Classification and Clustering of Functional Eye-Tracking Data for Autism Spectrum Disorders.

摘要

Eye-tracking experiments, in which the position of a subject's gaze is recorded over time, yield a wealth of data for psychological studies. However, the intrinsically high-dimensional nature of this data makes it difficult to analyze with traditional statistical methods, and the lack of smoothness of eye-tracking paths as functions of time limits the use of functional data analysis methods as well. In this work, some solutions to statistical analysis of scanpaths are presented. First, an algorithm to reduce the dimensionality of video scanpaths and prepare them for further analysis is described. Second, a hidden Markov model to describe static-image scanpaths and extract meaningful features of visual behavior is presented. Current methods for eye-tracking analysis frequently sacrifice short-term frame-by-frame structure of scanpaths in favor of simple averaging and often ignore the time-series aspect of this data; the proposed techniques in this work do not share these disadvantages, and thus offer improvements over existing analytic methods. The application of these methods to their respective types of eye-tracking data will enable sophisticated statistical analysis of a currently intractible type of data structure.