Signal Preprocessing And Analysis Of HRV
Short-term HRV analysis and long-term HRV analysis are two methods for evaluating the autonomous nervous system (ANS). First and foremost, identifying and removing artifacts, ectopic beats, and arrhythmia as well as assessing the clean and reliable HRV, is the first stage in analyzing the HRV signal. For measuring HRV, we need R peaks in the QRS complexes. Therefore, QRS complex extraction is of paramount importance. Different noises including motion artifacts, powerline interference and respiration, can affect the accuracy of R-peak detection. Pre-processing of ECG signal leads to a more accurate QRS detection and more reliable HRV. As seen in the fig. 8 diagram, the main pre-processing procedure are resampling, denoising, QRS detection, correction and detrending, which are discussed in the following sections.
(i) Resampling
Different recording conditions and analysis options have an impact on HRV measurements. For reducing observational errors and improving peak detection, selecting the appropriate sample rate for digitizing the signal is critical.
When the initial sample rate is quite high, we reduce the number of sample points per second by down sampling. However, a low sampling frequency may cause random errors in power of high frequency and distort the peak detection. As a result, the detailed information of HRV signal would be lost. On the other hand, in the era of telemonitoring and wearable devices, we must choose a lower sample rate to reduce the system’s computing expenses and complexity while also optimizing energy use because of limitations imposed by smaller device sizes and lower battery capacities [1].
The recommendations of the American Heart Association and the Association for the Advancement of Medical Instrumentation suggest that the minimum sampling rate for a digital ECG recording should be 500 samples per second (Hz), however, depending on the utilized method for QRS detection, a different sampling rate could be chosen. Accordingly, the table below lists some studies, their suggested sample rates, used interpolation method and features related to analysis domain.
Briefly, the mean amplitudes of the R-peaks tend to decrease as the sampling rate reduce [2]. Pizzuti et al. reported that they didn’t detect significant differences between the recordings sampled at rates of 250 Hz and 500 Hz, besides many studies have introduced this range as an optimal and acceptable range of sampling rate [1, 3]. However, in [4] and [5] between 500Hz and 1000Hz have been chosen as optimum ECG sampling frequency.
By comparing different sampling frequencies and evaluation of resultant HRV, Hejjel et al. suggested a sampling frequency of 1ms interval for HRV analysis which without interpolation gives an accurate time domain measure [6].
However, for spectral analysis, sampling rate dependent on healthy or patient cases could be different [7]. For instance, in the case of pathologically decreased HRV in patients, the spectral characteristics could be intensively affected by sampling frequency.
Some studies recommend utilizing interpolation for ECG data with low sampling frequency. Interpolation is a technique for increasing the quality of ECG signal and the accuracy of R-peak detection and then HRV assessment.
By comparing HRV measurements with and without cubic spline interpolation, a common interpolation method, a sampling rate of 71 Hz with interpolation has been suggested to be effective [4]. Using down-sampling to 100 Hz and even 50 Hz and then anti-aliasing filter, Mahdiani et al interpolated the ECG signal by cubic interpolation method and chose 71Hz for sampling rate. But this is probably not suitable for wearable devices because of the non-symmetry property of the R-peak waveform in some cardiovascular diseases [8].
To sum it up, low sampling frequency, i.e., below 250 Hz, can lead to a low quality and unreliable HRV which it can be ironed out by a well-chosen interpolation technique. In order to access accurate HRV and utilize it for mental state monitoring purposes, it seems that one of the most beneficial techniques has been proposed in [9] that is based on piecewise cubic Hermite interpolation (PCHIP) and piecewise cubic spline interpolation (SPLINE) for lower sampling frequency (128 Hz and lower). Moreover, methods based on cubic spline interpolation (CSI) algorithm, are efficient in terms of latency reduction and complexity, thus they are proper for real-time implementation.
In table 2, the studies are arranged according to the year they were published. As seen, depending on the utilized method for QRS detection, a different sampling rate could be chosen. Accordingly, the optimum range of sampling is 250-500 Hz, in which an accurate HRV in terms of HRV parameters in different domains can be assessed without using interpolation.
| Article | Goal and method | Sampling rate | interpolation | compared features | Subjects/Data |
| Habib et al. 2020 [10] | Accurate HRV QRS detection using CNN Acceptable range for sampling frequency |
100,250 Hz | interpolation distort high frequency signals | CNN Model accuracy | MIT-BIH, mix of healthy and arrhythmia |
| Petelczyc et al. 2020 [4] | accurate HRV Errors Due to Sampling Frequency |
512 Hz clinical |
_ | Time domain Frequency domain nonlinear parameters |
MIT-BIH, mix of healthy and arrhythmia |
| Kwon et al. 2018 [1] | Accurate HRV Acceptable range for sampling frequency |
250 Hz | linear interpolation | Time domain Frequency domain Poincare plot analysis |
ED patients |
| Ajdaraga & Gusev 2017 [3] | QRS detection Acceptable range for sampling frequency |
optimal range is 250-500 Hz sufficient accuracy: 120Hz |
_ | different Model accuracy | MIT-BIH, mix of healthy and arrhythmia |
| Mahdiani et al. 2015 [8] | Accurate HRV | 100 Hz | Cubic Spline interpolation | Time domain | healthy |
| Ellis et al. 2015 [11] | Accurate HRV reduce severity of false positive R peak |
125Hz without interpolation 71Hz with interpolation |
Cubic Spline interpolation | Time domain Frequency domain |
PTBDB, healthy |
| Sidek and Khalil 2013 [9] | enhance HRV, use interpolation | 128Hz | Hermite interpolation piecewise cubic spline interpolation |
_ | mix of healthy and patients MIT-BIH: NSRDB, SVDB, MITDB PAF: AFPDB |
| Ziemssen et al. 2008 [7] | influence of sampling on spectral analysis and baroreflex accurate TRS technique |
100Hz )normal HRV, higher HRV, RMSSD>10( |
_ | Frequency domain | EUROBAVAR |
| Hejjel and Roth 2004 [6] | Acceptable range for sampling interval | 1000Hz | _ | Time domain | healthy |
| Berntson et al. 1997 [5] | HRV Methods accurate HRV |
500–1000Hz 250Hz normal HRV |
_ | _ | _ |
| Malik et al. 1996 [12] | Standards of measurement | 250 – 500Hz 128Hz normal HRV and interpolation |
_ | _ | _ |
| Abboud and Barnea 1995 [13] | accurate HRV Errors Due to Sampling Frequency |
125Hz normal HRV higher HRV lower HRV: 1KHz |
_ | Frequency domain | mix of healthy and patients |
| Menz 1994 [14] | Accurate HRV Acceptable range for sampling frequency |
500Hz | _ | Amplitude of peaks | _ |
| Pizzuti et al 1985 [2] | Accurate HRV Comparison between 500Hz and 100Hz, 500Hz and 250Hz Acceptable range for sampling frequency |
250-500 Hz | _ | Amplitude of peaks intervals |
healthy |
(ii) Denoising
Generally, the ECG denoising methods have been classified into different techniques [11] illustrated in figure 9. As a matter of fact, most of the methods are suitable for the static ECG signals rather than ambulatory recordings. In wearable concept the most prominent noise is motion artifact. Motion artifact can be originated from both electrode movement (EM) and the muscle artefact (MA). Table 3 shows various types of noises that contaminate the ECG, their frequency range, and the sources from which they originate. According to the latest review [15] and conveying literature based on all noise removal techniques, the most efficient methods to enhance the SNR are as explained in table 4. Some of these approaches have been designed for wearable concept and some other have the potential to be implemented real-time in wearables.
