Detecting Emotional Disorder with Eye Movement Features in Sports Watching

Wei Qiang ^1,2 Lin Yang ³ Xucheng Zhang ⁴ Na Liu ³ Yanyong Wang ³ Jipeng Zhang ⁵ Yixin Long ⁶ Weiwei Xu ^3* Wei Sun ^4*

• ¹ Institute of Software Chinese Academy of Sciences, Beijing, China
• ² University of Chinese Academy of Sciences, Beijing, China
• ³ The First hospital of Hebei Medical University, Shijiazhuang, Hebei Province, China
• ⁴ Institute of Software, Chinese Academy of Sciences (CAS), Beijing, Beijing Municipality, China
• ⁵ Health Service Department of the Guard Bureau of the Joint Staff Department, Beijing, China
• ⁶ Capital Medical University, Beijing, China

Introduction: Digital technologies have significantly advanced the detection of emotional disorders (ED) in clinical settings. However, their adoption for long-term monitoring remains limited due to reliance on fixed testing formats and active user participation. This study introduces a novel approach utilizing common ball game videos-table tennis-to implicitly capture eye movement trajectories and identify ED through natural viewing behavior.Methods: An eye movement data collection system was developed using VR glasses to display sports videos while recording participants' eye movements. Based on prior research and collected data, four primary eye movement behaviors were identified, along with 14 associated features.Statistical significance was assessed using t-tests and U-tests, and machine learning models were employed for classification (SVM for single-feature analysis and a decision tree for significant features) with k-fold validation. The reliability of the proposed paradigm and extracted features was evaluated using intraclass correlation coefficient (ICC) analysis.Results: Significance tests revealed 11 significant features in table tennis videos, encompassing exploration, fixation, and saccade behaviors, while only 3 features in tennis videos, which served as a supplemental stimulus, were salient in the re-testing. GazeEntropy emerged as the most predictive feature, achieving an accuracy of 0.88 with a significance p-value of 0.0002. A decision tree model trained on all significant features achieved 0.92 accuracy, 0.80 precision, and an AUC of 0.94. ICC analysis further confirmed the high reliability and significance of key features, including GazeEntropy and fixation metrics (average, maximum, and standard deviation).Discussion: This study highlights the potential of ball game video viewing as a natural and effective paradigm for ED identification, particularly focusing on two key characteristics of ED:1 Sample et al.curiosity exploration and psychomotor function. Additionally, participant preferences for video content significantly influenced diagnostic performance. We propose that future in-home, longterm monitoring of psychological conditions can leverage interactions with daily digital devices, integrating behavioral analysis seamlessly into everyday life.