Uncertainty aware domain incremental learning for cross domain depression detection.

Area of Science:

• Computational psychiatry
• Machine learning applications in mental health

Background:

• Deep learning shows promise for Major Depressive Disorder (MDD) detection from text.
• Existing methods struggle with limited data, privacy, domain gaps, and uncertainty in real-world applications.

Purpose of the Study:

• To propose an Uncertainty-Aware Domain Incremental Learning framework for Cross-Domain Depression Detection (UDIL-DD).
• To overcome challenges of data privacy, domain gaps, class imbalance, and uncertainty in depression detection.

Main Methods:

• Developed Uncertainty-guided Adaptive Class Threshold Learning (UACTL) to measure prediction discrepancies and incorporate uncertainty.
• Implemented Data-Free Domain Alignment (DFDA) to approximate historical feature distributions without accessing prior data, mitigating catastrophic forgetting.

Main Results:

• The UDIL-DD framework demonstrated effectiveness in cross-domain depression detection.
• Experiments on four benchmark datasets (CMDC, DIAC-WoZ, MODMA, EATD) validated the method's robustness.

Conclusions:

• The proposed UDIL-DD framework offers a reliable approach for depression detection in real-world clinical scenarios.
• The integration of UACTL and DFDA effectively handles data privacy and domain shift challenges.