).
• CCIM dynamically adjusts fusion weights based on emotional consistency.
• EMSA uses emotion-aware attention for fine-grained cue extraction; CE-TM combines contrastive learning with temporal trend modeling.

Main Results:

• CEMTNet achieved superior performance on DAIC-WOZ and EATD-Corpus datasets.
• Achieved high accuracy (0.92), F1-score (0.92), and recall (0.93).
• Ablation studies and case analyses validated the effectiveness and robustness of CEMTNet's modules.

Conclusions:

• CEMTNet demonstrates significant improvements in multimodal depression recognition.
• The proposed network effectively handles emotional inconsistencies and dynamic emotional patterns.
• CEMTNet offers a promising advancement for early depression detection in public health.