The problem of detecting clinical events related to cardiac arrhythmias in long term electrocardiograms is a difficult one due to the large amount of irrelevant information that hides such events. This problem has been addressed in the literature by means of clustering or classification algorithms that create data partitions according to a cost function based on heartbeat features dissimilarity measures. However, studies about the type or number of heartbeat features is lacking. Usually, the feature sets used are relevant but redundant, which degrades algorithm performance. This paper describes a method for automatic selection of heartbeat features. This method is assessed using real signals from the MIT database and common features used in previous works.