EMD is a local and adaptive method in the frequency-time domain which is appropriate for nonstationary and non-linear signals. In accordance with researches, it could be implemented real-time also, can remove motion artifact and other types of noises including BW and PLI with quite high performance [20-22].
Deep-learning based methods have been widely used recently. A denoising autoencoder (DAE) is a machine learning model that recreates the input signal with the highest achievable level of accuracy using two nonlinear subparts called encoder and decoder [17, 40]. Moreover, the improved approach of DAE for ECG denoising named Generative Adversarial Network (GAN), is used in [15] and they reported SNR improvement of 41.36% for specifically MA artifact. Although they are offline for training stage, the denoising stage is online and can be used for real time denoising.
Wavelet-based methods have an advantage to be real time besides its efficiency in terms of SNR improvement [18, 27-37, 40]. One of the considerations that should be studied is choosing the appropriate mother wavelet and a thresholding technique. The selection of an appropriate wavelet basis function for denoising an ECG signal was carried out in [16] research.
Sparsity-based decomposition to denoise ECG signal breaks the signal into sparse parts and residue and utilizes the sparse parts in order to estimate clean signal [11]. Least Mean Squares (LMS) and Recursive Least Squares (RLS) are among this category’s techniques that have a good performance in online motion artifact removal [36, 27, 39, 43, 44].
One of the Bayesian-filter based denoisers for ECG is Kalman filter and its versions such as adaptive and extended Kalman filter (AKF and EKF). Being applied on multiple databases as well as being real-time and suitable for Holter monitoring, AKF and EKF successfully could remove MA artifact and enhance the SNR [14, 16].
Based on table 4, the most effective ECG denoising techniques that can be used in online ambulatory ECG monitoring using wearable devices are AKF [14], EKF [16], LMS adaptive [25], SWT [29], variable frequency demodulation [41] and RMS [42]. Therefore, because of low computational complexity and high performance in motion artifact removal, implementation of these methods can fulfill our enquires for ambulatory application.
| Noise | Range of frequency | Source |
| baseline wander (BW) | 0.15–0.3 Hz | respiration, body movements, bad electrode contact, and skin-electrode impedance |
| powerline interference (PLI) | 50/60 Hz | inductive and capacitive couplings of power lines |
| electromyographic (EMG/MA) | 0.01–1000 Hz | electrical activities in muscles |
| Motion artifact (large swing in the baseline) (MA) | 1-10 Hz | stretching, coughing, ambulation and changes in the electrode skin impedance |
| Channel Noise | All frequency components, such as AWGN [36] | Poor channel conditions |
| Electrode Contact Noise (EM) | Duration of 1 second [36] | loss of contact between the electrode and the skin |
Table 3. The type of noise, range of frequency and their sources
| Reference | Method | Noise and Artifact | advantage or disadvantage | Validation |
| Hesar HD, Mohebbi M 2020 [18] | Adaptive Kalman Filter (AKF) | WGN MA |
Suitable for Holter monitoring, Simple & fast preprocessing, Applied on 4 databases |
SNR improvement= 9.90 % |
| Wang J. et al. 2019 [19] | Generative Adversarial Network (GAN) based on deep learning | MA EM |
Offline training Online denoising |
SNR improvement (MA) = 41.36 % SNR improvement (EM) = 38.09% |
| Hesar HD, Mohebbi M 2017 [20] | New marginalized extended Kalman Filter (MP-EKF) | MA | Evaluated on arrythmia (2 databases) | SNR improvement= 16 % |
| Chiang HT et al. [21] | Fully convolutional network denoising autoencoder (FCN-based DAE) | EM+BW+MA | Higher performance than DNN/CNN based DAEs | SNR improvement= 15.49 % |
| Manas Rakshit, Susmita Das 2018 [22] | Dictionary learning-based sparse representation | MA PLI BW WGN |
2 databases | SNR improvement (MA) = 11.69 % |
| Oscar Hernández, Edgar Olvera 2009 [23] | Undecimated wavelet transform (UWT) | Mixture of various noises (EM, BW, MA, PLI…) | offline | SNR improvement= 17 % |
| Lee, McManus, Merchant, & Chon 2012 [24] | EMD | Motion and Noise Artifact | Real-time | Sensitivity=96.63% Specificity=94.73% |
| Blanco-Velasco et al. 2008 [25] | EMD | high frequency noise baseline wander |
NR | |
| Lakhe et al. 2020 [26] | Spectral trimming technique | motion artefact | Real time | RMSE = 17.24 CC = 0.91 SNR = 7.09 |
| Hossain et al. 2020 [27] | Variable frequency complex demodulation (VFCDM) algorithm | Muscle and motion artifact | Real time high frequency noise |
Sensitivity =99.14% P+ = 97.62% Accuracy = 96.80% |
