Wavelet classification of high frequency pupillary responses

This paper addresses the problem of classifying users with different visual abilities on the basis of their pupillary response while performing computer-based tasks. Multiscale Schur monotone (MSM) summaries of high frequency pupil diameter measurements are utilized as feature (or input) vectors in this classification. Various MSM measures, such as Shannon, Picard, and Emlen entropies, the Gini coefficient and the Fishlow measure, are investigated to assess their discriminatory characteristics. A combination of classifiers, motivated by a Bayesian paradigm, is proposed to minimize and stabilize the misclassification rate in training and test sets with the goal of improving classification accuracy. In addition, the issue of wavelet basis selection for optimal classification performance is discussed. The members of the Pollen wavelet library are included as competitors. The proposed methodology is validated with extensive simulation and applied to high-frequency pupil diameter measurements collected from 36 individuals with varying ocular abilities and pathologies. The expected misclassification rate of our procedure can be as low as 21% by appropriately choosing the Schur monotone summary and using a properly selected wavelet basis.