Feature extraction based on sparse graphs embedding for automatic depression detection

Background and Significance: Automatic detection of depression is crucial in today's fast-paced, depression-prone society. However, the current diagnosis still relies on manual assessment by psychologists using the Patient Health Questionnaire-9 (PHQ-9), resulting in delays in early detection and treatment. This paper proposes an automatic feature extraction method called Sparse Graphs Embedding (SGE) for depression detection. The further goal is to integrate SGE into the healthcare system, enabling automatic depression diagnosis during physical examinations for early detection and early treatment. Method and Novelty: The method addresses several key challenges, including the maintenance of local information in the fNIRS signal space, the removal of outliers, and the rejection of redundant information between features. Specifically, it constructs two weighted graphs embedded in the between-class scatter and within-class scatter, preserving the neighborhood relationships within the data. To mitigate sensitivity to outliers, the ℓ_2,1-norm is embedded in both the between-class and within-class scatter as a fundamental measure. Additionally, sparse orthogonality projections are employed to extract effective features for depression detection and eliminate redundant information. Results: Experimental results demonstrate promising detection rates of 86.1%, 82.8%, and 92.2% using functional near-infrared spectroscopy (fNIRS) under positive, neutral, and negative emotional stimuli, respectively. These rates surpass those achieved by other state-of-the-art methods. The promising results showcase the potential of fNIRS, and the weighted graph embedding approach sheds new light on depression detection.