Through the Youth Eyes: Training Depression Detection Algorithms with Eye Tracking Data

Abstract

Depression is a prevalent mental health disorder, and early detection is crucial for effective intervention. Recent advancements in eye-tracking technology and machine learning offer new opportunities for non-invasive diagnosis. This study aims to assess the performance of different machine learning algorithms in. predicting depression in a young sample using eye-tracking metrics. Eye-tracking data from 139 participants were recorded with an emotional induction paradigm in which each participant observed a set of positive and negative emotional stimuli. The data were analyzed to find differences between groups, where the most significant features were selected to train prediction models. The dataset was then split into training and testing sets using stratified sampling. Four algorithms support vector machines (SVM), random forest (RF), a multi-layer perceptron (MLP) neural network, and gradient boosting (GB) were trained with hyperparameter optimization and 5-fold cross-validation. The RF algorithm achieved the highest accuracy at 84%, followed by SVM, GB, and the MLP neural network. Performance metrics such as accuracy, recall, F1-score, precision recall area under the curve (PR-AUC), and Matthews Correlation Coefficient (MCC) were also used to evaluate the models. The findings suggest that eye-tracking metrics combined with machine learning algorithms can effectively identify depressive symptoms in the young, indicating their potential as non-invasive diagnostic tools in clinical settings.