Mixtures of independent component analyzers for microarousal detection

This paper presents a study of the application of new variants of the Sequential Independent Analysis Mixture Models (SICAMM) to the modeling and classification of electroencephalographic (EEG) signals. The real application approached was the detection of microarousals in EEG signals from sleep apnea patients. In addition, the proposed methods were tested on synthetic data with probability density changing in time in order to imitate the intrinsic nonlinearity and nonstationarity of the EEG signals. Thus, sequential dependence and sensitivity analyses with controlled simulated data are included. The SICAMM-based methods were compared with Dynamic Bayesian Networks (DBN) implementing Gaussian mixture models and static versions of the methods. We demonstrate that the proposed methods obtain the best performance in microarousal detection and their parameters adapt better for EEG signal dynamic modeling.