Joint segmentation of multiple features with application to analysis of epilepsy and music.

Abstract

Multivariate data is segmented into parts, called segments, with common characteristics. The segments are assumed to have an under- lying model structure. It is of interest to see whether the character- istic changes originate from a subset of features or a unique feature. Various approaches can be taken when constructing such a method for multiple features. Three di erent methods are devised and com- pared via a simulation study. The methods use a penalized likelihood in di erent ways to estimate the number of segments. Two of these methods exhibit positive results and are examined further. They are shown to have di erent capabilities. One method favors the detection of coordinated segment changes at the expense of nding those that originate in a unique feature. The other method has an overall better performance, i.e., it is better at locating each and every characteristic change. The two methods are applied in two real life settings, one measuring physical changes in coordination with various music and the other measuring a range of physical changes in epilepsy patients. The methods capture the trends in the data but are not able to detect precisely when the music changes or the beginning and the end of a seizure.