Mobile Mental Health Screening in EmotiZen via the Novel Brain-Inspired MCoG-LDPSNet

Area of Science:

• Computational neuroscience
• Machine learning for mental health

Background:

• Millions face barriers to mental health care due to stigma and wait times.
• Mobile mental health apps offer accessible screening but struggle with imbalanced data.
• Existing machine learning models often yield biased predictions in imbalanced datasets.

Purpose of the Study:

• To develop an accurate and imbalance-aware model for mobile mental health screening.
• To address the limitations of current machine learning methods in handling class imbalance.
• To improve the reliability and calibration of predictions in mental health applications.

Main Methods:

• Proposed MCoG-LDPSNet, a brain-inspired model with dual encoding pathways.
• Introduced a novel Loss-Driven Parametric Swish (LDPS) activation with adaptive gain.
• Utilized a benchmark mental health corpus and social media text for evaluation.

Main Results:

• MCoG-LDPSNet achieved AUROC of 0.9920 and G-mean of 0.9451 on a benchmark corpus.
• Demonstrated superior performance over traditional and state-of-the-art deep learning models.
• Achieved an AUROC of 0.9937 after transfer learning to social media text.
• Integration into the EmotiZen App led to significant symptom reductions (anxiety 28.2%; depression 42.1%).

Conclusions:

• MCoG-LDPSNet offers an accurate, imbalance-aware solution for scalable mobile screening of anxiety and depression.
• The model's adaptive gain mechanism enhances minority-class pattern detection.
• The findings support the use of MCoG-LDPSNet for improving mental health accessibility and outcomes.