Session S34.2

Premature Ventricular Beat Detection by Using Spectral Clustering Methods

B Ribeiro*, A Marques, J Henriques, M Antunes

University of Coimbra
Coimbra, Portugal

Arrhythmias are a disturbance in the rhythm of the heart that can range from a mild skipped beat to a life-threatening failure to pump. Among the ventricular arrhythmias, the premature ventricular contraction (PVC) is of great importance since if its occurrence is higher than normal it increases the risk of sudden death in patients.
In this paper, we the look at the spectral properties of features extracted from segmented ECG signals containing Normal (N) and premature ventricular beats (V) prior to apply classification methods for reliable PVC detection. In a first stage, feature extraction based on signal basic analysis which computes not only intervals and amplitudes on each beat, but also description of wave morphology was performed. For this research work, each beat is described by set of 18 parameters. Extracted parameters that describe the basic shape of the beat such as: average wave amplitudes, durations and areas have been computed. In a second stage, the eigen decomposition of data allows finding structure in records which is optimal to attain high performance of classification. In a third stage, Support Vector Machines (SVM) which are benchmarked against several techniques have been chosen for PVC detection.
Experimental data used to test the proposed approaches were taken from MIT-BIH Arrhythmia Database. The training, test and validation sets consist each of 19391 samples of 18 extracted features. By applying SVM Recursive Feature Elimination (SVM RFE) where the weight magnitude is used as ranking criterion we reduced the feature dimension to smaller sets. Then, with newly constructed dimension input features space we combine spectral clustering with SVM classifiers for attaining superior performance.
We obtained high beat detection performance with sensitivity of 95.73% and a positive predictability of 99.74%. Moreover, premature ventricular contraction beats were detected using an original approach combining a spectral clustering method followed by a classification strategy. The performances obtained allow us to point out the advantages of our approach according to the state of the art. The results obtained validate our approach for real world application.

(Abstract Control Number: 55)