Elsevier

Measurement

Volume 45, Issue 3, April 2012, Pages 474-487
Measurement

Delineation of ECG characteristic features using multiresolution wavelet analysis method

https://doi.org/10.1016/j.measurement.2011.10.025Get rights and content

Abstract

A discrete wavelet transform (DWT) based feature extraction technique in the QT segment of digitized electrocardiograph recordings is proposed. At first, the signal is denoised by decomposing it using DWT technique and discarding the coefficients corresponding to the noise components. A multiresolution approach along with an adaptive thresholding is used for the detection of R-peaks. Then Q, S peak, QRS onset and offset points are identified. Finally, the T wave is detected. By detecting the baseline of the ECG data, height of R, Q, S and T wave are calculated. For R-peak detection, proposed algorithm yields sensitivity and positive predictivity of 99.8% and 99.6% respectively with MIT BIH Arrhythmia database, 99.84% and 99.98% respectively with PTB diagnostic ECG database. For time plane features, an average coefficient of variation of 3.21 is obtained over 150 leads tested from PTB data, each with 10,000 samples.

Highlights

► Use of DWT for multiresolution analysis of the ECG data. ► Denoising is performed by discarding detailed coefficients D1 and D2. ► QRS zone is selected by D4 + D5 and adaptive thresholding. ► Fiducial points are determined by amplitude and slope reversal point detection. ► Average sensitivity and predictivity of 99.8% and 99.6% with mit-db obtained.

Introduction

Electrocardiogram (ECG) is widely used for diagnosing many cardiac diseases, which are one of the prime causes of mortality all over the world. The origin of ECG is the electrical activation of heart muscle cell causing sequence of depolarization and repolarization of its membrane. The electrical pulses generated due to this electrical activation are propagated along the cell fiber and transmitted to adjoining cells. The result is generation of electrical impulses, which travels through the cardiac surface. These electrical impulses can be detected by surface electrodes, amplified and displayed as the ECG. From electrical point of view, the heart is situated at the center of the electrical field it generates. The intensity of its electric field diminishes with the distance from its origin. A 12-lead electrode system is used for ECG recording, exploring an overall view of the heart’s electrical activity. ECG waveform consists of five different component waves, namely P, Q, R, S and T wave followed by a conditional U wave. A typical ECG beat is shown in Fig. 1. The durations and intervals of the constituent waves and amplitudes of the wave peaks reveal clinically significant information to the cardiologists for diagnosis [1].

Computerized processing of ECG for assisted diagnosis is an established area of biomedical research from long time. Performance of an automatic ECG analyzing system depends upon the reliable and accurate detection of the QRS complex and P and T waves along with the measurement of the QT segment. Detection and measurement of characteristic waves are related to diagnosis of various cardiac functions. For example, QRS detection is necessary to determine the heart rate and is used as reference for beat alignment. Likewise, ST segment elevation or depression is related to Myocardial Infarction.

The automatic delineation of the ECG is widely studied and algorithms are developed for QRS detection and wave detection [2], [3], [4]. A real time QRS detection algorithm, implemented in assembly language is developed by Pan and Tomkins [5]. Some early software based QRS detectors are presented [6], [7], [8]. ECG signal is normally corrupted with several noises, some of which are of physiological origin and others external. These are power line frequency interference, baseline drift, electrode contact noise, polarization noise, muscle noise, electrosurgical noise and the internal amplifier noise. So, denoising of the signal is a prerequisite to the accurate analysis. The noise sensitivity of nine different QRS complex detectors analyzed by Friesen et al. [8].

Frequently, for computerized determination of wave peaks and corresponding onset and offset points, QRS detection is the starting point. QRS location is represented by R peak index, or QS peak index in case of positive R absent. In time domain processing of samples, the individual fiducial points are determined with respect to the corresponding QRS index of the same beat by window search of appropriate width using magnitude and slope based criteria. Since ECG morphology varies among different age groups, communities and subcontinents, in addition to numerous types of cardiac diseases, there exists no golden rule for detection of the wave boundaries accurately. In an attempt to develop more accurate algorithms for extraction of features different approaches are used. Some algorithms are based on mathematical models [9], [10]. A simple, mathematical based method along with the concept of data structure is used to obtain the complex [11]. A new mathematical based QRS detector using CWT is explored in [30]. Some other approaches like matched filters [12], ECG slope criteria [13], second order derivatives [15], wavelet transforms [16], [17] are also studied. In [16], a multi-scale QRS detector including a method for monophasic P and T waves was proposed. Method based on evolutionary optimization process [18] is reported for wave delineation. ECG beats are classified by neuro-fuzzy networks where predefined feature vectors are used [19]. A rule mining based method is developed [20], where ischemic beats are identified by extraction of features followed by feature discretization and rule mining. ECG features was also extracted using linear predictive coding in [21]. Natalia et al. [29] has analyzed first derivative based QRS detection algorithm. A wavelet based soft decision techniques is proposed for identification of patients with cognitive heart failure [32], it uses power spectral density and soft computing techniques for the purpose. Feature measurement of ECG beats based on statistical classifier is explored in [33]. Integration of independent component analysis and neural networks for ECG beat classification is also an effective technique for classification of ECG beats [34]. A method of analysis of Myocardial Infarction is explored using discrete wavelet transform in [35].

