Blink characterization using curve fitting and clustering algorithms

Abstract

The motion of the upper eyelid during blinking can be important in identifying possible diseases and syndromes that affect the eye. Hypothesized lid motion functions are fit to the dynamic position of the center of the upper lid under four experimentally controlled conditions in a pilot study. The coefficients of these nonlinear fits are used to classify blinks. Agglomerative hierarchical and spectral clustering were used to attempt an automatic distinction between partial and full blinks as well as between normal and abnormal blinks. Results for both approaches are similar when the input data is suitably normalized. Clustering finds outlying blinks that do not fit the model functions for lid motion well and that differ from the majority of blinks in our sample; however, those blinks may not be outliers based on easily observed data such as blink amplitude and duration. This type of analysis has potential for studying blink dynamics under normal and pathological conditions, but more work is needed with larger sets of data from blinks.