Analysis of a Differential Noise Detection Filter in …docs.mipro-proceedings.com/dc/23_dc_5477.pdfAnalysis of a Differential Noise Detection Filter in ECG Signals Era Ajdaraga ,

Analysis of a Differential Noise Detection Filter inECG Signals

Era Ajdaraga∗, Marjan Gusev†

University of Ss.Cyril and Methodius, Skopje, Macedonia,Faculty of Computer Science and EngineeringEmail: ∗[email protected],

†[email protected]

Abstract—Noise detection presents a big challenge in wearableheart monitor technology and beat detection. The distorted signalis hard to interpret and as a consequence, valuable informationmay be lost. In this paper, we present our research for developinga noise detection filter based on a differential filter. Our analysisoffers in-depth evaluation of the optimal values of two criticalparameters - window size and Signal-to-Noise threshold. Ourfinal goal is to minimize errors in QRS detection and beatclassification caused by noise-distorted data.

Index Terms—Noise Detection, ECG, Telemedicine

I. INTRODUCTION

Wearable health monitors are quickly becoming ubiquitousand as a result there is a rapidly growing amount of data thatis digitally processed and analyzed. The integrity of this datais of utmost importance, especially when dealing with highlysensitive information such as health data.

ECGalert [1] is an instance of a wearable heart monitorthat transfers health data to the cloud for further processing.A wearable sensor samples the heart activity, while the datais collected by the user’s mobile phone over Bluetooth. Oncethe user connects to the Internet and the recording has beencompleted, it is automatically uploaded to the cloud where thedata is subject to detailed analysis. The ECGalert system cancorrectly classify several types of beats, such as• Normal (N),• Ventricular (V),• Supraventricular (S), and• Fusion of Ventricular and normal beats (F);

as well as the following types of rhythm episodes:• Bigeminy,• Trigeminy,• Ventricular ectopies,• Supraventricular (atrial) ectopies,• Atrial fibrillation and flutter,• Ventricular tachycardia and fibrillation.ECGalert uses an improved version of the Hamilton QRS

detection algorithm [2] for its analysis. We have found thatcorrect detection and classification depend on the quality of therecording. Noise may cause valuable information (containedwithin the signal) to be lost. This may result in false identi-fication of a QRS complex, or even missing a QRS complexdetection, which ultimately leads to false diagnosis. Therefore,

noise detection and identification of uninterpretable segmentsis the first step to eliminating errors caused by noise.

Unreadable, or uninterpretable, in this sense, means that thenoise level is so high that efficient filtering techniques usedin QRS detection algorithm can not perform well and revealsufficiently good results, where the sensitivity and positivepredictivity of QRS detection is higher than 99.8%.

Noise is usually classified as:• high frequency noise, usually caused by the environment

and nearby electro-magnetic field, and• low frequency noise, usually caused by muscles and body

movement.Elimination of very low frequency noise can be efficiently

realized by a high pass filter with cut-off frequency of 0.5 or1 Hz. This will eliminate the baseline drift that usually occursin the signal.

Elimination of high frequency noise that occurs due to theelectricity and other electro-magnetic fields can be efficientlyrealized by a low pass filter with cut-off frequency of 30 Hzor above.

The spectrum of an ECG signal, where the QRS complexesare usually lays between 3-15 Hz [3] and a deeper analysisshows which filters can be used for noise elimination [4]. Thebest results are obtained by implementing wavelet filteringmethods [5]

The real problem can be identified by analyzing the typicalpower spectra of an ECG signal, as presented in [3]. Elimi-nating the very low and very high frequencies by a bandpassfilter [4] will still reveal a signal where the noise corruptsthe original ECG signal. In fact almost all QRS detectionalgorithms work on detecting the high slope of a QRS complex[6], [7], which is always corrupted by noise which appearsbetween 3-10 Hz.

