Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 7, No.

2, July 2012

An online ECG QRS Detection Technique

Department of Applied Physics, University of Calcutta, Calcutta, India 1 Email: rgaphy@caluniv.ac.in, mmitra@ieee.org 2 Camellia School of Engineering and Technology, Calcutta, India 2 Email: hk_chatterjee123@yahoo.co.in pass and high pass filers, each operation being carried out separately, while some other use the high pass only [1]-[2]. The filtered signal is compared with a threshold for QRS detection. Digital differentiation based methods exploit the fact that QRS wave is having the maximum slope change in the samples. Most of the real time QRS detectors use a differentiator at initial stage of processing. A DSP based differentiator followed by squaring method is reported [3]. Differentiation and an adaptive threshold can be used on the differentiated signal to identify the QRS regions [4]. Digital differentiation technique followed by a Support Vector Machine (SVM) classifier to identify the QRS achieved 99.75% accuracy [5]. Pan and Tompkins reported a processor based real-time QRS identification mechanism which used digital analysis of slope, amplitude and width of QRS complexes after pre-processing [6]. Later, it was modified by a introducing an adaptive threshold [7]. A comparative study on different first derivative based QRS detection approaches is reported [8]. Algorithms involving advanced computational techniques for QRS detection use Hilbert Transform, Neural Networks, Wavelets, and Hidden Markov Models [9]-[13]. In real time acquisition applications, one or two ECG leads are used. Microcomputer based systems are now increasingly being used for QRS detection in ambulation appliances [14]-[15], Holters and bedside monitors in intensive care units (ICU). In one such application [16], self adaptive threshold based QRS detection is described. In another application [17], a combination of differential threshold and amplitude detection is used for localization of QRS regions quickly. A morphological operator based peak value extractor is used for QRS detection [18]. Some other approaches include zero crossing counts and syntactic methods [19]-[20]. In recent years there has been considerable use of embedded systems for ECG denoising [21] and QRS detection [22]-[23]. In [22] a decision rule to analyze quadratic spline wavelet coefficients to detect the QRS is described. In [23] a QRS detection algorithm is described which calculates the threshold of the next peak detection cycles from the median of eight previously detected peaks. Adaptive lifting scheme, an advanced wavelet tool is used for QRS detection in [24]. The method shows lower execution time compared to DSP platform with TMS320VC5509A, both run in simulation platform. The evaluation criteria used in the QRS detection algorithm are Sensitivity (Se) and Positive Predictivity (P+), defined as:
Se

Abstract- A simple technique, based on measuring the amplitude-span and slope of QRS in Electrocardiograph (ECG) data is described in this paper. Detection of QRS is done in two phases, viz., training and detection. At first, the dataset is searched by sliding a window of 96 ms width, from which an amplitude and slope threshold criteria is learned. From these criteria a QRS template is defined in terms of some signatures. In the detection phase, a QRS is located by matching a QRS template in a sliding 96 ms window. MATLAB simulation, using proposed technique with ECG data from Physionet yields over 98% accuracy. The algorithm is implemented on a standalone embedded system using 8051 microcontroller, which processes 10-12 beats stored in the external on-board RAM. The technique is suitable for online computation of heart rate. Index Terms ECG, QRS, Embedded system.

I. INTRODUCTION An Electrocardiogram (ECG) represents small electrical potential generated by heart muscles. A typical ECG wave includes repeated sequences of P, QRS and T waves associated with each beat, which represent polarization of the atria and the ventricles in a sequential manner. The clinical bandwidth used for recording the standard 12-lead ECG is 0.05-100 Hz. QRS energy is centered on 15 Hz. 60 to 100 bits per minute constitute standard heart rate. Heart rate slower than this leads to Bradycardia and greater than this leads to Tachycardia. Till date, many computer based algorithms are available for automated analysis of ECG samples for assisted diagnosis. QRS complex, representing the ventricular depolarization of the heart, is the most prominent segment of the ECG wave. Many of the ECG analysis algorithms use QRS finding as their starting point due to its characteristic shape. Heart rate variability is another important parameter that can be directly analyzed from prolonged QRS recording. QRS detection is a challenge not only due to variability of the QRS wave shape but also influence of different artifacts. Over the last few decades, many methods have been prescribed for QRS identification. Most of the QRS detectors use a pre-processor to enhance the QRS complex and suppressing the other waves of ECG. This is followed by a decision rule that identifies and often characterizes the QRS. Early approaches for QRS detection used simple derivative based methods and digital filters. Since a typical QRS wave frequency range is between 10-25 Hz, a simple digital filter with cut-off frequencies at extrema of the interval would attenuate other wave components and some artifacts. This band pass operation is implemented in combination of a low 2012 ACEEE DOI: 01.IJRTET.7.2.1 1