| Chandrakar & Sharma 2015 [28] | adaptive window technique (Threshold in time series for detect R) |
power line interference EMG interference baseline wandering motion artifact |
Real time Fast process Easy and cheap |
NR |
| Kim et al., 2012 [29] | LMS adaptive filter | Motion artifact baseline wandering |
Real time Need noise reference channel |
Sensitivity=78.03% |
| Hanine et al. 2017 [20] | RLS Adaptive Filtering | EMG interference | Real time Need noise reference channel |
NR |
| P. S. Addison, 2005 [31] | CWT DWT SWT WPT |
NR | CWT: high computational cost | NR |
| El B’charri, Latif, Elmansouri, Abenaou, & Jenkal, 2017 [32] | DT-WT Modified unified threshold Semi-soft function |
Combined noise (highest weight for MA) | Real time Overcome shortcomings like aliasing All kinds of noises |
For MA: SNR imp. = 7.22 MSE = 0.00277 |
| Berwal et al 2019 [33] | SWT Multi-resolution thresholding |
motion artifact EMG interference power-line interference baseline wandering |
separation of low and high frequency motion artifacts | correlation coefficient=0.9337 NMSE=0.012 |
| Nagai, Anzai, & Wang, 2017 [34] | SWT Haar wavelet (Db1) (Energy of signal for QRS detection and energy threshold) |
motion artifact | Real time | correlation coefficients=0.88 |
| Bhoraniya & Kher, 2014 [35] | DWT Bior6.8 Symlet4 |
motion artifacts | NR | NR |
| Chen et al. 2006 [36] | DWT 3 levels hard threshold Db4 (QRS adaptive threshold) |
motion artifact power-line interference baseline wandering |
Real time | correct detection rate=99.5% |
| Mukhopadhyay et al 2012 [37] | DWT 8 levels Db6 |
NR | NR | NR |
| Yanık et al., 2020 [38] | Wavelet analysis Db6 wavelet zero thresholding |
EMG interference power-line interference baseline wandering |
NR | NR |
| Wang et al. 2019 [39] | Wavelet fibr Sym4 Db4 Soft threshold |
EMG noise high frequency noise |
Enhance P, T wave detect high frequency noise |
NR |
| Boutaa, Bereksi-Reguig, & Debbal, 2008 [40] | Multiresolution analysis using wavelets Daubechies wavelet universal threshold Soft thresholding (Threshold in wavelet for QRS) |
power interference EMG interference motion artifact baseline wandering |
NR | Sensitivity=99.88% Predictivity=99.89% |
| Shimauchi et al. 2021 [41] | CWT Hann-windowed complex sinusoid Morlet complete morlet |
EMG interference motion artifact baseline wandering |
semi real time | F1 score=0.98, 0.95 |
| Lázaro et al. 2020 [42] | PCA, NLMS (Adaptive Weiner), an innovative channel selection method | EMG | Armband, real-time | Simultaneous Holter Monitoring Accuracy = 90.79% Sensitivity = 92.05% Specificity = 90% |
| Reljin et al. 2020 [43] | Redundant Convolutional Decoder-Encoder (R-CED) | White/pink/blue/purple/brown/motion offline |
Armband | Simultaneous Holter Monitoring SNR, Cross correlation, Ratio of power SNR= 91.86% ratio of power= 0.71 _ 0.03 cross-correlation= 0.77 _ 0.06 |
| Berwall et al. 2018 [44] | DWT (Biorthogonal Spline WT) | NR | Wearable, Arrhythmia detection | Sensitivity = 99.31% +P = 99.19% DER (detection error rate) = 1.49% |
| Hossain et al., 2021 [45] | variable-frequency complex demodulation (Pan and Tompkins QRS detection) |
Motion artifact | Long term monitor and wearable | P+ = 92.9577% SNR= 1.4595 ± 0.6326 |
| An & K Stylios, 2020 [88] | Adaptive filter RLS impedance pneumography as reference signal |
Motion artifact | Real time Good performance for normal/abnormal ECG Need noise reference channel |
Correlation coefficient= 0.9877 Mean square error= 0.0042 R-square= 0.9750 SNR= 19.9919 dB |
| Seol, Lee, & Lee, 2018 [46] | Adaptive filter, RMS LMS |
Motion artifact | Real time Need noise reference channel |
MSE(RLS)=0.0166 MSE(LMS)=0.0160 |
| Brij N. Singh et. Al. 2006 [68] | DWT (Daubechies 8) + Hybrid Sure thresholding scheme | All kinds of noises | Real time Retain necessary diagnostic information by peak preservation |
RMSE = 0.062 |
Table 4. Literature review of noise type, effective removal
CC=Correlation Coefficient
(iii) QRS detection
Peak detection is one of the most challenging steps of ECG preprocessing. Generally, in peak detection, two anomalies can be remarkable. First due to peak detector errors and second due to ectopic beats. We should minimize peak detector errors by choosing the best detection algorithm.
The challenges we face, are due to variety of artifacts and different cardiac abnormalities, such as isolated QRS-like artifacts and premature ventricular contractions (PVC) [44]. Since we use R waves to determine the HRV, these abnormalities affect the analysis of HRV especially in the spectral domain more than in other domains. Table 5 illustrates a literature review on previous proposed QRS detection methods. Also, their computational load and validation parameter are reported.