Therefore, the real problem is to identify if the noise level ishigh enough that it will prevent the correct detection of QRScomplexes. A differential filter is proposed in this research toanalyze the level of noise in the signal. The problem is that wedo not know if the signal contains QRS complexes, and we donot know their position. The detection method uses analysisof a given signal segment and the energy of signal level withinthe signal segment. The QRS complex itself generates a highlevel of energy, and we have to conclude if there is noise or

314 MIPRO 2019/DC

not. If the analyzed segment is small enough, then we facethe problem that there are none, one or two QRS complexes.To solve this, the segment should be large enough to containseveral QRS complexes. Analyzing a segment that is too largewill lose the efficiency to detect the interval where noise exists.

Accuracy and reliability are determining factors in theusability of a health-monitoring product and this proves to beour motivation. In this paper, we present our findings regardingthe implementation of a differential filter, as well as a thoroughanalysis of the optimal values for window size and Signal-to-Noise threshold, that will give a clear indication that the signalcan be efficiently processed.

The structure of this paper is as follows. Related work ispresented in Section II and our methodology is explainedin Section III. Section IV presents the results. Finally, adiscussion is given in Section V.

II. RELATED WORK

Velayudhan et al. [8] present a comprehensive summaryof the various noise-detection and denoising techniques thatcan be applied to an ECG signal. They find that the adaptivefiltering technique outperforms the other approaches when thesignal has a low SNR, however the filtered residue techniqueis superior when working with moderate SNR.

Jekova et al. [9] analyze the performance of five differentSR/VF detection algorithms to determine the ones that performbest with noisy data. They concluded that TCIs, VF filter andspectral analysis were the most appropriate algorithms. Basedon the presented data it can be concluded that the VF algorithmhad the best results in terms of performance stability.

Sameni et al. [10] propose a novel framework for achievinghigh-resolution ECG by implementing the Bayesian approach.In this framework they review the performance of conven-tional filtering approaches where they conclude that UnscentedKalman Filter basend on the Unscented Transform is the mostrobust approach to severe nonlinearities of the input timeseries.

Jekova et al. [11] present a thorough analysis of the mostcommon noise types present in ECG recordings, as well as anovel approach to evaluating the level of noise corruption inECG data. Their proposed solution incorporates six parallelnoise tests which aim to assess the corruption level of specificnoises. This approach uses 13 parameters to determine thepresence of noise.

Lee et al. [12] propose a real-time approach for the detectionof artifacts caused by motion and noise (MN). Their algorithmworks in two stages. Firstly, empirical mode decompositionof the signal is performed, and then they attempt to identifynoisy segments in the resulting high-pass-filtered signal byevaluating the randomness of the calculated values. The resultis a threshold for the separation between clean and MN-corrupted data.

Determining the level of SNR in an analyzed ECG signalpresents another challenge in the interpretation of ECG record-ings.

III. METHODOLOGY

Our work is based on the differential filter, and our goal is toidentify the optimal values for which the algorithm performsbest.

A. Algorithm

The algorithm we propose is a first-difference operator. Wecalculate the difference between two adjacent samples with

y[i] = x[i]− x[i−1], (1)

where y is the resulting vector which contains the differencebetween all adjacent samples in the original vector x. Once allthe values in vector y are calculated the following summationof values is performed

sum =n

∑i=0|y[i]| (2)

This sum is continuously recalculated after the slidingwindow has shifted z amount of times, i.e. z new sampleshave been inserted into the filter.

Experimental analysis allowed us to determine the optimalthreshold which differentiates the original, clean signal fromthe noisy signal.

B. Data

The data we used to conduct this research was obtainedfrom the MIT-BIH Noise Stress Test Database which is freelyavailable on the PhysioNet [13] website. Originally, the datawas digitized at 360 samples per second per channel with 11-bit resolution, however we have resampled it to 125Hz and16-bit resolution.

