Dual-Stream Multiple Instance Learning for Depression Detection With Facial Expression Videos.

Area of Science:

• Computer Science
• Psychiatry
• Machine Learning

Background:

• Depression is a prevalent mental illness causing significant individual harm.
• Objective physiological signals are increasingly linked to depression, necessitating automated detection methods.
• Facial expressions offer a potential avenue for depression classification, aiding online diagnosis and rehabilitation.

Purpose of the Study:

• To investigate the classification of depression using facial expressions via a weakly supervised learning approach.
• To develop and evaluate a novel Multiple Instance Learning (MIL) dual-stream aggregator for enhanced depression detection.
• To establish a new framework for psychiatric disorder detection using raw facial expression data.

Main Methods:

• A weakly supervised learning approach employing Multiple Instance Learning (MIL) was utilized.
• Data comprised 150 videos from 75 depressed and 75 healthy subjects.
• A novel MIL dual-stream aggregator (ADDMIL) incorporated instance-level max-pooling and bag-level attention weights.

Main Results:

• The ADDMIL method achieved 74.7% accuracy and 74.5% recall on the dataset.
• This represents a significant improvement over the baseline, with a 10.1% increase in accuracy and 9.8% in recall.
• The results surpassed the best reported accuracy for MIL-based methods by 2.1%.

Conclusions:

• Multiple Instance Learning (MIL) demonstrates substantial potential for depression classification through facial expressions.
• The proposed weakly supervised learning approach offers a novel framework for depression detection.
• This method could be adapted for the detection of other psychiatric disorders.