The most reputed and the gold standard algorithm for QRS detection, which has the capability of real-time processing scenario, is Pan-Tompkin’s method. This approach is based on discrete analysis of slope, amplitude, and width and detects QRS complex of ECG signals real-time with a sensitivity of 99.76%[50].
| article | method | Computational load | Validation |
| Poli et al. 1995[47] | Genetic Algorithm | NR | (%)=99.60 +P (%)=99.51 |
| Hamilton and Tompkins 1986[48] | BPF Adaptive threshold |
NR | (%)=99.69 +P (%)=99.77 |
| Zhang and Lian 2009[49] | Multiscale morphology | NR | (%)=99.81 +P (%)=99.80 |
| Leong et al. 2012[50] | Quadratic spline wavelet | NR | (%)=99.31 +P (%)=99.70 |
| Nallathambi and principe 2014[51] | Pulse train | NR | (%)=99.58 +P (%)=99.55 |
| Pan-Tompkins 1985 [52] | BPF + FD + squaring + MA Threshold |
medium | (%)=99.76 +P (%)=99.56 Acc (%)=99.3 |
| Li et al. 1995[53] | WT + digital filter Singularity + threshold |
medium | Acc(%)>99.8 (%)=99.84 +P (%)=98.89 |
| Afonso et al. 1996[54] | Filter bank Threshold |
high | (%)=99.59 +P (%)=99.56 |
| Moares et al. 2002[55] | LPF + FD + spatial velocity Threshold |
medium | (%)=99.88 +P (%)=99.69 |
| Martinez et al. 2004[56] | WT Threshold + ZC |
medium | (%)=99.66 +P (%)=99.56 |
| Chiarugi et al. 2007[57] | BPF + first derivative Threshold |
low | (%)=99.76 +P (%)=99.81 |
| Zheng and Wu 2007[58] | DWT + cubic spline + interpolation+ MA Threshold |
High | (%)=99.68 +P (%)=99.59 |
| Arzeno et al. 2008[59] | First derivative + Hilbert transform Threshold |
medium | (%)=99.24 +P (%)=99.29 |
| FD + squaring + BPF Threshold |
(%)=99.58 +P (%)=99.57 |
||
| Threshold’s variation | (%)=99.58 +P (%)=99.57 |
||
| Bsoul et al. 2009[60] | BPF, subtraction the regression line WT(haar) |
low | Acc (%)=99.8 |
| Engelse and Zeelenberg 1979 [61] | FD+digital filter+Threshold | (%)=98.42 +P (%)=98.39 |
|
| Pal | WT Threshold |
medium | (%)=99.6 +P (%)=99.51 |
| Benitez [62] | FD + HT Threshold |
medium | (%)=99.64 (%)=99.81 +P (%)=99.83 |
| Elgendi 2013[63] | BPF + SD + squaring Threshold |
Low | (%)=99.87 +P (%)=99.78 |
| Thirrunavukkarasu et al. 2019 [64] | WT(Db4)+Shannon energy | – | |
| Benitez 2000 [62] | FD + HT | medium | (%)=99.81 94 (SNR 6 db) +P (%)=99.83 |
| Pal S. 2010 [65] | Multiresolution wavelet transform | _ | Acc (%)=96 |
| H. Dikhaus [66] | Averaging and filtering Late potential (LP) scalogram of wavelet transform with FFT |
_ | (%)=86 Specifity = 81% |
| Miranda MV. [67] | feature extraction based on morphology analysis, fiducial point localization, CWT and DWT | _ | (%)=96 _ |
| Chen et al. 2005[36] | moving-averaging incorporating with wavelet denoising | Low (real-time) | (%)=99.55 +P (%)=99.49 |
| Christoph [68] | Multiple MA + FD 3 combined adaptive thresholds: an adaptive steep slope threshold, an adaptive integrating threshold and an adaptive beat expectation threshold |
Low (real-time) | Algorithm I: 99.69 Specifity = 99.65 Algorithm II: 99.74 Specifity = 99.65 |
| Kalidas and Lakshman 2017 [69] | SWT (Db3) + squaring +MA | Low (real-time) | (%)=99.88 +P (%)=99.84 |
Table 5. QRS detection algorithms, their computational load and validation parameter
FD=first derivative, SD=second derivative, BPF=band pass frequency, LPF=low pass filter, MA=moving average, WT=wavelet transform, DWT=discrete wavelet transform, SWT=stationary wavelet transform, CWT=continues wavelet transform, HT=Hilbert transform, ZC=zero crossing
The wavelet transform is now widely used to reduce noise and detect peaks. Comparing the Pan and Tompkins algorithm with the Wavelet transform and concluded that while the Pan and Tompkins algorithm is the most widely used method for QRS detection, wavelet is better for denoising and also faster than the Pan and Tompkins algorithm [70].
According to the literature the most impressive methodology for QRS complex detection is a robust and numerically efficient algorithm proposed by Elgendi et al. [63], which has applied the approach on 11 datasets and achieved a sensitivity of 99.87%, as well as its low implementation complexity. The open-source python package, Neurokit2, has implemented some of the most effective ECG processing and QRS detection methods. In addition, with respect to the research objectives, the other methods by low computational load can be implemented real-time, besides their high validation criterion i.e., sensitivity, positive predictive value, specificity and/or accuracy.
(iv) Signal correction
After peak detection and computing the RR-interval time series, correction the time series and the erroneous and ectopic beats in order to extract reliable features is necessary. Frequency domain features and some other features that are extracted from HRV are very sensitive to sample loss and ectopic beats [90]. Several time domain characteristics, such as PNN50, are also affected by ectopic beats [91]. As a result, owing to the fact that the presence of ectopic beats on ECG can drastically change Heart Rate Variability (HRV), many studies have focused on replacement and correction of the ectopic beats.
The following is a list of commonly used methods for detecting and then correcting ectopic beats:
Given the importance of being real-time in the case of wearable devices, time and morphological approaches, wavelet as well as template matching due to their low complexity are among the best options for ectopic detection [73].
Temporal features are mostly based on QRS duration, RR interval duration and QRS pattern, however morphological parameters are measured by cross correlation and time laps [74]. In a study by Lipponen et al [71], two variable thresholds for detecting and correcting abnormal beats (extra beats, missed beats, long or short beats, and ectopic beats) were used.
Also, wavelet-based approaches are used for detecting and correcting premature ventricular contractions (PVC). Daubechies wavelets (Db4) is shown to be appropriate for denoised ECG reconstruction [73]. According to features of HRV, DWT hard thresholding strategy is adequate for time-domain measures but problematic for spectrum measures, yet the DWT detection algorithm combined with linear interpolation is an automated combination suitable for big datasets. Based on [73], the choice of decomposition level and mother wavelet are disadvantages of the wavelet method, but it is simple and robust to noise [72].
