Comparison between different similarity measure functions for optimal clustering AEPs Independent Components
Norma Castañeda-Villa, Juan Manuel Cornejo-Cruz, Pilar Granados-Trejo · 2015
A critical part of applying Independent Component Analysis (ICA) to any neurophysiological data is the selection of relevant independent Components (ICs); i. e., to decide which ICs have neurological meaning. Standard ICA implementation supposes a square mixing matrix; this results in as many ICs as EEG channels. In this work, responses to repetitive auditory stimuli are the most important signals (Auditory Evoked Potentials, AEPs); so the ICs of interest should be repetitive and time-locked with the stimuli. In this paper an update of a previously proposed procedure for the objective selection of ICs using Mutual Information (MI) and cluster analysis is presented. This time, four different similarity functions are evaluated and three inter/intra-cluster quality criteria are explored to determine optimal cluster numbers to both synthetic AEPs and data from normal hearing children, so that to identify ICs related with the auditory response. The numbers of clusters and the similarity function that yield best results in both datasets, in other words optimal clustering AEPs ICs, were 8 and Euclidean link-clustering average respectively.