Session S92.3

Nonparametric Density-Based Clustering for Cardiac Arrhythmia Analysis

JL Rodriguez-Sotelo*, D Peluffo-Ordoņez, D Cuesta-Frau, G Castellanos-Dominguez

Universidad Nacional de Colombia
Manizales, Columbia

It is often necessary to apply a data mining scheme to long term ECG Holter records. This is because of the large number of heartbeats and the presence of signal artifacts (electrode noise, patient movement, baseline wander). Additionally, there is a great wave morphology variability due to patient pathology and physiology, signal acquisition system, and lead. Therefore, to facilitate the task of Holter record analysis, it is very convenient to apply a non-supervised data mining method.
Clustering methods are the usual tool to perform this data mining task in Holter records. However, these methods also pose some inherent problems such as cluster initialization, cost function selection, data dimensionality and local convergence.
We describe in this work a clustering method applied to Holter records which is based on a new approach that reduces the problems stated above. It is a partitional clustering method based on Parzen windows. It does not require Gaussian assumptions, and it employs a soft member function, which improves the final partition obtained. Besides, to avoid the local convergence problem, centroid initialization is carried out using the JH-means algorithm.
This method was applied to a dataset drawn from the MIT/BIH arrhythmia database, with heartbeat types normal, left and right branch bundle blocks, ventricular extrasystoles and atrial premature beats. Heartbeat features were selected using the Q-alpha algorithm. Clustering quantitative results in terms of sensitivity and specificity achieved were higher than 95%.

(Abstract Control Number: 236)