TP TP FN

(1)

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 7, No. 2, July 2012
TP

threshold was used as, (2) x (i ) x (i 1) threshold (5) In the training period each sample was checked for validation using equation (5) and were discarded in case if violation. Hence at the end of training period, the following parameters were obtained.
( ampl ) max max. aplitude span in 96 ms window ( slp ) max max. 8-point average slope

TP FP where, TP (True-Positive) stands for correctly found R-peaks, FN (False-Negative) for missed R-peaks and FP (False -Positive) stands for the number of misdetections. This paper describes a QRS detection algorithm based on amplitude span and slope of the QRS. Simulation has been performed with over 400 single lead normal and abnormal records from PTB diagnostic ECG database (ptb-db) and 20 records from MIT-BIH arrhythmia database (mit-db) under Physionet. Finally the algorithm was implemented on 8051 microcontroller (MCU) based standalone embedded system.

(6) (7)

II.

MATERIALS AND METHODS

A typical ECG beat comprises of some equipotential segments connected by wave peaks, among which QRS is with highest slope and normally, the tallest. The QRS morphology varies among different age groups, communities and subcontinents. The proposed QRS detection technique exploits the following QRS signatures, viz., (i) the amplitude span, which is R to S (or Q) height remains almost constant throughout a particular lead data and (ii) QRS region is having the maximum slope, (actually, inter-sample difference), (iii) the width of QRS segment is nominally 96 ms, within which, Q (or S) to R distance is maximum 60 ms. Hence, to implement a technique which relies only such signatures essentially require training with data before the actual QRS detection is performed. In the proposed algorithm, two such signatures were extracted in the training period from the acquired (and buffered) data, typically from 8-10 beats. These signatures were used to define a QRS template, which was then used for exact localization of the QRS peaks in the entire data. Thus the entire operation can be divided into two parts, viz., training period and detection period. During the training period, a sliding window of 96 ms width was moved by 20 ms in steps through the entire dataset to capture these two signatures. The objective was to capture a QRS zone which could be defined using the following parameters. The sliding shift 20 ms (empirically determined) was taken to ensure capture of QR or QS peaks in a single window. The (i) Maximum amplitude span; (ii) Maximum positive slope averaged over 8 points, The slope at a sample i was computed as: slp (i ) x (i 1) x (i ) (3) To minimize the high frequency noise, 8-point average slope was used and calculated as,
slp (i )
i x (i 8) x (i ) 8

The following rules for the detection of probable QRS neighbourhood are formulated as: (a) Amplitude span in moving 96 ms window should exceed 80% of maximum amplitude span computed over the entire dataset. This was named as amplitude threshold criteria (ATC) and the value could be denoted by (ampl)th. Hence, ATC: ampl(k)96 > (ampl)th (8) where, ampl(k)96 is amplitude span over 96 samples for kth position and (ampl)th = 0.8 X (ampl)max (b) Average (8-point) slope should exceed 80% of maximum slope computed over the dataset. This was named as slope threshold criteria (STC) and the value could be denoted by (slp)th. Hence, STC: slp(k)96 > (slp)th (9) where, slp(k)96 is slope at kth position over 96 ms window and (slp)th = 0.8 X (slp)max In the detection period, a new search was initiated in a 96 ms window with a sliding step of 20 ms to get a match of ATC. If for a particular window in the detection phase, ATC was found to be satisfied, a probable QRS neighborhood was assumed. Figure 1 shows different types QRS templates with respective amplitude span match with a moving 96 ms window (shown in green dashed line). It is also imperative that to meet with the ATC criteria, this window must contain either R or Q or S peak as its local maxima or minima. However, tall T peak for some abnormal ECG record may also meet the ATC. Now, at this position of the window, at first, any of the R or Q or S peak is to be determined by a slope reversal search initiated from the local maxima or minima. This is shown in Figure 2.

