Modeling of Eye Movement System Using Orthogonal Visual Stimuli and Volterra Polynomial Representations

Abstract

This paper presents an approach to the modeling and diagnostic analysis of the human eye movement system (EMS) based on nonlinear integral models and Volterra polynomial representations in the form of multidimensional transient characteristics. Experimental “input-output” data were obtained using eye-tracking technology during identification experiments with three test visual stimuli positioned at different distances relative to the initial fixation point. The proposed methodology enabled the construction of adequate EMS models that combine high identification accuracy with a relatively simple mathematical representation. Two EMS models were formed from the experimental data: Model1, derived from horizontal experiments, and Model2, obtained from vertical experiments. Diagnostic feature spaces were generated using the transient characteristics of models for further psychophysiological state classification by machine learning methods. Two categories of feature spaces were investigated: heuristic features and features obtained from approximation and detail coefficients produced by wavelet decomposition of EMS transient characteristics. To improve dataset representativeness, augmentation with Gaussian noise levels of 1%, 3%, and 5% was performed. Classification efficiency was evaluated using Stratified K-Fold cross-validation with 8 folds for the original datasets and 32 folds for the augmented datasets. An exhaustive computational search was applied to determine the most informative feature combinations according to the probability of correct recognition (PCR) criterion. Pairwise feature combinations were analyzed for all datasets, whereas triplet combinations were additionally investigated for augmented data. The obtained results indicate that dataset augmentation considerably improves classification performance compared to the original data. In the heuristic feature spaces, the maximum PCR increased from 87.5% for the original datasets to 98.44% after augmentation for feature pairs, while the analysis of triplet combinations on augmented datasets produced PCR values approaching 100% within the applied cross-validation procedure. In the wavelet-based feature spaces, PCR values increased from 68.75% to 85.94% for augmented feature pairs and up to 95.31% for triplet combinations. The obtained findings confirm the effectiveness of the proposed intelligent information technology for psychophysiological state diagnostics based on quadratic Volterra EMS models and machine learning techniques. The proposed approach can be considered a basis for further development of machine learning methods for eye-tracking-based psychophysiological state assessment.