There is a growing interest in the analysis of beat-to-beat variations of the morphology (BBM) of cardiac waves in electrocardiograms (ECG). Such analyses are confronted with the low BBM-to-noise ratio. An ECG clustering technique is introduced that brings the benefits of signal averaging to BBM analysis and recovers the beat-to-beat pattern of BBM. ECG clustering aligns waves and sorts them into clusters. The precision of the alignment was enhanced by sub-sample alignment. Kohonen's self-organising neural networks identified the clusters of the cardiac waves during training. The subsequent clustering of a wave results in a label for the closest cluster, a distance to the cluster and optimal alignment. Furthermore, ECG clustering avoids base-line variations and amplitude modulation sufficiently to be applied to the QRS wave in the raw ECG. The technique is demonstrated on 14 subjects with coronary heart disease and no myocardial infarction, myocardial infarction, or inducible ventricular tachycardia. ECG clustering is a general-purpose technique for beat-to-beat analysis, where the variations are cyclic as in the sinus rhythm. Results show that beat-to-beat variations in the QRS morphology are in general cyclic, with a main period of about four cardiac cycles. All calculations were performed with the Cardio software.