Extensions of Independent Component Analysis Mixture Models for classification and prediction of EEG signals

Abstract

[EN] This paper presents two applications of Independent Component Analysis Mixture Modeling (ICAMM) for the classification and prediction of data. The first one of these extensions is Sequential ICAMM (SICAMM), an ICAMM structure that takes into account the sequential dependence in the feature record. This algorithm can be used to classify input observations in a given set of mutually-exclusive classes. The performance of SICAMM is tested with simulations and compared against that of the base ICAMM algorithm and of a Dynamic Bayesian Network (DBN). All three methods are also used to classify real electroencephalographic (EEG) signals to compute hypnograms, a clinical tool used to help in the diagnosis of sleep disorders. The second extension of ICAMM is PREDICAMM, an estimation algorithm that makes use of the ICAMM parameters in order to reconstruct missing samples from a set of data. This predictor is used to reconstruct real EEG data from a working memory experiment, and its performance is compared to that of a classical predictor for EEG signals: sphere splines. Prediction performance is measured with four error indicators: signal-to-interference ratio, KullbackLeibler divergence, correlation, and mean structural similarity index. Both extensions of the base ICAMM algorithm have achieved a higher performance than other methods