These records were obtained in such a way that two recordsfound within the MIT-BIH arrhythmia ECG database, 118 and119, were used as a basis where noise of the types typicallyobserved in ECG recordings were added to the files, generatingvariations of the files with different SNRs. The name of therecord means how much noise is added, by identifying thesignal to noise ratio measured in dB. There are 6 recordswithin this database for 118 and 119 labeled by e24, e18,e12, e06, e00 and e 6, correspondingly to SNR levels of 24dB,18dB, 12dB, 06dB, 0dB, and −6dB.

SNR level of 24dB means dominant signal, level of 0dBmeans equal level of noise and signal and level of −6dB meansthat noise is dominant in relation to the signal. Although theoriginal signal was sampled using an 11-bit resolution (valueswithin range 0 to 2047 and zero at 1023), due to high levelof noise, every file was resampled to 16-bit resolution (valuesbetween 0 and 16383, and zero at 8191).

Noise was not added to the whole signal but in segments.The first segment starts at the fifth minute, lasts two minutes,then in alternate segments of 2 minutes without and with noise.

MIPRO 2019/DC 315

C. Evaluation indicators

The following statistical parameters are used to evaluate theperformance of the algorithm: sensitivity, positive predictivevalue and 1/F score.

A careful definition of true positives (TP), false positives(FP), false negatives (FN) and true negatives (TN) is neededto calculate these statistical parameters. In this paper, we definetwo evaluation measures, the first one concerns the ability todetect:• QRS complex,• a high noise level that prevents QRS detection algorithm

to perform satisfactorily.The majority of authors use only the first evaluation measure

to detect a QRS complex, defining TP to be an identified QRScomplex within 150ms from the true QRS complex, FP to be adetected QRS that does not exist, FN to be inability to detect aQRS complex, and TN correctly identifying the non-existenceof a QRS complex. According to the standard 60601-2-47 [14],a QRS is correctly detected if it is within 150ms of the reallocation of the QRS complex.

Sensitivity (SEN), also known as hit rate or true positiverate (TPR), measures the proportion of correctly identifiedpositive cases TP in regards to the actual number of positivecases (altogether those correctly and incorrectly detected QRScomplexes), calculated by (3).

SEN =T P

T P+FN(3)

Positive Predictive Value (PPV), also known as precision,is the proportion of positive results that are true positives inregards to the total number of detected QRS complexes, thatis positive results (both true and false positives) acquired withthe detection method, calculated by (4).

PPV =T P

T P+FP(4)

Sensitivity refers to the ability of the algorithm to detectall existing events, while positive predictive rate means howmuch of the detected events are correct.

Presenting just one of these statistical measures may misleadreal performance. High sensitivity means that the algorithmwill reveal all existing events (detecting all relevant QRScomplexes) but at the same time it may also reveal a lot ofmisses (detecting false positives, QRS complexes that do notexist). This is why the Fscore as a harmonic mean of boththese values will reveal a more comprehensive and accuratestatistical measure to evaluate the performance. (5) presentsthe way it is calculated.

1F score

=1

SEN+

1PPV

(5)

In this research we use an improved version of the Hamil-ton’s QRS detector [7], which performs better than mostof the existing QRS detectors [15] achieving QRS detec-tion sensitivity of 99.1% and positive predictive value of99.9%. This algorithm has shown the following performance

TABLE IPERFORMANCE OF THE IMPROVED QRS DETECTION ALGORITHM TESTED

ON THE MIT-BIH NOISE STRESS TESTING DATABASE

Record SEN PPV 1/F score118 100.00% 99.91% 99.95%

118e24 100.00% 99.87% 99.93%118e18 99.47% 97.42% 98.43%118e12 98.02% 91.40% 94.59%118e06 95.87% 82.85% 88.89%118e00 87.18% 76.47% 81.47%118e 6 84.55% 69.88% 76.25%

119 100.00% 99.95% 99.97%119e24 99.90% 99.90% 99.90%119e18 99.80% 96.59% 98.17%119e12 99.30% 87.92% 93.26%119e06 96.48% 79.21% 87.00%119e00 89.58% 71.72% 79.66%119e 6 83.34% 66.08% 73.71%

50%

60%

70%

80%

90%

100%

118 118e24 118e18 118e12 118e06 118e00 118e_6

Performance

SEN PPV Fscore

Fig. 1. Performance of the QRS detection algorithm evaluated on record 118MIT-BIH noise stress testing database (SNR level indicated after the recordname)

in regards to the MIT-BIH noise stress testing database, andincluding original records from MIT-BIH arrhythmia database,as presented in Table I.

