Gaze Behavior based Depression Severity Estimation

Depression is a severe mental illness that affects millions of people's lives Every year, more than 700,000 individuals choose to commit suicide due to depression. Early screening of patients with depression is challenging, making it critical to find methods that can aid doctors in quickly iden...

Full description

Saved in:
Bibliographic Details
Published in2023 IEEE 4th International Conference on Pattern Recognition and Machine Learning (PRML) pp. 313 - 319
Main Authors Shu, Tao, Zhang, Feixiang, Sun, Xiao
Format Conference Proceeding
LanguageEnglish
Published IEEE 04.08.2023
Subjects
Behavioral sciences
Depression
Depression detection
Depression severity estimation
Estimation
Gaze behavior
Machine learning
Medical services
Pattern recognition
Reliability
Online AccessGet full text

Cover

More Information
Summary:Depression is a severe mental illness that affects millions of people's lives Every year, more than 700,000 individuals choose to commit suicide due to depression. Early screening of patients with depression is challenging, making it critical to find methods that can aid doctors in quickly identifying potential patients. Currently, a significant number of depression severity estimation tasks focus on text, speech, and vision, but the exploration of the relationship between eye gaze behavior and depression is incomplete. This paper aims to explore the mapping relationship between eye gaze behavior and depression. We propose an integrated method to estimate depression severity using eye gaze behavior data. We formulate a more reasonable representation of gaze direction by expressing it as the angle of deflection. We employ a deep residual shrinkage network with channel-wise thresholds (DRSN-CW) to process the input gaze sequence and estimate the depression severity score. Moreover, we introduce label distribution learning to represent more fine-grained score information and alleviate the overfitting problem associated with small datasets, enhancing the reliability of depression severity estimation. We train and validate our model on the Distress Analysis Interview Corpus - Wizard of Oz (DAIC-WOZ) dataset, achieving an F1 score of 0.769, and reducing the mean absolute error (MAE) to 3.971 and root mean square error (RMSE) to 5.141. The results demonstrate significant improvement in the depression severity estimation task compared to previous tasks.
DOI:10.1109/PRML59573.2023.10348319