(4)

Since non-QRS regions of the ECG wave used to be more flat and with lower amplitude span (except for tall T waves) Figure 1. Different QRS forms captured by sliding 96 ms window and threshold, the sliding window had been used to If pmax and pmin are the indexes of maximum and minimum points continuously compute the amplitude span and average slope in a 96 ms window, where the ATC criteria were found to met to store the tallest and sharpest QRS. To minimize high with. Then average slope over 8 samples (i.e., 16-point aver frequency spike at sample i, an empirically determined 2 2012 ACEEE DOI: 01.IJRTET.7.2.1

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 7, No. 2, July 2012 age slope), at 20 samples upside and 18 samples downside w.r.t pmax and pmin was calculated. Computed values could be denoted as: slp_r20- : 20 sample downside average slope w.r.t pmax slp _r18+ : 18 sample upside average slope w.r.t pmax slp_q20- : 20 sample downside average slope w.r.t pmin and slp_q18+ : 18 sample upside average slope w.r.t pmin Now, two conditions may arise as follows: Condition 1: slp_r20- > 0 and slp _r18+ < 0; (10) we may conclude that there is a slope sign reversal around p max and, condition 2: slp_q20- < 0 and slp _q18+ > 0; (11) Both conditions (1) and (2) may also get simultaneously satisfied. But it does not ensure a location of R-peak, since a positive or negative T-peak may also satisfy conditions (1) and (2). To ascertain the position of QRS, STC criteria were used in the following way: For condition (1) [or (2)] being satisfied, a new window of 96 ms was formed around pmax (or pmin), keeping that at the midpoint. From the starting point of the new window formed, for each sample (say index n) 8point average slope was computed at n and (n+25) index positions, with an objective to locate the position of upside and downside of positive R-wave. The following conditions are tested: Condition 3: (slp)n >= slpth (12) Condition 4: (slp)n+25 < 0 (13) where, (slp)n is slope at nth index etc. Condition 3, actually being the STC, was used to reject out any falsely detected probable T-peak, because QRS complex used to be sharper than the T wave. If condition (3) was satisfied, then condition (4) was searched to find a positive R peak. In case conditions (3) and (4) are found to be satisfied, existence of a positive R peak was concluded. It was also verified by a slope direction reversal within next 70 ms along the array, confirming occurrence of an S peak. The correct position of R-peak was determined by finding the local maximum within next 50 samples from sample n. For QS peaks (i.e., no positive R peak), condition (3) has to be satisfied only, not the condition (4). In that case, it has to be concluded that a positive R-peak does not exist and search for QS peak has to be initiated. The local minimum within the current window (of 96 ms width) has to be taken as the QS peak. Once the R-peak (or QS peak) was determined, the next QRS search was initiated at 400 samples upside the data array, based on the assumption that the minimum R-R interval is 400 ms. For ECG data with tall T peaks, equation (10) and (11) may get initially satisfied, but condition (3) will not match within 96 ms window. Hence, in that case the next search have to be initiated from 200 samples upside the data array. Due to noise problem one of more R-peak may have been missed, resulted into erroneous R-R interval calculation. In such a case the average R-R interval was computed as follows: If the following represent detected R-peak indexes: 516, 1510, 2517, 4516, 5517, 6517, 8512, 9517, 12517, 16512, Then at first, all the R-peak indexes were taken in an array. All the R-R intervals were computed and a new array was 2012 ACEEE DOI: 01.IJRTET.7.2.1 3 generated using following formula: rri = ri+1 - ri for i =1 to n-1, where n being the number of peaks determined. From this array the minimum value, say rr m was taken, and compared with other elements of this array to determine missed R peaks. If rri exceeds rrm by 80, occurrence of some missed R-peak was concluded. The corresponding positions of R were discarded and only the valid R-R intervals were taken for calculation of average R-R interval. The algorithm was implemented with a standalone embedded system based on 8051 MCU. The system was tested with single lead synthetic ECG data, each consisting of 10000 samples. The digitized samples at 8 bit resolution were generated in a desktop PC and were delivered through the serial port for serial storage in an on-board RAM. Then standalone system performed the QRS detection task on these pre-stored data. III. TESTING AND RESULTS The algorithm was initially validated with MATLAB using synthetic ECG data from Physionet [25]. Physionet offers digital recordings of different physiological signals and related data for use of biomedical research. PTB Diagnostic ECG database (ptb-db) is available in Physiobank under Physionet.