Analyzing the QRS detection algorithm presented in Fig. 1and 2, we conclude that if the level of noise is lower thanSNR of 12dB then the achieved QRS detection performanceis higher than 98%, which is the performance achieved bystate-of-the-art QRS detection algorithms.

Therefore, the new evaluation measure we introduce is theability to detect if the noise level is low enough, such that thesignal to noise ration (SNR) is higher than 12dB and QRSdetection performs well.

In this sense, we calculate the sensitivity, positive predictiverate and 1/F score to detect if the SNR is higher than 12dBin order to enable quality detection of QRS complexes in theECG signal. This will be achieved by calculating:• TP (true positive) for a segment that the tested algorithm

correctly detects a high level of noise,• FP (false positive for a segment for which the tested

algorithm wrongly detects a high level of noise for asignal with a small level of noise,

316 MIPRO 2019/DC

50%

60%

70%

80%

90%

100%

118 118e24 118e18 118e12 118e06 118e00 118e_6

Performance

SEN PPV Fscore

Fig. 2. Performance of the QRS detection algorithm evaluated on record 119MIT-BIH noise stress testing database (SNR level indicated after the recordname)

• FN (false negative) for a segment for which the testedalgorithm does not detect a signal with a high level ofnoise,

• TN (true negative) for a segment that the tested algorithmdoes not detect since the signal does not have a high levelof noise.

The corresponding statistical measures SEN, PPV and 1/Fscore are calculated by the same equations (3), (4) and (5)correspondingly.

The results will be evaluated according to the calculation ofthe number of episodes and duration of segments with highlevels of noise. The number of segments with high level ofnoise in the existing signal is 7 and total duration of noise is13 minutes and 5 seconds per record.

IV. RESULTS

Our research aims to identify the optimal values for win-dow size and Signal-to-Noise threshold. We assumed that adependency between these two parameters and the accuracyof the algorithm exists. The first parameter that we reviewedis window size, i.e. the amount of samples the filter can holdat any given time. The values we experimented with were250, 625 and 1025, which correspond to 2, 5 and 10 secondintervals sampled at 125Hz.

Our analysis concludes that a filter length of 625 is mostoptimal when working with files sampled at 125Hz. We havealso experimented with values for z, which as we previouslymentioned is the number of samples that will be inserted intothe filter before another calculation of sum is conducted. Ourresearch shows that from three analyzed values: 25, 50 and100 - 50 and 100 prove to perform better than 25. Figures 3and 4 visualize the data on which we based our conclusionsfor filter length and z. The data is presented so that the x-axis visualizes the number of noisy episodes detected by ouralgorithm, while the y-axis presents the file and filter lengthwhich has been used in format file name-filter length.

The correct number of noisy episodes that are present inthe noise-corrupted files is 7. As it was previously mentioned,

TABLE IIACHIEVED RESULTS FOR PERFORMANCE TO DETECT HIGH LEVEL OF

NOISE IN THE ECG SIGNAL.

Record Threshold SEN PPV 1/Fscore118e 6 20000 0.99 0.96 0.95119e 6 20000 0.98 0.97 0.96118e00 20000 0.97 0.97 0.93119e00 20000 0.90 0.98 0.88118e06 45000 0.99 0.72 0.70119e06 45000 0.98 0.96 0.95118e12 45000 0.95 0.71 0.68119e12 45000 0.79 0.97 0.76118e18 70000 0.99 0.47 0.46119e18 70000 0.97 0.54 0.51118e24 70000 0.94 0.46 0.42119e24 70000 0.83 0.50 0.41

noise was not added to the whole signal but in segments. Thefirst segment starts at the fifth minute, lasts two minutes, thenin alternate segments of 2 minutes without and with noise. Sothe total number of noisy episodes for each file is 7.

