Electrocardiogram Beat Classification Using Data Filtration Technique and Support Vector Machine

This work addresses the automatic classification of arrhythmia beats into four generalized classes as described by the Association for the Advancement of Medical Instrumentation (AAMI) standard. We propose a method that includes time-series, statistical and frequency features of RR-interval, DWT, and EMD analysis of QRS morphology. Also, a data filtration technique using support vector selection and under-sampling is applied to find those features as well as data points having significant prediction capabilities. While testing the above combination on MIT-BIH arrhythmia database, adopting the inter-patient paradigm, we achieved 70%, 99.79%, 64.5%, and 80.55% Se and 61.76%, 94.64%, 83.22%, and 77.48% PPV for F, N, SVEB, and VEB classes respectively. Further, the proposed method reduced the classifier’s complexity through feature selection and computation time by data reduction while maintaining the generalization capability of the model. Another finding includes the significant contribution that RR-interval, 180–360 Hz and 0–45 Hz band power, and non-linear statistical characteristics have in distinguishing the arrhythmia classes. The feature and data selection criterion used is F -score and one-class classification by RBF-SVM respectively. The classifier used for building the final model is SVM with the cubic kernel.