Figure 2. Determination of local maxima or minima in 96 ms window as R or Q peak

Ptb-db database contains 549 high-resolution 15-lead ECGs (12 standard leads together with Frank XYZ leads) records and clinical summaries for each record. From one to five ECG records are available for each of the 294 subjects at 16 bit resolution and 1 kHz sampling. The R-peak detection algorithm was applied on different single lead ECG data, each containing 30000 samples, which were normalized in 8 bit resolution level. Table 1 shows some of the Sensitivity (Se) and Positive Predictivity (P+) figures for R peak detection, expressed in percentage scale, out of total 400 different normal and abnormal leads tested. An average sensitivity figure of 98.12% and Predictivity figure of 99.76% was obtained. Figure 5(a)-(c) represent a graphical representation of detected R-peaks (indicated by red vertical lines) in the quantized

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 7, No. 2, July 2012
TABLE I. SENSITIVITY AND POSITIVE PREDICTIVITY WITH PTB-DB ( IN PERCENTAGE)

(Narration: N Normal, M: Myocarditis, MI-Myocardial Infarction, MI-Inf- MI Inferior, MI-Ant- MI Anterior, Inf Lat- Infero Latera, AntLat- Antero-lateral)

data plot for some normal and abnormal ECG data from ptbdb in MATLAB simulation. The algorithm was also validated with MIT-BIH arrhythmia data files (mit-db) under Physionet. The mit-db files contain 48 half-hour excerpts of two-channel ambulatory ECG recordings at 360 Hz sampling, obtained from 47 subjects.
TABLE II. SENSITIVITY AND POSITIVE PREDICTIVITY WITH MIT-DB ( IN PERCENTAGE)

Since the developed algorithm was applicable for ECG records at 1 kHz, an interpolation technique was used to upsample the mit-db data to a 1 kHz sampling level. 50000 samples from a single lead are used for validation of the algorithm. Table 2 shows the Sensitivity (Se) and Positive Predictivity (P+) figures for mit-db data. An average sensitivity of 94% and predictivity of 100% was obtained with 20 different records. Figure 5(d) represent R-peak detection using a single lead of arrhythmia data under mitdb, containing 50000 samples in MATLAB. The detected Rpeaks are indicated by red vertical lines. As evident from the plots shown in Figure 5, the algorithm is immune from baseline modulation, if any, in the ECG data. Testing of the algorithm in embedded system platform was achieved in two stages. A. Normalization and delivery of ECG data to the standalone system A separate application software was developed which, at first digitized the single lead millivolt level ECG samples as obtained from Physionet database, with 8 bit resolution and have saved in a text file. The second function was to deliver these quantized samples to the embedded system through the serial port, using an event driven programming technique. A typical serial port session [26] involving data output 2012 ACEEE DOI: 01.IJRTET.7.2.1 4

through the serial port involves the following steps: i) Fixation of initial communication settings like baud rate, number data bits, parity and stop bit etc ii) From computer memory, data byes are to be sent to the output buffer of the USART (of the serial port) by using commands in MATLAB. iii) The USART have to transmit continuously the data bytes to the external device using the TxD line, which is connected to the RxD pin of the MCU. iv) An OutputEmpty event is generated when the output buffer goes empty, triggering the corresponding event described by BytesAvailableFcn and OutputEmptyFcn. The calling function loads the next block of data to the USART for delivery to the external device. The steps (ii) to (iv) were repeated till all the ECG data were delivered to the standalone embedded system. A GUI based front end was provided to perform the entire operation, i.e., choosing the ptb-file and lead for quantization, and then delivering the serial data to the standalone system. After all samples have been delivered, the software have to wait for the QRS locations to be sent back from the standalone embedded system to the PC using serial port for automatic storage in a text file. B. QRS Detection from digitized samples The standalone embedded system was consisted of an 8051 microcontroller, on-board RAM and MAX-232 level converter as the principal components. Figure 3 shows block diagram of the testing hardware. At first, 10000 samples were stored in the temporary RAM.

