A Noise Tolerant Method for ECG Signals Feature Extraction and Noise Reduction

Abstract

A noise tolerant template model technique for ECG feature extraction based on an individual-specific training approach is presented in this paper. Baseline wander, electrode motion artifacts, and electromyographic interference were added with varying signal-to-noise ratios (SNRs) to a dataset of approximately 3000 beats of different ECG recordings from the QT database to validate the performance of our technique. All of the QRS-complex, P-and T-waves detectors achieved an average sensitivity of 96.11%, positive predictivity of 83.8% and accuracy detection rate of 81.9% for SNRs between 24dB and -6dB, outperforming four recent beat detection algorithms evaluated with respect to the same types of noise. Furthermore, the ability of our technique to achieve efficient noise reduction including in-band noise, while preserving the morphological and clinical information of the original signal, is described.