Regarding real-time correction of ECG heartbeats as a necessary prerequisite for online monitoring, a real-time automated point-process-based method for detection and correction of erroneous and ectopic beats (extra beat, missed beat, misplaced beat, two misplaced beats, and resetting ectopic beat) is proposed in [75]. The algorithm was tested over two datasets from the Fantasia normal rhythm database and the benchmark MIT-BIH. Comparing this method with 3 methods in [76-78], they found that their method represents an improvement with 99.985% specificity, 100% sensitivity to missed and additional beats, 96.01% sensitivity for artificially misplaced beats, and 98.73% positive prediction value for true arrhythmic occurrences.
The most commonly used interpolation methods for beat correction are: zero-order, first order (linear interpolation), spline and polynomial interpolations. Interpolation techniques that replace ectopic beats reduce HF content and suggested Lomb periodograms for spectral analysis that require no resampling, however choi et al, noticed that in most cases and studies, the use of interpolation reduces errors [89]. In another approach by Wen [89], a trend-predict correction method was adopted to detect ectopic beats. Comparing trend-predict correction method (TPC) with the spline interpolation method and middle value replacement, the result of the TPC method was better than the others. Also, if we have missing RR-interval [80] and [81] suggested using interpolation methods and bootstrapping. [80] found that data without using interpolation methods and bootstrapping is optimal for time domain HRV analysis. In another work, Dong et al. compared four methods to correct ectopic beats for the analysis sample entropy for continuous monitoring of physiological signal.: KeepSampEn (their new method), SkipSampEn (remove the missing values and connect the remaining points), LinearSampEn (linear interpolation) and BootSampEn (bootstrapping). They calculated percentage errors and made pairwise comparisons by paired T-test and came to the conclusion that their new method has the best performance in comparison to others.
Likewise beat correction, resampling is another consideration to reduce errors are RR-interval time series. In [82], different resampling frequencies of RR-interval time series are compared and they suggested a resampling rate of 4 Hz for HR < 90 bpm, 6 Hz for 90<HR<117 bpm and 8 Hz for HR>117 bpm. They used linear interpolation for resampling. The RR-interval resampling rate at 1 Hz and 4 Hz to validate the use of 1 Hz for wearable devices has been compared by authors in [83]. They suggested using 1 Hz linear interpolation, a rectangular window and the Yule-Walker method for estimating sympathetic activity. Also, they compared interpolation methods and found that linear interpolation and cubic spline interpolation showed the best results for time-domain and high frequency related variables, and nearest neighbor interpolation (zero order) showed the best results for low frequency components [84].
In order to evaluate and compare HRV detection and correction algorithms, a common way is to generate abnormal beats or simulate them by removing beats, and then calculate “false positive” and “accuracy” for each method. In addition, some HRV parameters (such as Mean RR, SDNN, RMSSD, LF power, and HF power [71]) to investigate the effects of different simulated artefacts and detection/removal approaches.
(v) Detrending
Studies have shown the non-stationary trend in RR-interval time series affects the frequency analysis of HRV. Thus, in order to get acceptable results in HRV analysis, detrending is of significantly importance. In the following, various types of trends and their effect on HRV in addition to trend removal methods are explained.
Comparing three methods of detrending RR-interval signals, Li L. et al. showed that detrending could reduce errors. The time cost of EMD and SPA were higher than the wavelet. Also, the Wavelet had better detrending performance [85]. Also, in [86] it has been shown that trends in signals can affect SDNN in time domain analysis and LF and VLF in frequency domain analysis more than others. Moreover, by comparing two methods, Ensemble EMD (EEMD) and SPA, it has been reported that EEMD has better performance in detrending the cosine trends [87].
References
[1] O. Kwon et al., “Electrocardiogram sampling frequency range acceptable for heart rate variability analysis,” Healthcare informatics research, vol. 24, no. 3, pp. 198-206, 2018.
[2] G. Pizzuti, S. Cifaldi, and G. Nolfe, “Digital sampling rate and ECG analysis,” Journal of biomedical engineering, vol. 7, no. 3, pp. 247-250, 1985.
[3] E. Ajdaraga and M. Gusev, “Analysis of sampling frequency and resolution in ECG signals,” in 2017 25th Telecommunication Forum (TELFOR), 2017, pp. 1-4: IEEE.
[4] M. Petelczyc, J. J. Gierałtowski, B. Żogała-Siudem, and G. Siudem, “Impact of observational error on heart rate variability analysis,” Heliyon, vol. 6, no. 5, p. e03984, 2020.
[5] G. G. Berntson et al., “Heart rate variability: origins, methods, and interpretive caveats,” Psychophysiology, vol. 34, no. 6, pp. 623-648, 1997.
[6] L. Hejjel and E. Roth, “What is the adequate sampling interval of the ECG signal for heart rate variability analysis in the time domain?,” Physiological measurement, vol. 25, no. 6, p. 1405, 2004.
[7] T. Ziemssen, J. Gasch, and H. Ruediger, “Influence of ECG sampling frequency on spectral analysis of RR intervals and baroreflex sensitivity using the EUROBAVAR data set,” Journal of clinical monitoring and computing, vol. 22, no. 2, pp. 159-168, 2008.
[8] S. Mahdiani, V. Jeyhani, M. Peltokangas, and A. Vehkaoja, “Is 50 Hz high enough ECG sampling frequency for accurate HRV analysis?,” in 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2015, pp. 5948-5951: IEEE.
[9] K. A. Sidek and I. Khalil, “Enhancement of low sampling frequency recordings for ECG biometric matching using interpolation,” Computer methods and programs in biomedicine, vol. 109, no. 1, pp. 13-25, 2013.
[10] A. Habib, C. Karmakar, and J. Yearwood, “Choosing a sampling frequency for ECG QRS detection using convolutional networks,” arXiv preprint arXiv:2007.02052, 2020.
[11] R. J. Ellis, B. Zhu, J. Koenig, J. F. Thayer, and Y. Wang, “A careful look at ECG sampling frequency and R-peak interpolation on short-term measures of heart rate variability,” Physiological measurement, vol. 36, no. 9, p. 1827, 2015.
[12] T. F. o. t. E. S. o. C. t. N. A. S. o. P. Electrophysiology, “Heart rate variability: standards of measurement, physiological interpretation, and clinical use,” Circulation, vol. 93, no. 5, pp. 1043-1065, 1996.
