An important goal of clinical electrophysiological studies is estimation of the source of rhythm disturbances (arrhythmia) in the heart. 15% of ventricular arrhythmias are known to originate from the outer surface of the heart (epicardium). One localization approach targeting the epicardium uses multielectrode catheters placed in the coronary veins. However, epicardial measurement sites from these catheters are limited to locations reached via the coronary veins. This study investigates the feasibility of several pattern classification and neural network approaches for localization of the source of ventricular arrhythmias from sparse measurements acquired from within coronary veins. Specifically, we studied Kohonen self-organizing maps and fuzzy C-means clustering methods for the construction of the target vector in neural networks from experimental high-resolution activation-time patterns. We also studied two neural network techniques, probabilistic neural networks and backpropagation networks, for the training and test procedures. The results of this study showed that it was possible to localize the arrhythmia source in a relatively small region for approximately 70% of cases. This study, in general, showed that the combination of probabilistic neural networks, Kohonen self-organizing maps and fuzzy C-means clustering approaches is feasible in catheter-based epicardial arrhythmia source localization.