The Effect of White Noise and False Peak Detection on

HRV Analysis

G. Manis

, A. Alexandridi

, S. Nikolopoulos

and K. Davos

University of Ioannina, Dept. of Computer Science,

P.O. Box 1186, Ioannina 45110, Greece

Foundation of Biomedical Research of the Academy of Athens

Sor. Ephessiou 4, 114 27 Athens, Greece

National Technical University of Athens

Dept. of Electrical and Computer Engineering

Zografou Campus, Zografou 15773, Athens, Greece

Abstract. Heart rate variability (HRV) is an established measure for cardiac

health. Its use is widespread and many methods have been developed for its

analysis. Little emphasis, however, has been given to the specific influence of

noise from the electrocardiogram (ECG) on the heart rate (HR) series. There

are explicit factors of noise that have been extensively studied on the ECG and

much work has been published on their limitation or elimination. Despite all

these solutions, however, often noise does end up in the ECG and is inevitably

included in the derived HR series. It is of interest to investigate how this influ-

ences subsequent HRV analysis. We propose that the noise into the resulting

HR series: Shifted R-peak (white noise) and false peaks. In this paper, we dem-

onstrate how these two scenarios affect the outcome of the HRV analysis.

1 Introduction

HRV is computed from the variations in temporal distance between R-peaks of the

electrocardiogram (ECG) [1,2]. In order for these peaks to be extracted from the

original recorded signal, the QRS complex must be properly and accurately identi-

fied. The cancellation or elimination of noise in the ECG has been examined compre-

hensively by the medical and engineering communities and there has been a plethora

of suggestions on general as well as specific noise reduction methods [3]. Although

the problem has been greatly reduced, QRS points are still often measured with some

amount of error.

Since the heart rate (HR) signal is detected from the ECG, it would be of interest to

investigate how erroneous R-R intervals influence the subsequent HRV analysis

methods in their accuracy. Previous studies have addressed this issue [4,5], but gen-

eral observations are still lacking in the literature. In this paper, we investigate the

performance of the most common linear HRV analysis methods on HR series that

have been contaminated with varying amounts of noise.

Manis G., Alexandridi A., Nikolopoulos S. and Davos K. (2005).

The Effect of White Noise and False Peak Detection on HRV Analysis.

In Proceedings of the 1st International Workshop on Biosignal Processing and Classiﬁcation, pages 161-166

DOI: 10.5220/0001195301610166

 SciTePress

2 Noise Artifacts

The most common noise artifacts present in the ECG have been well highlighted in

[6] and these will be examined in this paper. They include power line interference,

electrode contact noise, motion artifacts, muscle contractions and baseline drift and

amplitude modulation caused by respiration.

In the clinical and research context, there are three settings through which ECGs

are digitally recorded: ECG monitoring, Holter devices and the treadmill stress test.

Noise found in the ECG is commonly either filtered or discarded [6]. In the case of R-

peak detection, noise that is not properly eliminated usually produces one of two

errors in the HR series: false R peaks or lost R peaks.

Peak detection errors usually result in the introduction of white noise to the HR se-

ries. As false peaks are introduced or points are removed from the original ECG se-

ries, the true temporal occurrence of the R-peaks is slightly shifted in the resulting

HR series. It is reasonable to claim then, that if the noise reduction method is not

finely tuned and R-peaks are not identified properly, there is a larger amount of white

noise present in the computed HR series. Below we list some of the most common

forms of ECG noise, most of them described in more detail by [6].

a) Power line interference. This type of noise inserts 50 Hz or 60 Hz into the sig-

nal as well as harmonics that appear as sinusoids or a combination of sinusoids [6]. A

sample of methods to reduce power line interference may be found in [7,8] and [9].

b) Motion artifacts. These are modulations in the baseline that are caused by

change in electrode impedance at the site of contact of the electrode with the skin [6].

Often this is caused by respiration, perspiration or physical activity. A sample of

noise reduction techniques may be found in [10-12].

c) Electrode contact noise. This occurs when the electrode contact with the skin is

lost, usually leading to a loss in measurement [6].

d) Muscle contractions. These introduce milli-volt potentials to the signal [6]. This

is also a common occurrence in Holter recordings and treadmill tests. A sample of

muscle contraction noise reduction methods may be found in [13,14].

e) Baseline drift and respiration. This often has the appearance of a sinusoidal

component at the frequency of respiration [6]. A sample of noise reduction methods

may be found in [12, 15, 16].

f) Sampling. This kind of noise is due to quantization (digital recording). An ac-

ceptable minimum sampling rate found in most cases is 300 samples per second.

3 Methods

In the previous section we discussed the most common and significant sources of

ECG noise. In this paper, our interest is how this noise affects the HR series. From

the above discussion, we conclude that noise is inserted into the HR series in one of

three ways: Shifted R-peak (which is commonly white noise), false peak detection

and lost peaks. The third of these interferences does not really influence the outcome

of the HRV analysis. Therefore, the results of the first two cases are reported here.

