The most common way to diagnose cardiac dysfunctions is the ECG signal analysis, usually starting with the assessment of the QRS complex as the most significant wave in the electrocardiogram. Many methods for automatic heartbeats classification have been applied and reported in the literature but the use of different ECG features and the training and testing on different datasets, makes their direct comparison questionable. This paper presents a comparative study of the learning capacity and the classification abilities of four classification methods--Kth nearest neighbour rule, neural networks, discriminant analysis and fuzzy logic. They were applied on 26 morphological parameters, which include information of amplitude, area, interval durations and the QRS vector in a VCG plane and were tested for five types of ventricular complexes--normal heart beats, premature ventricular contractions, left and right bundled branch blocks, and paced beats. One global, one basic and two local learning sets were used. A small-sized learning set, containing the five types of QRS complexes collected from all patients in the MIT-BIH database, was used either with or without applying the leave one out rule, thus representing the global and the basic learning set, respectively. The local learning sets consisted of heartbeats only from the tested patient, which were taken either consecutively or randomly. Using the local 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 with the global learning set.