[13] S. Abboud and O. Barnea, “Errors due to sampling frequency of the electrocardiogram in spectral analysis of heart rate signals with low variability,” in Computers in Cardiology 1995, 1995, pp. 461-463: IEEE.
[14] M. D. MENZ, “Minimum sampling rate in electrocardiology,” Journal of Clinical Engineering, vol. 19, no. 5, pp. 386-394, 1994.
[15] S. Chatterjee, R. S. Thakur, R. N. Yadav, L. Gupta, and D. K. Raghuvanshi, “Review of noise removal techniques in ECG signals,” IET Signal Processing, vol. 14, no. 9, pp. 569-590, 2020.
[16] B. N. Singh and A. K. Tiwari, “Optimal selection of wavelet basis function applied to ECG signal denoising,” Digital signal processing, vol. 16, no. 3, pp. 275-287, 2006.
[17] A. Velayudhan and S. Peter, “Noise analysis and different denoising techniques of ECG signal-a survey,” IOSR journal of electronics and communication engineering, vol. 1, no. 1, pp. 40-44, 2016.
[18] H. D. Hesar and M. Mohebbi, “An adaptive Kalman filter bank for ECG denoising,” IEEE journal of biomedical and health informatics, vol. 25, no. 1, pp. 13-21, 2020.
[19] J. Wang et al., “Adversarial de-noising of electrocardiogram,” Neurocomputing, vol. 349, pp. 212-224, 2019.
[20] H. D. Hesar and M. Mohebbi, “An adaptive particle weighting strategy for ECG denoising using marginalized particle extended Kalman filter: An evaluation in arrhythmia contexts,” IEEE journal of biomedical and health informatics, vol. 21, no. 6, pp. 1581-1592, 2017.
[21] H.-T. Chiang, Y.-Y. Hsieh, S.-W. Fu, K.-H. Hung, Y. Tsao, and S.-Y. Chien, “Noise reduction in ECG signals using fully convolutional denoising autoencoders,” IEEE Access, vol. 7, pp. 60806-60813, 2019.
[22] M. Rakshit and S. Das, “Hybrid approach for ECG signal enhancement using dictionary learning-based sparse representation,” IET Science, Measurement & Technology, vol. 13, no. 3, pp. 381-391, 2019.
[23] O. Hernández and E. Olvera, “Noise cancellation on ECG and heart rate signals using the undecimated wavelet transform,” in 2009 International Conference on eHealth, Telemedicine, and Social Medicine, 2009, pp. 145-150: IEEE.
[24] J. Lee, D. D. McManus, S. Merchant, and K. H. Chon, “Automatic motion and noise artifact detection in Holter ECG data using empirical mode decomposition and statistical approaches,” IEEE Transactions on Biomedical Engineering, vol. 59, no. 6, pp. 1499-1506, 2011.
[25] M. Blanco-Velasco, B. Weng, and K. E. Barner, “ECG signal denoising and baseline wander correction based on the empirical mode decomposition,” Computers in biology and medicine, vol. 38, no. 1, pp. 1-13, 2008.
[26] A. Lakhe et al., “Spectral trimming technique: a new approach for suppressing motion artefacts in stress electrocardiography,” Journal of Medical Engineering & Technology, vol. 44, no. 6, pp. 338-345, 2020.
[27] M.-B. Hossain, J. Lázaro, Y. Noh, and K. H. Chon, “Denoising wearable armband ECG data using the variable frequency complex demodulation technique,” in 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2020, pp. 592-595: IEEE.
[28] C. Chandrakar and M. Sharma, “A real time approach for ECG signal denoising and smoothing using adaptive window technique,” in 2014 9th International Conference on Industrial and Information Systems (ICIIS), 2014, pp. 1-6: IEEE.
[29] H. Kim et al., “Motion artifact removal using cascade adaptive filtering for ambulatory ECG monitoring system,” in 2012 IEEE Biomedical Circuits and Systems Conference (BioCAS), 2012, pp. 160-163: IEEE.
[30] M. El hanine, E. Abdelmounim, R. Haddadi, and A. Belaguid, “Real Time EMG Noise Cancellation from ECG Signals using Adaptive Filtering,” in Proceedings of the 2nd International Conference on Computing and Wireless Communication Systems, 2017, pp. 1-6.
[31] P. S. Addison, “Wavelet transforms and the ECG: a review,” Physiological measurement, vol. 26, no. 5, p. R155, 2005.
[32] O. El B’charri, R. Latif, K. Elmansouri, A. Abenaou, and W. Jenkal, “ECG signal performance de-noising assessment based on threshold tuning of dual-tree wavelet transform,” Biomedical engineering online, vol. 16, no. 1, pp. 1-18, 2017.
[33] D. Berwal, C. Vandana, S. Dewan, C. Jiji, and M. S. Baghini, “Motion artifact removal in ambulatory ECG signal for heart rate variability analysis,” IEEE Sensors Journal, vol. 19, no. 24, pp. 12432-12442, 2019.
[34] S. Nagai, D. Anzai, and J. Wang, “Motion artefact removals for wearable ECG using stationary wavelet transform,” Healthcare technology letters, vol. 4, no. 4, pp. 138-141, 2017.
[35] D. V. Bhoraniya and R. K. Kher, “Motion artifacts extraction using DWT from ambulatory ECG (A-ECG),” in 2014 International Conference on Communication and Signal Processing, 2014, pp. 1567-1571: IEEE.
[36] S.-W. Chen, H.-C. Chen, and H.-L. Chan, “A real-time QRS detection method based on moving-averaging incorporating with wavelet denoising,” Computer methods and programs in biomedicine, vol. 82, no. 3, pp. 187-195, 2006.
[37] S. Mukhopadhyay, S. Biswas, A. B. Roy, and N. Dey, “Wavelet based QRS complex detection of ECG signal,” arXiv preprint arXiv:1209.1563, 2012.
[38] H. Yanık, E. Değirmenci, B. Büyükakıllı, D. Karpuz, O. H. Kılınç, and S. Gürgül, “Electrocardiography (ECG) analysis and a new feature extraction method using wavelet transform with scalogram analysis,” Biomedical Engineering/Biomedizinische Technik, vol. 65, no. 5, pp. 543-556, 2020.
