Agrupamiento no supervisado de latidos ECG usando características WT, Dynamic Time Warping y k-means modificado

Abstract

Heartbeat clustering is convenient in Holter record analysis and interpretation for data mining. Nevertheless, this clustering has to address several problems due to factors such as long signal length (high computational burden), noise and artifacts (patient movements, baseline wander, electrodeskin contact variability, powerline interference), and waveform variability because of patient's physiology and pathology. Therefore, it is very important to choose a suitable clustering configuration sufficiently robust against these problems. Such configuration includes the core of the clustering method itself, centroid initialization, dissimilarity measure, feature extraction and selection, and computational cost optimization techniques. To this aim, we present in this work an improved version of the k-means clustering algorithm. First, a preclustering stage is employed to reduce the initial heartbeat set using Dynamic Time Warping (DTW) and a suitable conservative threshold. In the clustering stage, a local search heuristic (kmedians) is analyzed. Feature extraction from heartbeats is carried out by using WT coefficients (from biortogonal 2.2 wavelet) and trace segmentation. The modified k-means considers both the dynamical and nonlinear behavior of ECG signal and provide best performance that k-means standard algorithm. As a result, the modified k-means algorithm is less sensitive to the presence of outliers and hence clustering error declines (up 8% in average) compared to the k-means standard algorithm. Non supervised analysis results are presented for a set of 44872 heartbeats with 16 different types of arrhythmia from MIT́s arrhythmia database. © Springer-Verlag Berlin Heidelberg 2007.