Detection and Classification in Electrocardiac Signals

Signal processing can be used to condition medical signals to facilitate their interpretation, and to extract clinically important information. The purpose of the present doctoral thesis is to put forward solutions to certain problems encountered in the processing of electrocardiac signals. Paper I examines the properties of a maximum likelihood method for aligning vectorcardiographic loops, with respect to different noise levels by using a simulation model. Loop alignment is performed in order to compensate for the influence of extracardiac noise which can be modeled by various geometrical and temporal transformations (scaling, rotation and time synchronization). It is shown that accurate loop alignment can be done at low to moderate noise levels, however, at a certain noise level, depending on loop morphology, the performance breaks down abruptly. Papers II and III deal with the detection of body position changes by analyzing the properties of ECG signals. It is well-known that such changes may falsely be interpreted as ischemia. The developed detector is based on the hypothesis that body positions are manifested in the VCG as sudden shifts in the electrical axis and, therefore, can make use of the rotation angles that result from the alignment of successive VCG loops. The results show, using different ECG databases, that reliable detection can be achieved. Papers IV and V address the problem of event detection and clustering in the electrogram. In implantable CRM devices, detection and classification of events must be performed with algorithms having a low computational complexity. A model of the electrogram waveforms is presented which allows for efficient feature extraction using the dyadic wavelet decomposition. A GLRT-based detector, relying on the wavelet coefficients, is derived and found to produce accurate detection; the performance in terms of probability of a missed detection and probability of a false alarm is less than 0.007 and 0.001, respectively. The wavelet coefficients are also serve as the basis for clustering of different electrogram morphologies. Clustering based on the leader-follower algorithm show that reliable clustering can be achieved, resulting in a reasonably low number of clusters