[39] Z. Wang, J. Zhu, T. Yan, and L. Yang, “A new modified wavelet-based ECG denoising,” Computer Assisted Surgery, vol. 24, no. sup1, pp. 174-183, 2019.
[40] M. Boutaa, F. Bereksi-Reguig, and S. Debbal, “ECG signal processing using multiresolution analysis,” Journal of medical engineering & technology, vol. 32, no. 6, pp. 466-478, 2008.
[41] S. Shimauchi, K. Eguchi, R. Aoki, M. Fukui, and N. Harada, “RR Interval Estimation for Wearable Electrocardiogram Based on Single Complex Wavelet Filtering and Morphology-Based Peak Selection,” IEEE Access, vol. 9, pp. 60802-60827, 2021.
[42] J. Lázaro, N. Reljin, M.-B. Hossain, Y. Noh, P. Laguna, and K. H. Chon, “Wearable armband device for daily life electrocardiogram monitoring,” IEEE Transactions on Biomedical Engineering, vol. 67, no. 12, pp. 3464-3473, 2020.
[43] N. Reljin, J. Lazaro, M. B. Hossain, Y. S. Noh, C. H. Cho, and K. H. Chon, “Using the redundant convolutional encoder–decoder to denoise QRS complexes in ECG signals recorded with an armband wearable device,” Sensors, vol. 20, no. 16, p. 4611, 2020.
[44] D. Berwal, A. Kumar, and Y. Kumar, “Design of high performance QRS complex detector for wearable healthcare devices using biorthogonal spline wavelet transform,” ISA transactions, vol. 81, pp. 222-230, 2018.
[45] M.-B. Hossain, S. K. Bashar, J. Lazaro, N. Reljin, Y. Noh, and K. H. Chon, “A robust ECG denoising technique using variable frequency complex demodulation,” Computer Methods and Programs in Biomedicine, vol. 200, p. 105856, 2021.
[46] T. Seol, S. Lee, and J. Lee, “A wearable electrocardiogram monitoring system robust to motion artifacts,” in 2018 International SoC Design Conference (ISOCC), 2018, pp. 241-242: IEEE.
[47] R. Poli, S. Cagnoni, and G. Valli, “Genetic design of optimum linear and nonlinear QRS detectors,” IEEE Transactions on Biomedical Engineering, vol. 42, no. 11, pp. 1137-1141, 1995.
[48] P. S. Hamilton and W. J. Tompkins, “Quantitative Investigation of QRS Detection Rules Using the MIT/BIH Arrhythmia Database,” IEEE Transactions on Biomedical Engineering, vol. BME-33, no. 12, pp. 1157-1165, 1986.
[49] F. Zhang and Y. Lian, “QRS Detection Based on Multiscale Mathematical Morphology for Wearable ECG Devices in Body Area Networks,” IEEE Transactions on Biomedical Circuits and Systems, vol. 3, no. 4, pp. 220-228, 2009.
[50] C. Ieong et al., “A 0.83-
$mu {rm W}%%EDITORCONTENT%%lt;/tex>
QRS Detection Processor Using Quadratic Spline Wavelet Transform for Wireless ECG Acquisition in 0.35- $mu{rm m}%%EDITORCONTENT%%lt;/tex> CMOS,” IEEE Transactions on Biomedical Circuits and Systems, vol. 6, no. 6, pp. 586-595, 2012.
[51] G. Nallathambi and J. C. Príncipe, “Integrate and fire pulse train automaton for QRS detection,” (in eng), IEEE Trans Biomed Eng, vol. 61, no. 2, pp. 317-26, Feb 2014.
[52] J. Pan and W. J. Tompkins, “A real-time QRS detection algorithm,” IEEE transactions on biomedical engineering, no. 3, pp. 230-236, 1985.
[53] C. Li, C. Zheng, and C. Tai, “Detection of ECG characteristic points using wavelet transforms,” IEEE Transactions on biomedical Engineering, vol. 42, no. 1, pp. 21-28, 1995.
[54] V. X. Afonso, W. J. Tompkins, T. Q. Nguyen, and S. Luo, “Filter bank-based ECG beat detection,” in Proceedings of 18th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 1996, vol. 3, pp. 1037-1038: IEEE.
[55] J. Moraes, M. Freitas, F. Vilani, and E. Costa, “A QRS complex detection algorithm using electrocardiogram leads,” in Computers in cardiology, 2002, pp. 205-208: IEEE.
[56] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna, “A wavelet-based ECG delineator: evaluation on standard databases,” IEEE Transactions on Biomedical Engineering, vol. 51, no. 4, pp. 570-581, 2004.
[57] F. Chiarugi, V. Sakkalis, D. Emmanouilidou, T. Krontiris, M. Varanini, and I. Tollis, “Adaptive threshold QRS detector with best channel selection based on a noise rating system,” in 2007 Computers in Cardiology, 2007, pp. 157-160: IEEE.
[58] H. Zheng and J. Wu, “Real-time QRS detection method,” in HealthCom 2008-10th International Conference on e-health Networking, Applications and Services, 2008, pp. 169-170: IEEE.
[59] N. M. Arzeno, Z.-D. Deng, and C.-S. Poon, “Analysis of first-derivative based QRS detection algorithms,” IEEE Transactions on Biomedical Engineering, vol. 55, no. 2, pp. 478-484, 2008.
[60] A. A. R. Bsoul, S.-Y. Ji, K. Ward, and K. Najarian, “Detection of P, QRS, and T components of ECG using wavelet transformation,” in 2009 ICME International Conference on Complex Medical Engineering, 2009, pp. 1-6: IEEE.
[61] W. A. Engelse and C. Zeelenberg, “A single scan algorithm for QRS-detection and feature extraction,” Computers in cardiology, vol. 6, no. 1979, pp. 37-42, 1979.
[62] D. S. Benitez, P. Gaydecki, A. Zaidi, and A. Fitzpatrick, “A new QRS detection algorithm based on the Hilbert transform,” in Computers in Cardiology 2000. Vol. 27 (Cat. 00CH37163), 2000, pp. 379-382: IEEE.
