Presentation Information
[A-8-11]Real-Time Gaze Anomaly Detection for Edge Devices Using an Echo State Network
〇SILU WANG1, Kantaro Fujiwara2, Takeaki Yajima1 (1. Graduate School of Information Science and Electrical Engineering, Kyushu Univ., 2. Graduate School of Medicine and Faculty of Medicine, The University of Tokyo.)
Keywords:
Gaze anomaly detection,Echo State Network,Mahalanobis Distance,Free viewing task,Psychiatric disorder assessment
This study investigated a gaze anomaly detection method using eye movement data during a Free viewing task for the objective assessment of psychiatric disorders. Although eye movements may reflect changes in brain function, Free viewing data contain complex mixtures of saccades and fixations, making it difficult to detect abnormalities using simple features alone. In the proposed method, gaze coordinate data were preprocessed, and movement-related time-series components were input into an Echo State Network (ESN). The reservoir internal states obtained from the ESN were then evaluated using Mahalanobis Distance (MD). The distribution of internal states from healthy controls was defined as the normal pattern, and test data were classified as patient data when the MD score exceeded a threshold. Cross-validation using data from 10 healthy controls and 10 patients showed that the Movement Mode + train_p60 condition achieved a sensitivity of 0.84, specificity of 0.60, and balanced accuracy of 0.72. This performance was higher than that of the Morita score, which combines conventional eye movement features, and the Summary PCA + LR method. These results suggest that ESN-MD can represent temporal gaze movement patterns during Free viewing and may extract useful information for detecting patient groups.