Clustering is advisable technique for analysis and interpretation of long-term ECG Holter records. As a nonsupervised method, several challenges are posed due to factors such as signal length (very long duration), noise presence, dynamic behavior and morphology variability (different patient physiology and/or pathology). This work describes an improved version of the k-means
IntroductionLong-term analysis of the ECG signals is a sound technique to assess patient state and evolution in a number of diseases: cardiac arrhythmias, transient ischemic episodes and silent myocardial ischemia, which are not readily detected in a short-time electrocardiogram [1,2]. In this case, record analysis is performed off-line due to their long duration (hundreds of thousands of beats to examine), keeping in mind not to skip any beat, since the diagnosis might depend on just a few of them. Factors such as signal length, noise, dynamic behavior of signal and variability in the waveform by patient's physiology and pathology have to be considered, as well [3]. In this sense, unsupervised computer-aided analysis of Holter registers is a suitable choice regarding off-line analysis because it does not require a previous knowledge of heartbeat classes [4]. However, there are still some open problems because of the heartbeat duration and morphology variability.Recently, some solutions have been proposed. Namely, [5] studies the ECG heartbeat dynamics, which are represented by non-parametric feature extraction methods and compared by a dissimilarity measure, based on DTW.I n [6], ECG variability is considered regarding its morphology and duration, by means of a time-frequency representation of the signal using several WT families. Although, this work is aimed at classifying some cardiac pathologies, there is no clear way to extend the results to other common pathologies. Clustering techniques are among the most used non-supervised processing techniques, but selection of an specific clustering algorithm has to take into account several issues: computational cost, partition optimality, outlier detection, local or global convergence, and simplicity. It has also to be robust against highly unbalanced partitions, for instance, set of objects with a great difference in the number of instances of each class, etc.The main goal of present work is to improve the performance of a heartbeat clustering method for ECG registers, using WT coefficients as features. In this case, the nonstationarity nature of the WT allows to extract discriminant data from signal morphology [6]. The unsupervised analysis is performed by the heuristic search method Jmeans, which solves the minimization problem with a solution close to the global optimum [7]. A comparison is carried out between features, obtained by the WT coefficients. In order to remove time length differences, heartbeats are previously nonuniformly resampled by means of trace segmentation. Computational cost is relatively low, thus because a preprocessing stage removes obvious redundant heartbeats...