Figure 3. Block diagram of the developed system

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 7, No. 2, July 2012 ACKNOWLEDGMENT The authors acknowledge the technical support received from the UGC SAP DRS-I project at Department of Applied Physics, University of Calcutta, India. REFERENCES
[1] M. L. Ahlstrom and W. J. Tompkins, Automated high-speed analysis of holter tapes with microcomputers, IEEE transaction on Biomedical Engineering, vol. 30 (10), pp. 651657, 1983. [2] M. Okada, A digital filter for the QRS complex detection, IEEE transaction on Biomedical Engineering, vol. 26 (12), pp. 700-703, 1979. [3] V. Arora, R. Chugh, A. Gagneja and A. Pujari, Digital ECG and its analysis, in: SPIT-IEEE Colloquium 2007-2008, Mumbai, India, pp.1-7, 2008. [4] J. Fraden and M. R. Neuman, QRS wave detection, Medical and Biological Engineering and Computing, vol. 18 (12), pp. 125-132, 1985. [5] S. S. Mehta and N. S. Lingayat, Identification of QRS complexes in 12-lead electrocardiogram, Expert Systems with Applications, vol. 36 (1), pp. 820-828, 2009. [6] J. Pan and W. J. Tompkins, A real-time QRS detection algorithm, IEEE transaction in Biomedical Engineering, vol. 32 (3), pp. 231-236, 1983. [7] P. S. Hamilton and W. J. Tompkins, Quantitative investigations of QRS detection rules using the MIT-BIH arrhythmia database, IEEE transaction on Biomedical Engineering, vol. 33 (12), pp. 1157-1165, 1986. [8] N. M. Arzeno, Z. D. Deng and C. S. Poon, Analysis of firstderivative based QRS detection algorithms, IEEE transaction on Biomedical Engineering, vol. 55 (2), pp. 478-484, 2008. [9] V. Di-Virgilio, C. Francaiancia, S. Lino and S. Cerutti, ECG fiducial points detection through Wavelet Transform, in: 17th IEEE Annual Conference of Engineering in Medicine and Biology Society, Montreal, Quebec, Canada, Vol. 2, pp. 10511052, 1995. [10] X. Xu and Y. Liu, Adaptive Threshold for QRS complex detection based on Wavelet Transform, in: 27 th Annual International Conference of the Engineering in Medicine and Biology Society, IEEE-EMBS 2005, Shanghai, China, pp. 7281-7284, 2005. [11] Y. H. Hu, W. J. Tompkins, J. L. Urrusti and V. X. Afonso, Applications of artificial neural networks for ECG signal detection and classification, Electrocardiology, vol. 26 (suppl) pp. 66-73, 1993. [12] D. A. Coast, R. M. Stern, G. G. Cano and S. A. Briller, An approach to cardiac arrhythmia analysis using hidden markov models, IEEE transaction on Biomedical Engineering, vol. 37 (9), pp. 826-836, 1990. [13] D. S. Benitez, P. A. Gaydecki, A. Zaidi and A. P. Fitzpatrick, A new QRS detection algorithm based on the Hilbert Transform, in: Computers in Cardiology, Cambridge, USA, pp. 379-382, 2000. [14] M. Adnane, Z. Jiang and S. Choi, Development of QRS detection algorithm designed for wearable cardiorespiratory system, Computer Methods and Programs in Biomedicine, vol. 93 (1), pp. 20-31, 2009. [15] F. Zhang, J. Tan and Y. Lian, An effective QRS detection algorithm for wearable ECG in body area network, in: Biomedical Circuits and Systems Conference (BIOCAS 2007), Montreal, Quebec, Canada, pp. 195-198, 2007.

