Joint-IVA for identification of discriminating features in EEG: Application to a driving study

Abstract

We propose a new method, joint independent vector analysis (jIVA), for obtaining discriminating features, i.e., interpretable signatures from medical data that can be used to study differences between multiple conditions or groups. The method is especially attractive for event related studies of electroencephalogram (EEG) data as it enables one to take advantage of the cross information across multiple channels effectively while enabling the use of information from multiple epochs. We introduce the general model and then demonstrate its successful application to EEG data collected during a driving experiment. As opposed to traditional analysis techniques that only detect differences, we identify statistically significant differences in measured band power showing when and how the differences occur for two experimental conditions across the same group of subjects. We compare jIVA features to those produced from competing data-driven approaches and demonstrate the advantages of jIVA as it fully leverages the statistical dependencies across multiple electrodes, and note its promise as a powerful data-driven method of obtaining informative features of multiset data.