Effective schizophrenia recognition using discriminative eye movement features and model-metric based features

Abstract

Eye movement abnormalities have been effective biomarkers that provide the possibility of distinguishing patients with schizophrenia from healthy controls. The existing methods for measuring eye movement abnormalities mostly focus on synchronic parameters, such as fixation duration and saccade amplitude, which can be directly obtained from eye movement data, while lack of considering more thorough features. In this paper, to better characterize eye-tracking dysfunction, we create a dataset containing 100 images with eye movement data of 40 patients and 30 healthy controls via a free-viewing task, and propose two types of features for effective schizophrenia recognition, i.e. the hand-crafted discriminative eye movement features and the model-metric based features via utilizing the computational models of fixation prediction and the metrics of evaluating their prediction performance. Using the proposed features, two commonly used classifiers including support vector machine and random forest have been trained for classification between patients and controls. Experimental results demonstrate the effectiveness of the proposed features for improving classification performance, and the potential that our method can serve as an alternative and promising approach for the computer-aided diagnosis of schizophrenia.