Source localization of ventricular arrhythmias using self-organizing neural networks

Body surface potential mapping (BSPM) data obtained during endocardial stimulation at multiple ventricular pacing sites show a broad spectrum of potential distributions. In this study, BSPM sequences are analysed using a neural network approach based on self-organisation that provides a noninvasive estimation of the site of origin of stimulated ventricular activation. The Self-Organizing Map (SOM) network used in this study is arranged as a two-dimensional lattice of neurons, each of them representing a particular distribution of body surface potentials. For the training of the SOM network, 123-channel BSPM recordings were obtained from 86 endocardial pacing locations in 19 patients with a previous myocardial infarction. Ventricular activation patterns from different pacing sites are visualized as time traces on the trained SOM. Classification of the activation patterns with respect to the endocardial pacing location is performed by Learning Vector quantization. The localisation results are visualized on a realistic model of the endocardial surfaces of the right and left ventricles