Number of arrhythmia beats determination in Holter electrocardiogram: how many clusters?

Holter signals correspond to long-term electrocardiograph (ECG) registers. Manual inspection of such signals is difficult because of the enormous quantity of beats involved. Throughout the literature several methods of automatically detecting and separating the significant beats using unsupervised learning were proposed. An important part of the unsupervised learning problem is determining the number of constituent clusters which best describe the data. In this paper we concentrate on the problem of the number of arrhythmia beats-clusters selection presented in Holter ECG. We apply and compare several criteria for assessing the number of clusters and we show that, with a Gaussian mixture model, the approach is able to select 'an optimal' number of arrhythmia beats and so partition a Holter ECG. The following criteria has been examined: Bayesian selection method, Akaike's information criteria, minimum description length, minimum message length, fuzzy hyper volume, evidence density and partition coefficient. We conclude that only minimum description length and Bayesian selection method are suitable for our real-world electrocardiogram data. In order to validate the procedure, an experimental comparative study is carried out, utilizing records from the MIT database.