Classification of QRS complexes: comparison of different methods

Early detection and classification of the QRS changes are of great interest in real-time monitoring such as in cardiac critical care units or operating room environments. The beat-to-beat classification of the QRS complex will permit to follow the hearth evolution and to detect arrhythmias.The learning capacity and the classification ability for normal beats (N), premature ventricular contractions (PVC), left bundled branch blocks (LBBB), right bundled branch blocks (RBBB), and paced beats (PB) clustering by four classification methods were compared. The methods were neural networks (NN), K-th nearest neighbour rule (Knn), discriminant analysis (DA) and fuzzy logic (FL). Twenty-six morphology feature parameters, which include information of amplitude, area, specific interval durations and measurement of the QRS vector in a VCG plane, were defined. They were measured for all the QRS complexes in the annotated MIT-BIH arrhythmia database. All of the 48 ECG recordings from the MIT-BIH arrhythmia database have been examined. One global, one basic, two local and two mixed learning sets were used. In fact one a small-sized learning set was formed, which contained the five types of QRS complexes collected from all patients in the MIT-BIH database and it was used either with or without applying the leave one out rule, thus representing either the global or the basic learning set respectively. The local learning sets consisted of heartbeats only from the tested patient, which were taken either consecutively or randomly. The mixed learning sets were obtained by combining the local sets with the basic one without applying the leave one out rule. Using the local and mixed learning sets the assessed methods achieved high accuracies, while the small size of the basic learning set was balanced by reduced classification ability. Expectedly, the worst results were obtained by using the basic learning set and the leave one out rule. The performances of Knn and NN methods were the best with randomly collected mixed and local learning sets, while for FL the randomly collected learning was the most suitable. The DA method achieved the best accuracy with the mixed learning set, containing consecutive heartbeats of the tested patient.