Machine Learning Effectiveness of a Psychophysiological State Classification System based on Eye Tracking Technology

This study investigates a modeling approach for the human eye movement system (EMS) based on Volterra models represented as multidimensional transient characteristics obtained from experimental "input-output" EMS data using innovative eye-tracking technology. The respondent was examined both in the morning and in the evening, which enabled the evaluation of the variability in EMS transient characteristics. EMS identification was carried out using first- and second-order transient functions determined by the least squares method with step test signals of varying amplitudes. Based on the resulting EMS transient characteristics, several heuristic feature spaces were constructed, including those obtained before and after wavelet filtering of empirical data, as well as after wavelet filtering of the transient characteristics themselves. An additional feature space was formed using the wavelet decomposition coefficients of the EMS model characteristics. Statistical machine learning methods were applied to build classifiers, particularly the Bayesian classifier and the support vector machine (SVM) method. The informativeness of the features was analyzed, and feature pairs with the highest probability of correct recognition (PCR) of the psychophysiological state were identified. The Bayesian classifier achieved a PCR close to 100%, while the SVM-based classifier reached a maximum PCR of 93.75%.