Improving depression prediction using a novel feature selection algorithm coupled with context-aware analysis.

Area of Science:

• Computational psychiatry
• Machine learning in healthcare

Background:

• Developing machine learning (ML) models for depression prediction from long-term recordings is crucial yet challenging for clinical diagnosis.
• Accurate depression diagnosis is vital for timely intervention and treatment.

Purpose of the Study:

• To develop a novel, high-performance machine learning-based depression prediction method.
• To identify key features and topics indicative of depression from multimodal data.

Main Methods:

• A two-stage feature selection algorithm was applied to high-dimensional features (over 30,000) from the DAIC-WOZ dataset.
• Context-aware analysis integrated audio, video, and semantic features.
• The proposed method was compared against seven reference models.

Main Results:

• The method achieved excellent depression classification performance (F1-score: 0.96/0.67, Precision: 1.00/0.63, Recall: 0.92/0.71 on development/test sets).
• Promising results were obtained for depression severity estimation (RMSE: 4.43/5.11, MAE: 3.22/3.98).
• Audio features were predominant for classification, while all three feature types contributed equally to severity estimation.

Conclusions:

• The developed pipeline for depression recognition, along with identified topics and features, can support the diagnosis of psychological distress.
• Further inclusion of depression samples and optimization of the feature selection process are recommended.