A new discrimination algorithm, based on the analysis of ventricular electrogram (EGM) onset, was proposed in order to discriminate between supraventricular and ventricular tachycardias (SVTs and VTs) in implantable cardioverter defibrillators (ICDs). Due to the absence of a detailed statistical model for the ventricular activation, this algorithm was based on a support vector method (SVM) learning machine plus bootstrap resampling to avoid overfitting. This SVM classifier was trained with available arrhythmia episodes, so that it can be viewed as containing a statistical model for the differential diagnosis. However, the black-box model character of any learning machine presents problems in a clinical environment. A solution is the extraction of the statistical information enclosed in the black-box model. The SVM could be appropriate for this purpose, given that the support vectors represent the critical samples for the classification task. In this article we propose two SVM-oriented analyses and their use in building two new differential diagnosis algorithms based on the ventricular EGM onset criterion.