A QRS estimator using linear prediction approach
Introduction
The electrocardiographic (ECG) signals recorded from the body surface are often used to diagnose abnormalities in the heart's conduction system, which either manifests in the form of rhythm changes or gradual variations in the morphological features. The ECG signals are constituted by a sequence of component-waves separated by regions of zero electrical activity, called iso-electric regions. Under normal conditions, the component-waves repeat themselves in a rhythmic manner with a periodicity determined by the frequency of impulse generation at the sino-atrial (SA) node. The effect of a single impulse excitation as reflected on the body surface is depicted in the ECG cycle shown in Fig. 1. The time evolution of the surface ECG reflects the electro-physiological effects brought about by the propagation of a current-wave resulting from the periodic impulse excitation at the SA node along well defined conduction pathways laid out in the cardiac musculature. The most important of these effects is the concomitant contraction and relaxation of the atrial and ventricular muscles. As shown in Fig. 1, a single ECG beat is made up of three distinct component-waves designated as P-, QRS- and T-waves, respectively. The P-wave characterises the depolarisation of the atrial musculature leading to atrial contraction and occurs at the final third of ventricular filling [5]. The QRS complex corresponds to the ventricular depolarisation followed by the occurrence of the T-wave during ventricular repolarisation, resulting in the consequent relaxation of the ventricular muscles. The interval between the offset point of the QRS complex (called the J-point) and the onset of the following T-wave (Ton) is called the ST-interval. A precise estimation of the ST-segment waveform is often required for the differential diagnosis of Coronary artery disease (CAD) [9]. The extraction of the ST-segment features necessitates a procedure for blanking out the P- and T-waves with an effort to delineate the QRS pulses from the surface ECG.
Over the past two decades, much of the effort, both in hardware and software, has understandably been concentrated on the detection of QRS complexes alone 6, 8, 11. However, the problem of accurately delineating the QRS pulse shape remained elusive until Murthy et al. proposed a rational function representation of the ECG signal in the DCT domain and used a pole-zero model to carry out the delineation of component-waves [7]. Murthy's work brings out the inherent difficulties involved in a direct time domain parametric modelling of the ECG. The advantage gained by performing a DCT is that, the higher order energy coefficients are brought closer to the origin and consequently, the time domain signal gets mapped into a decaying set of coefficients. As a result of this transformation, the DCT of an individual component-wave is approximated by a set of damped cosinusoids with closely spaced frequencies. Each damped cosinusoid, referred to as a fractional component is approximated by a biphasic rational function with second order numerator and denominator polynomials. The delineation of component-waves is accomplished by decomposing the rational function into partial fractions with second order numerator and denominator coefficients. The pole angles in each partial fraction expression is shown to have a one to one relationship with the peak location of each one of the component-waves present. The desired set of component-waves are then reconstructed in the DCT domain by grouping together the subset of those partial fraction expressions which contain poles with their corresponding angles related to the peak locations of the component-waves to be separated out. A major drawback of this approach is that the selective separation of the component-waves require a prior knowledge of their respective temporal locations in the ECG record. Moreover, the subset of parameters needed for the reconstruction is chosen from a partial fraction expansion of the rational function model. This involves lengthy computations when a large number of ECG beats are involved. In view of these difficulties, we now propose a QRS delineation method which works independently on each lead and performs the task of separating out the QRS morphology. The advantage of using the present method lies in the fact that it is well suited not only for noisy ECGs but also for those containing ectopic and missed beats. Some of the other important features of the proposed method are that it does not require a prior knowledge of the position of QRS pulses and is capable of providing a more accurate description of the QRS morphology irrespective of its temporal location in the recorded ECG.
Section snippets
Theory
Considering N samples of an ECG record y(n), a sequence of length 2N is constructed with , and zero elsewhere. A 2N-point discrete Fourier transform (DFT) of the sequence is defined asThe sequence obtained by even extending y(n) is denoted by The discrete cosine transform (DCT) of the sequence y(n) is defined as the first N points of the 2N-point DFT of ye(n),
QRS estimator
We refer to the term `QRS subspace' to denote a set of narrow band signals represented by the largest set of eigenvalues λ1>λ2>⋯>λu, with u⩽p and the complementary subspace containing the P and T wave information, as represented by the remaining set of the p−u eigen values. We start with a given prediction matrix X∈RL of size L×N−p (L=p+1) which is full rank and spanned by the columns of any unitary matrix in RL×L, the Euclidean space of square and real valued L dimensional matrices. The same
Summary of the procedure
A closed procedure including all steps beginning with the data record (ECG) with N samples and ending with the reconstructed QRS cycles is as follows:
- 1.
Find the N-point DCT of the data record in accordance with the definition for DCT as given in Eq. (3).
- 2.
Decide the value of p: As discussed earlier, choosing the best value for p is crucial to achieving the best performance of the estimator. If the choice of p is too low (lower than the number of peaks in the Fourier spectrum of the DCT
Conclusions
It is demonstrated that the component-waves present in the time domain ECG record get mapped into a set of narrow band signals in the DCT domain. The narrow band signals are selectively estimated by performing an SVD operation on a prediction matrix of size as determined by the appropriate number of these signals with which the component-waves are approximated in the DCT domain. Since the time domain features of the QRS components are manifested by pulses having very short duration with sharp
References (12)
- I. Dologlou, G. Carayannis, Physical interpretation of signal reconstruction from reduced rank matrices, IEEE Trans....
- G.H. Golub, C.F. Van Loan, Matrix Computations, Johns Hopkins University Press, Baltimore, MD,...
- R. Kumaresan, D. Tufts, Singular value decomposition and improved frequency estimation using linear prediction, IEEE...
- R. Kumaresan, D. Tufts, Estimating the parameters of exponentially damped sinusoids and pole zero modeling in noise,...
- J.R.Levick, An Introduction to Cardiovascular Physiology, Butterworths, London,...
- et al.
A robust digital QRS detection algorithm for arrhythmia monitoring
Comput. Biomed. Res.
(1983)
Cited by (5)
Cepstral transformation technique for dissociation of wide QRS-type ECG signals using DCT
1999, Signal ProcessingR-Peak detection using wavelet transforms technique
2015, UPB Scientific Bulletin, Series C: Electrical Engineering and Computer ScienceImproved fast filtering algorithm with low distortion for dynamic electrocardiogram
2012, 2012 IEEE 5th International Conference on Advanced Computational Intelligence, ICACI 2012QRS complex detection of ECG signal by using teager energy operator
2008, 2nd International Conference on Bioinformatics and Biomedical Engineering, iCBBE 2008Neural network classifier based on the features of multi-lead ECG
2005, Lecture Notes in Computer Science