Optimizing depression detection in clinical doctor-patient interviews using a multi-instance learning framework.

Area of Science:

• Artificial Intelligence
• Psychiatry
• Computational Linguistics

Background:

• Depression diagnosis relies on subjective methods like self-rating scales and interviews, which can be unreliable.
• Existing automatic depression detection (ADD) methods using interview data often suffer from information redundancy and loss.
• There is a need for more objective, efficient, and interpretable methods for depression detection.

Purpose of the Study:

• To develop a novel multiple instance learning (MIL) framework for enhanced automatic depression detection from textual interview data.
• To improve text representation and information extraction from long interview texts.
• To enhance the interpretability of depression detection models.

Main Methods:

• Application of the multiple instance learning (MIL) framework to transcribed interview data for depression detection.
• Development of an ensemble model (multi-MTRB) combining MT5 and RoBERTa for feature extraction and confidence scoring.
• Introduction of hyper-parameters to manage text sentiment uncertainties and LIME techniques for in-depth interpretation.

Main Results:

• The proposed MIL-based method achieved excellent performance on the DAIC-WOZ and E-DAIC datasets.
• Achieved an F1 score of 0.88 on the DAIC-WOZ dataset and 0.86 on the E-DAIC dataset.
• Demonstrated high interpretability by identifying specific sentences indicative of depression.

Conclusions:

• The MIL framework offers a promising, interpretable, and accurate approach for automatic depression detection using textual interview data.
• The multi-MTRB model effectively captures local features and alleviates sample imbalance issues.
• This method provides a significant advancement over previous approaches in ADD for text-based analysis.