The second parameter is Signal-to-Noise threshold forwhich we have experimentally determined the optimal value.Fig. 5 visualizes the effect of noise on the value of sum.We have experimentally concluded that 20000 is an appro-priate threshold for differentiating between clean and noisysignal when SNR is between -6 and 0. The threshold forthe remaining files has also been experimentally deduced.This experimental determination of the value of thresholdwas realized by testing the performance of the system todifferent values, and determining which value proved to be anappropriate differentiating line between the clean and noisysignal.

Table II contains the results from all of the conductedexperiments.

V. DISCUSSION

Our research shows that a filter length of 625 is optimalwhen working with data sampled at 125Hz. Conducted ex-periments also prove that 50 samples is an optimal value forthe recalculation interval. For the sake of readability, the datapresented in Table II has only been processed once with afilter length of size 625 and z=50.

Experimental data showed that a threshold of 20000 willresult with satisfactory performance of the algorithm in termsof SEN, PPV and 1/F score if the data has an SNR between -6and 0. This was further confirmed with the results presentedin Table II. As the SNR increases, it is evident that itbecomes harder to differentiate between the clean and noisysignal, hence the poor performance of the algorithm as SNRapproaches 24.

Our analysis presents a new tool for determining the SNRin an ECG recording. As 1/F score decreases below 0.70, theSNR increases - which means that the effect of noise on theoriginal signal is minimal. This allows us to approximate thevalue of SNR in an ECG recording and use this measureas a tool for validating the diagnosis given by ECGalert.For example, if our system detects multiple episodes in one

MIPRO 2019/DC 317

20 40 60 80 100 120 140 160 180

No. of episodes

100 50 25

0 10 20 30 40 50 60 70 80 90

118e_6-250

118e_6-625

118e_6-1025

118e00-250

118e00-625

118e00-1025

118e06-250

118e06-625

118e06-1025

118e12-250

118e12-625

118e12-1025

118e18-250

118e18-625

118e18-1025

118e24-250

118e24-625

118e24-1025

No. of episodes

100 50 25

Fig. 3. Visualization of the effect of window size and recalculation interval on the number of detected noise episodes in 118

0 10 20 30 40 50 60 70

118e_6-250

118e_6-625

118e_6-1025

118e00-250

118e00-625

118e00-1025

118e06-250

118e06-625

118e06-1025

118e12-250

118e12-625

118e12-1025

118e18-250

118e18-625

118e18-1025

No. of episodes

119e18-1025

119e24-250

119e24-625

119e24-1025

100 50 25

0 20 40 60 80 100 120 140 160 180

119e_6-250

119e_6-625

119e_6-1025

119e00-250

119e00-625

119e00-1025

119e06-250

119e06-625

119e06-1025

119e12-250

119e12-625

119e12-1025

119e18-250

119e18-625

119e18-1025

119e24-250

119e24-625

119e24-1025

No. of episodes

100 50 25

Fig. 4. Visualization of the effect of window size and recalculation interval on the number of detected noise episodes in 119

318 MIPRO 2019/DC

0

10000

20000

30000

40000

50000

60000

70000

80000

90000

1 11 21 31 41 51 61 71 81 91 101

111

121

131

141

151

161

171

181

191

201

211

221

231

241

251

261

271

281

291

301

311

321

331

341

351

361

371

381

391

401

411

421

431

441

451

461

471

481

491

501

511

521

531

541

551

561

571

581

591

601

611

621

631

641

651

661

671

681

691

701

SUM

Fig. 5. Visualization of the effect of noise on the value of sum

recording, we may want to confirm that the SNR is highenough so that noise does not get wrongly classified as aQRS complex. By using 1/F score as a measure of recordingquality, we add an automated layer of diagnosis verificationto our system.

