Classification and compression of ICEGS using Gaussian mixture models

Implantable cardioverter defibrillators (ICD) administer high voltage shock therapies to terminate dangerous cardiac arrhythmias. Improving the functionality of these devices to include online diagnosis based on intracardiac electrogram (ICEG) morphology and to log dangerous signals is important for their more widespread use. It is essential that the ICD implement a signal compression scheme due to the limited memory in the device. We have fitted Gaussian mixture models to the ICEG signals in order to investigate to what extent, nonlinear data models are advantageous in this application compared to the traditional linear approaches used in the field and to explore the common features between classification and compression. Results of fitting the mixture models show that typically a single Gaussian per class for classifiers and single Gaussian prediction models for data compression are adequate data representations provided the data is preprocessed to remove non-stationary behaviour