The wavelet transform (WT) provides a description of a signal, decomposing it at different time–frequency resolution. WT is a well suited tool for analysis of non-stationary signals like ECG. The different wave components of ECG having separate frequencies, becomes clearly visible when subjected to multiresolution analysis. Moreover the various noise levels, which appear at different frequency bands and their contribution towards distortion of the signal, can be clearly identified.

In this paper, a discrete wavelet transform (DWT) based ECG feature extraction technique in the QT region is proposed. The developed algorithm extracts various clinical signatures from the ECG data by determining the fiducial points from a single lead data. Some of the extracted time-plane features are QRS width, QT interval, R height, T height. The algorithm validated using ECG records from PTB diagnostic ECG database (ptb-db) and MIT-BIH arrhythmia database (mit-db) under Physionet [23], [31] are used. After decomposing the signal using DWT, well localized frequency domain features are obtained at different levels of decomposition. QRS complex and T wave frequency bands are identified along with the frequency levels for different noises. The signal is denoisied by discarding those frequency bands corresponding to noise levels. A multiresolution approach along with thresholding is used for the detection of R-peaks in each cardiac beat. Hence, the heart rate is calculated. Then other fiducial points (Q and S) are detected by differentiation and slope criteria based search. QRS onset and offset points are detected. Finally, the T wave peak is detected and QT interval is measured. Baseline is also detected in the TP segment and height of R, Q, S and T waves are calculated.

Section snippets

Wavelet theory

Wavelet transform is a linear transform, which decomposes a signal into components that appears at different scales (or resolution). It is a decomposition of signal using a combination of a set of basis functions, obtained by means of dialation (scaling) and translation of a single prototype wavelet. The greater the scale factor, wider is the basis function and consequently, the corresponding components give the low frequency component of the signal and vice versa. In this way the temporal

Testing and result

The algorithm is validated with arbitrarily chosen ECG data from Physikalisch-Technische Bundesanstalt diagnostic ECG database (ptb-db) of Physionet [23], which contains 549 records from 290 subjects with 52 healthy controls and 148 Myocardiac infarction patients. This algorithm is tested on 12-lead ECG records, each of 10 min duration. Since ptb-db database contains 12-lead signal, it corresponds to 12 different patterns, each having different characteristic features. The R peak detection

Discussion

On decomposing the signal till level 10, the last level, A10, is observed to have baseline drift as shown in Fig. 2 and Fig. 5C. And the baseline drift correction is done by reconstruction of the signal without A10.

The algorithm employs use of some threshold values, for the purpose of peaks, QRS onset and offset detection. These values are empirically set after testing the algorithm on several lead records and calculating the actual possibility of the corresponding limit.

Conclusion

The proposed method uses multiresolution feature extraction using DWT, which can be used for feature extraction from ECG data. Additional features like T wave direction, pathological Q, etc. can be used for characterization of the ECG wave. This paper also explores multiresolution analysis for identification of various frequencies present in an ECG signal. The noise frequencies are also identified and eliminated. The proposed method yields a sensitivity of 99.84% and Predictivity of 99.92% with

References (35)

  • C. Saritha et al.

    ECG analysis using wavelets

    Bulg. J. Phys.

    (2008)
  • J. Pan et al.

    A real time QRS detection algorithm

    IEEE Trans. Biomed. Eng.

    (1985)
  • O. Pahlm et al.

    Software QRS detection in ambulatory monitoring – a review

    Med. Biol. Eng. Comput.

    (1984)
  • G.M. Friesen et al.

    A comparison of the noise sensitivity of nine QRS detection algorithms

    IEEE Trans. Biomed. Eng.

    (1990)
  • I. Murthy et al.

    Analysis of ECG from pole-zero models

    IEEE Trans. Biomed. Eng.

    (1992)
  • J. Vila et al.

    A new approach for TU complex characterization

    IEEE Trans. Biomed. Eng.

    (2000)
  • A.S.M. Koeleman et al.

    Beat-to-beat interval measurement in the electrocardiogram

    Med. Biol. Eng. Comput.

    (1985)
  • Cited by (146)

    • Capsule network assisted electrocardiogram classification model for smart healthcare

      2022, Biocybernetics and Biomedical Engineering
      Citation Excerpt :

      In SVEB detection, only 200, 202, 210 and 213 are slightly better, and the average F1 score of our model is slightly lower than Wu [41]. Because Wu [41] uses the K-Means algorithm for each record when making “common datasets,” the SVEB beat is selected based on the clustering results. This method contains more beat types to improve SVEB classification performance than our random selection method.

    • ECG Signal Denoising Using an Improved Hybrid DWT-ADTF Approach

      2024, Cardiovascular Engineering and Technology
    View all citing articles on Scopus
    View full text