VI. CONCLUSIONS

In this paper we have presented a differential noise detectionfilter, and we have provided in-depth analysis on the effects ofwindow size and Signal-to-Noise threshold on the sensitivity,positive predictive value and F value of the algorithm. Ourresearch proves that when using optimal parameters, the al-gorithm performs well in different environments with variableamounts of noise.

This approach provides us with the tools to verify that theprocessed ECG recordings are of good quality and can becorrectly processed by our system.

REFERENCES

[1] R. Trobec, I. Tomasic, A. Rashkovska, M. Depolli, and V. Avbelj,“Commercial ECG systems,” in Body Sensors and Electrocardiography.Springer, 2018, pp. 101–114.

[2] M. Gusev and E. Domazet, “Optimizing the impact of resampling onQRS detection,” in International Conference on Telecommunications.Springer, 2018, pp. 107–119.

[3] N. Thakor, J. Webster, and W. Tompkins, “Optimal QRS detector,”Medical and Biological Engineering and Computing, vol. 21, no. 3,pp. 343–350, 1983.

[4] M. Gusev and E. Domazet, “Optimal DSP bandpass filtering for QRSdetection,” in 2018 41st International Convention on Information andCommunication Technology, Electronics and Microelectronics (MIPRO).IEEE, 2018, pp. 0303–0308.

[5] A. Milchevski and M. Gusev, “Improved pipelined wavelet implemen-tation for filtering ECG signals,” Pattern Recognition Letters, vol. 95,pp. 85–90, 2017.

[6] B.-U. Kohler, C. Hennig, and R. Orglmeister, “The principles of softwareQRS detection,” IEEE Engineering in Medicine and Biology Magazine,vol. 21, no. 1, pp. 42–57, 2002.

[7] P. S. Hamilton and W. J. Tompkins, “Quantitative investigation ofQRS detection rules using the MIT/BIH arrhythmia database,” IEEEtransactions on biomedical engineering, no. 12, pp. 1157–1165, 1986.

[8] A. Velayudhan and S. Peter, “Noise analysis and different denoisingtechniques of ECG signal-a survey,” IOSR Journal of Electronics andCommunication Engineering (IOSR-JECE), pp. 40–44, 2018, e-ISSN:2278-2834, p- ISSN: 2278-8735.

[9] I. Jekova, A. Cansell, and I. Dotsinsky, “Noise sensitivity of three sur-face ECG fibrillation detection algorithms,” Physiological measurement,vol. 22, no. 2, p. 287, 2001.

[10] R. Sameni, M. B. Shamsollahi, C. Jutten, and G. D. Clifford, “A nonlin-ear bayesian filtering framework for ECG denoising,” IEEE Transactionson Biomedical Engineering, vol. 54, no. 12, pp. 2172–2185, 2007.

[11] I. Jekova, V. Krasteva, I. Christov, and R. Abacherli, “Threshold-basedsystem for noise detection in multilead ECG recordings,” Physiologicalmeasurement, vol. 33, no. 9, p. 1463, 2012.

[12] J. Lee, D. D. McManus, S. Merchant, and K. H. Chon, “Automaticmotion and noise artifact detection in Holter ECG data using empiricalmode decomposition and statistical approaches,” IEEE Transactions onBiomedical Engineering, vol. 59, no. 6, pp. 1499–1506, 2012.

[13] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C.Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E.Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: Components of anew research resource for complex physiologic signals,” Circulation,vol. 101, no. 23, pp. e215–e220, 2000 (June 13).

[14] International Electrotechnical Commission. (2012) Internationalstandard of IEC 60601-2-47, medical electrical equipment part2-47: Particular requirements for the basic safety and essentialperformance of ambulatory electrocardiographic systems. [Online].Available: https://webstore.iec.ch/publication/2666

[15] E. Domazet and M. Gusev, “Improving the QRS detection,” Journal ofTechnology and Healthcare, 2019, in press.

MIPRO 2019/DC 319