Cardiac arrhythmia discrimination using evolutionary computation

The use of Implantable Cardioverter Defibrillators (ICD) for cardiac arrhythmia treatment implies a search for efficiency in terms of discrimination quality and computational complexity, given that improved efficiency will automatically turn into more effective therapy and longer battery lifetime. In this work, we applied evolutionary computation to create classifiers capable of discriminating between ventricular and supraventricular tachycardia (VT/SVT) in episodes registered by ICDs. Evolutionary computation comprises several paradigms emulating natural mechanisms for solving a problem, all of them characterized by a population of individuals (possible solutions) which evolve generation after generation to provide fitter solutions. Genetic programming was the paradigm chosen here because its solutions, coded as decision trees, can be both computationally simple and clinically interpretable. For the experiments, we considered electrograms (EGM) from episodes registered by ICDs in spontaneous/induced tachycardia, previously classified as VT/SVT by clinical experts from several Spanish healthcare centers. Training data were 38 real-valued samples, arranged as the concatenation of two beat segments: a sinus rhythm template immediately previous to the arrhythmic episode (basal reference), and the arrhythmic episode template. Several low complexity trees provided low error rates and allowed physiological interpretation. The best tree yielded an error rate of 1.8%, with both sensitivity and specificity above 98%. This solution compares two samples from the end of the arrhythmic pulse with another two samples from the sinus rhythm, pointing out to a relevant discrimination role of the lasting EGM.