Figure 4. Hardware layout of the system

The data set was searched and each detected R- peak index was stored in two consecutive locations of the 128 byte internal RAM of the MCU. The stored indexes were delivered back to the PC using RS-232 communication. The R-peak indices were captured in a separate data file for verification with the actual R-peaks of the corresponding lead data. A snap shot of test layout of the system with detection of R-peaks indicated by the alteration of a signal between two different logic levels has been represented in Figure 4. CONCLUSIONS A QRS detection technique based on amplitude span and slope signature of QRS has been described. One advantage of the proposed algorithm is that it required the processing of a fewer samples at a time, hence utilizing minimum internal resources of processor. The method has shown fairly good results with noisy samples. Thus, no preprocessing of data was required. For online applications involving large amount of data (70-100 beats), a trade-off can be done between computational time and accuracy of detection. Here, training period can be confined to 10 beats, so as to reduce the overall computational time and latency of detection. However, training with more data would increase the accuracy. The simplicity of the algorithm provides its suitability to implement it in a low level microcontroller like 8051. In its present form the developed system accepts 8 bit digitized ECG data from serial port. However, it can be modified to interface directly with a standard ECG lead system with amplifier module to directly acquire live ECG from human patients. The standalone system can be used in a primary healthcare unit for heart rate computation of patients during initial check up. The existing system can be upgraded to a real-time QRS detector using a modification of the algorithm by using a 96 sample first-in-first-out stack. The first 2000 incoming data samples (expected to contain at least one QRS region) can be used for learning phase and the next incoming samples can be treated as detection phase. Additionally, this modified technique would not require an external RAM. 2012 ACEEE DOI: 01.IJRTET.7.2.1 5

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 7, No. 2, July 2012

[16] Z. H. Ying and H. K. Mean, Embedded real time QRS detection algorithm for pervasive cardiac care system, in: 9 th International Conference on Signal processing (ICSP), Beijing, China, pp.1-4, 2008 . [17] G. Zhengzhong, K. Fanxue and Z. Xu, Accurate and rapid QRS detection for intelligent ECG monitor, in: 3 rd International Conference on Measuring Technology and Mechatronics (ICMTMA), China, vol. 3, pp.298-301, 2011. [18] P. E. Trahanias, An approach to QRS detection using mathematical morphology, IEEE transaction in Biomedical Engineering, vol. 40 (2), pp. 201-205, 1993. [19] B. U. Kohler, C. Hennig and R. Orglmeister, QRS detection using zero crossing counts, Progress in Biomedical Research, vol. 8 (3), pp. 138-145, 2003. [20] J. K. Udupa and I. S. N. Murty, Syntactic approach for ECG rhythm analysis, IEEE transaction in Biomedical Engineering, vol. 27 (7), pp. 370-375, 1980. [21] R. Ramos, A. M. Lazaro, J. D. Rio and G. Oliver, FPGAbased implementation of an adaptive canceller for 50/60m-Hz

interference in Electrocardiography, IEEE transaction on Instrumentation and Measurement vol. 56 (6) pp. 2633-2740, 2007. [22] C. I. Ieong, M. I. Vai, and U. P. Mak, QRS recognition with programmable hardware, in: 2 nd Annual Conference on Bioinformatics and Biomedical Engineering, Shanghai, pp.2028-2031, 2008. [23] A. Shukla and L. Macchiarulo, A fast and accurate FPGA based QRS detection system, in: 30 th annual IEEE International Conference on Engineering in Medicine and Biology (EMBS 2008), Vancouver, Canada, pp. 4828-4831, 2008. [24] Y. Li, H. Yu, L. Jiang, L. Ma and Z. Li, Adaptive Lifting Scheme for ECG QRS complex detection and its FPGA implementation, in: 3 rd International Conference on Biomedical Engineering and Informatics (BMEI), Yantai, China, vol. 2, pp.721-724, 2010. [25] http://www.physionet.org/ <accessed 19.04.2011> [26] MATLAB documentation file apiext.pdf from http:// www.mathworks.com/

2012 ACEEE DOI: 01.IJRTET.7.2.1