Classification of schizophrenia based on RAnet-ET: resnet based attention network for eye-tracking

Objective.There is a notable need of quantifiable and objective methods for the classification of schizophrenia. Patients with schizophrenia exhibit atypical eye movements compared with healthy individuals. To address this need, we have developed a classification model based on eye-tracking (ET) data to assist physicians in the intelligent auxiliary diagnosis of schizophrenia.Approach.This study employed three ET experiments-picture-free viewing, smooth pursuit tracking, and fixation stability-to collect ET data from patients with schizophrenia and healthy controls. The ET data of 292 participants (133 healthy controls and 159 patients with schizophrenia) were recorded. Utilizing ET data in picture-free viewing, we introduce a Resnet-based Attention Network for ET (RAnet-ET) integrated with the attention mechanism. RAnet-ET was trained by employing multiple loss functions to classify patients with schizophrenia and healthy controls. Furthermore, we proposed a classifier for handling multimodal features that combines specific features extracted from the well-trained RAnet-ET, 100 ET variables extracted from three ET experiments, and 19 MATRICS Consensus Cognitive Battery scores.Main results.The RAnet-ET achieved good performance in classifying schizophrenia, yielding an accuracy of 89.04%, a specificity of 90.56%, and an F1 score of 87.87%. The classification results based on multimodal features demonstrated improved performance, achieving 96.37% accuracy, 96.87% sensitivity, 95.87% specificity, and 96.37% F1 score.Significance.By integrating attention mechanisms, we designed RAnet-ET, which achieved good performance in classifying schizophrenia from free-viewing ET data. The synergistic combination of specific features extracted from the well-trained RAnet-ET, MCCB scores, and ET variables achieved exceptional classification performance, distinguishing individuals with schizophrenia from healthy controls. This study underscores the potential of our approach as a pivotal asset for the diagnosis of schizophrenia.