Classification of supraventricular and ventricular beats by QRS template matching and decision tree

This study presents a two-stage heartbeat classifier. The first stage makes initial assignment of beats towards continuously updated beat templates of the predominant rhythm, and calculates a set of features, tracking the morphology and RR-interval variation, and correlation to noise robust average beat templates. The second stage implements a decision tree for classification of supraventricular (SVB) and ventricular beats (VB). The training process on 3 large ECG databases (AHA, EDB, SVDB) applies splitting and pruning of the tree to different levels. A solution with 150 decision nodes and error cost <;0.01 is selected for unbiased test-validation with MIT-BIH database, showing: specificity=99.7% for SVBs, sensitivity=95.9%, positive predictivity=95.1% for VBs. Decision trees combine high performance, rapid interpretation and easy configuration of the complexity.