162

In order to investigate the above, we inserted white noise (1%-20% of the mean

value of the signal). It was then of interest to see how these various amounts of noise

affected the discrimination between patient groups as well as the statistical signifi-

cance of the results (p-value). The p-value indicates the possibility that the two data

sets are not different. A similar method was applied in order to test the affect of false

peaks on the HRV analysis. N false peaks were inserted to the timeseries, with N

ranging from 1% to 20% of the total signal length. N new points were created each

one by replacing an existing point with two new ones, with the sum of the two new

points being equal with the value of the replaced point, generating in that way the

effect of a false peak detection.

The HRV analysis methods which are applied in this study are the most commonly

used in clinical and research settings [6]: the SDNN, the SDANN, the SDNNindex,

the RMSSD, the pNN50 and the SDSD. We have also applied Local Linear predic-

tion [17]. The SDNN is the standard deviation of the NN interval, where NN is

equivalent to two successive R-peaks in the HR signal [6]. The SDANN is the stan-

dard deviation of the average NN interval that is calculated from short periods com-

monly 5 minutes in length and the SDNNindex is the mean of the 5-min SDNN com-

puted over 24h [6]. The RMSSD entails measuring the successive beat-to-beat inter-

vals and finding their distribution [6]. The pNN50 is the proportion resulting from the

division of the number of interval differences of successive NN intervals greater than

50ms and the total number of intervals. The SDSD is the standard deviation of differ-

ences between successive NNs [6]. Local Linear prediction [17] predicts future sam-

ples of a time series x

,…,x

by using a linear combination of the previous k sam-

ples, where k is a specific window length.

Two sets of data are analyzed in this study. The first set, henceforth referred to as

‘Data set 1’ is composed of HRV signals of healthy young subjects, both male and

female and a set of healthy elderly subjects, also both male and female. This data is

drawn from the ‘Fantasia’ database (Physiobank), [18]. Both the younger subjects

(21-34 years old) and the elderly (68-85 years old) have been thoroughly tested to

establish cardiac health. The recordings are 120 minutes long and all subjects were

resting in the supine position while watching the movie ‘Fantasia’. The second set,

henceforth referred to as ‘Data set 2’ is a collection of Holter recordings acquired

from heart failure subjects and an analogous control group. The data of the Holter

group is comprised of 24 hour recordings.

4 Results

Results are shown both in the case of white noise as well as in the case where false

peaks are inserted into the HR series. We are interested in investigating how the dif-

ferentiation between young and elderly subjects (Data Set 1) and control/patient sub-

jects (Data Set 2) is affected by these white noise and how the insertion of false peaks

effects the results of Data Set 1. The effects of false peaks on Data Set 2 are still un-

der investigation and will be included in future work by the authors. The possibility

of error in the classification is shown through the computation of the p-value

(ANOVA test) and is taken as significant when p < 0.05. This method is applied since

163

it is the most common among the clinical community in the statistical assessment of

classification studies. The statistical significance of the methods applied is shown in

Table 1. Due to space limitations, we only include one indicative figure of the results

of the SDNN method. The results of all the other methods are summarized in Table 2.

Fig. 1.1 White noise effects on the SDNN method on the discrimination of controls or younger

subjects (circles) and heart failure or elderly subjects (pluses). The statistical possibility of

error of classification for each data set is also shown

Fig. 1.2 False Peak effects on the SDNN method on controls (circles) and heart failure subjects

(pluses). The statistical possibility of error of classification is also shown

As seen in Fig. 1, the data of Data Set 1 is greatly affected by noise after the 12%

noise level. The same holds for Data Set 2 after roughly 6% noise is inserted. The

outcome of the method is statistically significant for Data Set 1 up to roughly 13%

noise insertion while for Data Set 2 all levels of noise produce statistically significant

164

results. The insertion of false peaks to Data Set 2 immediately affects the discrimina-

tion between the two subject groups (even at 1% noise), as shown in Figure 1.2. This

is also indicated by the fact that statistically significant results are found only when

no noise is inserted. The above results are summarized in Table 1 and Table 2.

5 Discussion

The effects of noise on these methods have been experimentally examined in this

study. It is of interest, however, to examine why these results have occurred. Al-

though the standard deviation (SDNN) is the method of choice among most clini-

cians, there has been little interest in its robustness to all the above noise factors. As

has been indicated here, it is clearly better than the others, as is the SDANN, which is

similar but applied to 5-min intervals of the data. The RMSSD, which is also most

common among clinicians, does not produce such encouraging results here, since it is

a measure of spread and not a direct measurement of the deviation. The poorer per-

formance of the SDNNindex is most likely attributed to the individual examination of

small segments of holter data, which is unstable and usually contains the most noise

of all recording equipment. Overall, we believe these are interesting results and we

are currently examining the effects of such noise on other classification methodolo-

gies, which, although not as popular in the clinical setting, do provide research inter-

est.

