Window-Based Quaternion Principal Component Analysis of Eye Gaze Dynamics for Depression Severity Prediction

Depression is a global health issue that necessitates severity identification for effective treatment. Although early detection and monitoring of mental abnormalities are crucial for mitigating adverse conditions, standard diagnostic approaches for depression, including self-rating scales and clinician-based interviews, fall short of meeting these needs. Wearable sensors offer an opportunity for unobtrusive evaluation of patients' status throughout their daily life, but it presents challenges in feature extraction and depression severity prediction. Studies have shown that eye gaze dynamics in term of 3-D rotations are a good indicator of depression. However, their traditional representation in Euler angles is computationally inefficient and overlooks the channel-wise interrelations, thereby reducing their discriminative power in medical diagnosis. This paper proposes a novel approach to address these challenges. It uses a window-based quaternion principal component analysis algorithm to extract features of eye gaze dynamics, preserving the channel-wise interrelations and temporal information. The proposed algorithm was tested and validated using a benchmark depression dataset from previous studies. Results showed that the algorithm achieved higher accuracy in depression severity prediction with a root mean square error of 5.08, outperforming existing state-of-the-art prediction methods that only use visual data. The results were further analyzed and discussed, providing valuable insights into the effectiveness and potential applications of wearable sensor for mental health diagnostics.