Session P79.6

A New Method for Unsupervised Analysis of ECG Beats Based on WT Features, DTW and J-Means

JL Rodríguez*, D Cuesta, G Castellanos

Universidad Nacional de Colombia
Manizales, Columbia

For Holter record analysis and interpretation, heartbeat clustering is necessary, nevertheless, its automation represents several challenges due to factors such as as signal length, noise and artifacts (patient movements, baseline wander, etc.), dynamic behavior of signal by poor contact between skin and electrode, and variability in the waveform by patient's physiology and pathology. Consequently, non-supervised analysis of ECG signals is the most appropriate, though, it carries other issues: computational cost, centroid initialization method, dissimilarity measure selection, high dimensionality of features; most of them are still open problems. In this work an improved version of k-means is presented. It addresses the iterative computation and centroid initialization problems. First, a preclustering stage is employed to reduce the initial set using Dynamic Time Warping (DTW) and a suitable threshold. In the clustering stage, three well-known local search heuristics (k-medians, H-means, J-means) are analyzed. Non supervised analysis results are presented for a set of 109871 heartbeats with 16 different types of arrhythmia from MIT's arrhythmia database. Feature extraction from heartbeats is done through the use of WT coefficients and trace segmentation. As result, the j-means algorithm is less sensitive to the presence of outliers, thence clustering error declines (up 8% in average) compared to the k-means simple algorithm. For this particular case of analysis, among the proposed techniques, the best performance was obtained with J-means, since it guarantees a local optimum by solving the problem of the minimum sum of squares.

(Abstract Control Number: 66)