[63] M. Elgendi, “Fast QRS detection with an optimized knowledge-based method: Evaluation on 11 standard ECG databases,” PloS one, vol. 8, no. 9, p. e73557, 2013.
[64] R. Thirrunavukkarasu, T. Meeradevi, A. Ravi, D. Ganesan, and G. P. Vadivel, “Detection R peak in electrocardiogram signal using daubechies wavelet transform and shannon’s energy envelope,” in 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS), 2019, pp. 1044-1048: IEEE.
[65] S. Pal and M. Mitra, “Detection of ECG characteristic points using multiresolution wavelet analysis based selective coefficient method,” Measurement, vol. 43, no. 2, pp. 255-261, 2010.
[66] H. Dickhaus, L. Khadra, and J. Brachmann, “Quantification of ECG late potentials by wavelet transformation,” Computer methods and programs in biomedicine, vol. 43, no. 3-4, pp. 185-192, 1994.
[67] M. V. G. Miranda, I. P. V. Espinosa, and M. J. F. Calero, “ECG signal features extraction,” in 2016 IEEE Ecuador Technical Chapters Meeting (ETCM), 2016, pp. 1-6: IEEE.
[68] I. I. Christov, “Real time electrocardiogram QRS detection using combined adaptive threshold,” Biomedical engineering online, vol. 3, no. 1, pp. 1-9, 2004.
[69] V. Kalidas and L. Tamil, “Real-time QRS detector using stationary wavelet transform for automated ECG analysis,” in 2017 IEEE 17th International Conference on Bioinformatics and Bioengineering (BIBE), 2017, pp. 457-461: IEEE.
[70] Y. Goyal and A. Jain, “Study of HRV dynamics and comparison using wavelet analysis and Pan Tompkins algorithm,” in 2012 ASE/IEEE International Conference on BioMedical Computing (BioMedCom), 2012, pp. 43-49: IEEE.
[71] K. Zhao, Y. Li, G. Wang, Y. Pu, and Y. Lian, “A robust QRS detection and accurate R-peak identification algorithm for wearable ECG sensors,” Science China Information Sciences, pp. 1-1, 2021.
[72] T. Bruggemann, D. Andresen, D. Weiss, J. Rose, A. Chorianopoulos, and R. Schroder, “Heart rate variability: how to exclude extrasystoles from the analysis?,” in Proceedings of Computers in Cardiology Conference, 1993, pp. 467-470: IEEE.
[73] D. Nabil and F. Bereksi Reguig, “Ectopic beats detection and correction methods: A review,” Biomedical Signal Processing and Control, vol. 18, pp. 228-244, 2015/04/01/ 2015.
[74] D. Nabil and F. B. Reguig, “Ectopic beats detection and correction methods: A review,” Biomedical Signal Processing and Control, vol. 18, pp. 228-244, 2015.
[75] L. Citi, E. N. Brown, and R. Barbieri, “A real-time automated point-process method for the detection and correction of erroneous and ectopic heartbeats,” IEEE transactions on biomedical engineering, vol. 59, no. 10, pp. 2828-2837, 2012.
[76] J. Mateo and P. Laguna, “Analysis of heart rate variability in the presence of ectopic beats using the heart timing signal,” IEEE Transactions on biomedical engineering, vol. 50, no. 3, pp. 334-343, 2003.
[77] J. Rand, A. Hoover, S. Fishel, J. Moss, J. Pappas, and E. Muth, “Real-time correction of heart interbeat intervals,” IEEE Transactions on Biomedical Engineering, vol. 54, no. 5, pp. 946-950, 2007.
[78] G. G. Berntson, K. S. Quigley, J. F. Jang, and S. T. Boysen, “An approach to artifact identification: Application to heart period data,” Psychophysiology, vol. 27, no. 5, pp. 586-598, 1990.
[79] F. Wen and F.-t. He, “An efficient method of addressing ectopic beats: new insight into data preprocessing of heart rate variability analysis,” Journal of Zhejiang University SCIENCE B, vol. 12, no. 12, pp. 976-982, 2011.
[80] K. K. Kim, H. J. Baek, Y. G. Lim, and K. S. Park, “Effect of missing RR-interval data on nonlinear heart rate variability analysis,” Computer methods and programs in biomedicine, vol. 106, no. 3, pp. 210-218, 2012.
[81] X. Dong et al., “An improved method of handling missing values in the analysis of sample entropy for continuous monitoring of physiological signals,” Entropy, vol. 21, no. 3, p. 274, 2019.
[82] D. Singh, K. Vinod, and S. Saxena, “Sampling frequency of the RR interval time series for spectral analysis of heart rate variability,” Journal of medical engineering & technology, vol. 28, no. 6, pp. 263-272, 2004.
[83] H.-K. Chen, Y.-F. Hu, and S.-F. Lin, “Methodological considerations in calculating heart rate variability based on wearable device heart rate samples,” Computers in biology and medicine, vol. 102, pp. 396-401, 2018.
[84] A. Choi and H. Shin, “Quantitative Analysis of the Effect of an Ectopic Beat on the Heart Rate Variability in the Resting Condition,” (in eng), Frontiers in physiology, vol. 9, pp. 922-922, 2018.
[85] L. Li, K. Li, C. Liu, and C.-Y. Liu, “Comparison of detrending methods in spectral analysis of heart rate variability,” Res J Appl Sci Eng Technol, vol. 3, no. 9, pp. 1014-1021, 2011.
[86] M. P. Tarvainen, P. O. Ranta-Aho, and P. A. Karjalainen, “An advanced detrending method with application to HRV analysis,” IEEE transactions on biomedical engineering, vol. 49, no. 2, pp. 172-175, 2002.
[87] C. Zeng and X. Xu, “Evaluation of Detrending Method Based on Ensemble Empirical Mode Decomposition for HRV Analysis,” J. Comput., vol. 9, no. 6, pp. 1325-1332, 2014.
[88] X. An and G. K Stylios, “Comparison of motion artefact reduction methods and the implementation of adaptive motion artefact reduction in wearable electrocardiogram monitoring,” Sensors, vol. 20, no. 5, p. 1468, 2020.
