ScholarWorks@동국대학교

초록

This study aimed to develop a pupillometry indicator-based machine learning model for depression screening based on pupillary time-series responses to images. Forty participants completed an experiment in which they were presented with a total of 55 emotional stimuli. Pupil responses to each stimulus were treated as independent samples per participant, from which a total of 45 pupil-based features were extracted. A t-test was performed to compare pupil-based features between the highest and lowest 10% groups, as determined by depression scores on the Korean version of the Center for Epidemiologic Studies Depression Scale. The images were subsequently ranked based on their effectiveness in distinguishing between these groups. The results demonstrated that maximum prediction accuracy was achieved using only the top two ranked stimuli. This stimulus optimization process significantly reduced the number of stimuli and measurement duration to under one minute while maintaining high predictive performance. Synthetic minority oversampling technique was applied to address the severe imbalance between the depressed and non-depressed groups. The model was trained and evaluated using a light gradient boosting machine classifier with group 3-fold cross-validation. The optimized model exhibited excellent predictive performance, achieving an accuracy of 89.0%, precision of 80.0%, recall of 75.0%, F1 score of 0.74, specificity of 93.0%, and an area under the curve of 0.85. This model can effectively screen individuals at risk for depression under concise measurement conditions, suggesting future clinical applications as an accessible pupillometry indicator-based tool for depression screening.

키워드