Table 1. Maximum noise level at which statistical significance of data is maintained

White Noise

Method

Data Set 1 Data Set 2

False Peaks

SDNN 13% All None

SDANN All All Varies

SDNNindex - 9% None

RMSSD 7% All None

SDSD 8% 18% None

LLP 11% All 4%

Table 2: Maximum noise level at which classification of data sets is maintained

White Noise

Method

Data Set 1 Data Set 2

False Peaks

SDNN 12% 6% None

SDANN All All 2%

SDNNindex - 2% 5%

RMSSD 6% 2% None

SDSD 5% 2% None

LLP 3% 2% None

165

References

1. Task force of the European Society of Cardiology and the North American Society of Pacing

and Electrophysiology: Heart rate variability: Standards of measurement, physiological in-

terpretation, and clinical use. Circulation (1996) 93:1043-1065

2. Teich MC, Lowen SB, Jost BM, Vibe-Rheymer K, Heneghan C.: Heart-Rate Variability:

Measures and Models. Nonlinear Biomed Signal Processing, Vol. II, Dynamic Analysis

and Modeling. M. Akay, IEEE Press, New York (2001) 159-213

3. Daskalov, I.K, Dotsinsky I.A., Christov I.I.: Developments in ECG Acquisition, Preprocess-

ing, Parameter Measurement, and Recording. In IEEE Eng Med Biol, (1998) 17(2):50-58.

4. Hilton M.F., Beattie J.M., Chappell M.J., Bates R.A.: Heart rate variability: measurement

error or chaos? Computers in Cardiology (1998) 24:125-128

5. Signorini MG, Marchetti F, Cerutti S.: Applying nonlinear noise reduction in the analysis of

heart rate variability. In IEEE Eng Med Biol Mag. (2001) 20(2):59-68

6. Friesen, G.M., Jannett, T.C., Jadallah, M.A., Yates, S.L., Quint, S. R. and Nagle, H.T., 1990.

A comparison of the noise sensitivity of nine QRS detection algorithms.In IEEE Trans

Biomed. Eng, 1990;37:85-98.

7. Adli, Yamamoto, Y.: Impedance balancing analysis for power-line interference elimination

in ECG signal. IMTC/98. Conference Proceedings (1998) 1:235 – 238

8. Hamilton, P.S.: A comparison of adaptive and nonadaptive filters for reduction of power line

interference in the ECG. IEEE Trans Biomed Eng (1996) 43(1):105 – 109

9. Ziarani, A.K., Konrad, A.: A nonlinear adaptive method of elimination of power line inter-

ference in ECG signals. IEEE Trans Biomed Eng (2002) 49(6):540 – 547

10. Hamilton, P.S., Curley, M.G., Aimi, R.M., Sae-Hau, C.: Comparison of methods for adap-

tive removal of motion artefact. In Comp in Cardiol (2000) 383 – 386

11. Tong, D.A., Bartels, K.A., Honeyager, K.S.: Adaptive reduction of motion artifact in the

electrocardiogram. Engineering in Medicine and Biology, 2002. EMBS/BMES Conference

(2002) 2:1403 – 1404.

12. Hamilton, P.S., Curley, M.G.: Adaptive removal of motion artifact [ECG recordings] Engi-

neering in Medicine and Biology society, Proceedings of the 19th Annual International

Conference of the IEEE (1997) 1:297 – 299

13. Sun, P., Wu, Q.H., Weindling, A.M., Finkelstein, A., Ibrahim, K.: An improved morpho-

logical approach to background normalization of ECG signals. IEEE Trans Biomed Eng

(2003) 50(1):117 – 121

14. Benitez, I., Lu, C.-C.: Portable real-time body surface Laplacian ECG mapping.

BMES/EMBS Conference (1999) 1:301

15. Pandit, S.V.: ECG baseline drift removal through STFT. Engineering in Medicine and

Biology Society, 1996. Bridging Disciplines for Biomedicine. Proceedings of the 18th An-

nual International Conference of the IEEE (1996) 4:1405 – 1406

16. Cuiwei L., Chongxun Z., Changfeng T.: Detection of ECG characteristic points using

wavelet transforms. IEEE Trans Biomed Eng (1995) 42(1):21 – 28

17. Manis G., Nikolopoulos S., Alexandridi A.: Prediction Techniques and HRV Analysis.

MEDICON (2004).

18. Porta A, Baselli G, Guzzetti S, Pagani M, Malliani A, Cerutti S.: Prediction of short car-

diovascular variability signals based on conditional distribution. IEEE Trans Biomed Eng.

(2000) 47(12):1555-64.

166