DepressionMIGNN: A Multiple-Instance Learning-Based Depression Detection Model with Graph Neural Networks.

Area of Science:

• Computational psychiatry
• Affective computing
• Machine learning for healthcare

Background:

• Global depression prevalence highlights the need for technological solutions in diagnostics.
• Sensor-based systems can augment limited resources for early depression detection.
• Existing methods often focus on momentary expressions, missing chronic disorder patterns.

Purpose of the Study:

• To propose a novel framework for depression detection using multimodal data (video, audio, text).
• To address depression as a chronic disorder by analyzing temporal behavioral patterns.
• To improve the accuracy of depression detection by considering long-term indicators.

Main Methods:

• Utilized benchmark datasets with multimodal data (video, audio, transcribed text).
• Developed a framework segmenting videos into utterance-level instances using Gated Recurrent Units (GRU) for contextual representation.
• Constructed graphs with utterance embeddings as nodes, connected by dual relationships (chronological and relevant information).
• Employed Graph Neural Networks (GNNs) to learn multi-dimensional edge relationships and align multimodal representations across temporal dependencies.

Main Results:

• Achieved superior performance on AVEC2014 with a Mean Absolute Error (MAE) of 5.25 and Root Mean Square Error (RMSE) of 6.75.
• Demonstrated strong results on AVEC2019 with a Concordance Correlation Coefficient (CCC) of 0.554 and RMSE of 4.61.
• Showcased significant improvements over existing methods focusing on momentary expressions.

Conclusions:

• The proposed framework effectively detects depression by analyzing temporal behavioral patterns from multimodal data.
• The graph-based approach integrating GRU and GNNs captures complex temporal dependencies for enhanced diagnostic accuracy.
• This method offers a significant advancement in leveraging technology for long-term depression assessment.