Heartbeat classification system using adaptive learning from selected beats

An adaptive system for the automatic processing of the electrocardiogram for the classification of heartbeats into beat classes that learns from selected beats is presented. A first set of beat labels is produced by the system by processing an incoming recording with an unadapted classifier. The beat labels are then ranked by a confidence measure calculated from the posterior probabilities estimates associated with each beat classification. An expert then validates and if necessary corrects a fraction of the least confident beats of the recording. The system adapts by first training a classifier using the newly annotated beats, and then combining the outputs with the unadapted classifier to produce an adapted classification system. The adapted system then updates the remaining beat labels of the recording. Data was obtained from the heartbeats obtained from the 44 non-pacemaker recordings of the MIT-BIH arrhythmia database classified into one of eleven classes. With no adaptation a classification accuracy of 63% was achieved. By adapting the classifier, classification accuracy could be increased to over 91%. Our results show that a significant boost in classification performance of the system is achieved even when a small number of selected beats are used for adaptation.