How to fix a wandering baseline on an ECG

4.3.6.5 Respiratory rate

Respiratory rate (RR), the number of breaths taken in a minute, is used for diagnosis and prognosis in a range of clinical settings (Charlton et al., 2018a). RR is a marker for clinical deteriorations in acutely-ill hospitalized patients, and an elevated RR is predictive of adverse events such as cardiac arrest (Schein et al., 1990) and death (Duckitt et al., 2007). RR is also used in the identification of pneumonia (Schein et al., 1990) and sepsis (Seymour et al., 2016). Despite its importance, RR is usually measured through manual breath counting outside of intensive care. This process is time-consuming and inaccurate (Lovett et al., 2005; Philip et al., 2015). Furthermore, existing methods for RR monitoring in wearables often require equipment such as chest-bands. Consequently, there is great potential for an unobtrusive method of RR monitoring using routine sensors such as a PPG sensor.

The PPG signal is subtly modulated by breathing, providing an opportunity to estimate RR from it. There are three main modulations, as shown in Fig. 4.23A: baseline wander (BW), amplitude modulation (AM), and frequency modulation (FM) (Charlton et al., 2017; Liu et al., 2020a). Most RR algorithms follow a standard structure consisting of three fundamental stages (Charlton et al., 2016), as shown in Fig. 4.23A:

Extraction of respiratory signals: One or more respiratory signals are extracted from the PPG signal. This is beneficial as the extracted respiratory signals are dominated by a respiratory modulation, making it easier to estimate RR from them than from the original PPG signal. Broadly, there are two approaches for extracting respiratory signals: feature-based extraction and filter-based extraction (Charlton et al., 2018a). In feature-based extraction, a feature is extracted from each pulse wave, such as the pulse wave amplitude (see Fig. 4.23B). In filter-based extraction, filtering is used to extract a respiratory signal, such as a band-pass filter with a passband corresponding to the range of plausible respiratory frequencies.

RR estimation: RR is estimated from a respiratory signal typically using either a time- or frequency-domain technique (see Fig. 4.23B). For instance, a time-domain technique could consist of identifying peaks in the signal (indicating breaths), and calculating RR from the number of peaks in a specified time. A frequency-domain technique could consist of calculating the frequency spectrum of the respiratory signal, and obtaining the RR as the frequency corresponding to the maximum power.

Fusion of RR estimates: Optionally, several RR estimates can be obtained from different respiratory signals and then fused to provide a single RR estimate. For instance, in Karlen et al. (2013), a RR estimate was calculated as the mean of estimates obtained from respiratory signals indicative of BW, AM, and FM (see Fig. 4.23B).

The techniques used in RR algorithms are described in further detail in Charlton et al. (2018a, 2016). Recently, respiratory quality indices have been proposed to determine whether the respiratory modulations in a segment of PPG signal are sufficiently strong to accurately estimate RR (Birrenkott et al., 2018).

The performance of algorithms to estimate RR from the PPG has been assessed in several studies. In studies of several algorithms from the literature, the best-performing algorithm was found to have limits of agreement of -5.1 to 7.2 breaths per minute when tested on healthy subjects in controlled conditions (Charlton et al., 2016), and -9.2 to 8.8 breaths per minute when tested on critically-ill hospitalized patients (Charlton, 2017). The limits of agreement indicate the ranges in which 95% of errors are expected to lie. Future research may provide additional evidence on the performance of algorithms in target settings, such as when used with pulse oximeters or smartphones for spot-check assessments in home monitoring, or when used in wearables for continuous monitoring during daily living.

6.5 Noise and Artifacts

An important reason behind the success of computer-based ECG analysis is the capability to improve poor signal quality by means of signal processing algorithms. This result has been achieved thanks to good knowledge of not only signal properties but also noise properties. It is therefore important to become familiarized with the most common types of noise and artifacts in the ECG before addressing methods in the next chapter which compensate for their presence. Below follows a list of common noncardiac noise sources of which the first three are of technical origin whereas the fourth is of physiological origin. Even parts of the cardiac activity can sometimes be viewed as a source of noise when detecting QRS complexes, see Section 7.4.

Baseline wander is an extraneous, low-frequency activity in the ECG (Figure 6.17(a)) which may interfere with the signal analysis, rendering the clinical interpretation inaccurate and misleading. For example, ECG measurements defined with reference to the isoelectric line cannot be computed because the isoelectric line is no longer well-defined. Baseline wander, which is often exercise-induced, may result from a variety of noise sources including perspiration, respiration, body movements, and poor electrode contact. The magnitude of the undesired wander may exceed the amplitude of the QRS complex by several times. Its spectral content is usually confined to an interval well below 1 Hz, but it may contain higher frequencies during strenuous exercise. Signal processing techniques for the removal of baseline wander are presented in detail in Section 7.1.

Electrode motion artifacts are mainly caused by skin stretching which alters the impedance of the skin around the electrode. Motion artifacts resemble the signal characteristics of baseline wander, but are more problematic to combat since their spectral content considerably overlaps that of the PQRST complex. They occur mainly in the range from 1 to 10 Hz [31, 32. In the ECG, these artifacts are manifested as large-amplitude waveforms which are sometimes mistaken for QRS complexes (Figure 6.17(b)). Electrode motion artifacts are particularly troublesome in the context of ambulatory ECG monitoring where they constitute the main source of falsely detected heartbeats.

Powerline interference (50/60 Hz) is caused by improper grounding of the ECG equipment and interference from nearby equipment [33]. Such interference can be removed in many situations by means of linear or nonlinear filtering, see Section 7.2.

The electrical activity of skeletal muscles during periods of contraction causes electromyographic noise (EMG noise), commonly seen in ECGs recorded during ambulatory monitoring or exercise. The main characteristics of such noise have already been presented on page 74 in connection with artifact rejection in EEG signal processing (different muscles are, however, active in producing the noise which corrupts the ECG signal), see also Section 5.1. Electromyographic noise can either be intermittent in nature, e.g., due to a sudden body movement (Figure 6.17(c)), or have more stationary noise properties. The frequency components of EMG considerably overlap those of the QRS complex while also extending into higher frequencies. As a result, difficulties in removing EMG noise from the EEG signal without introducing distortion are unfortunately also present in ECG signal processing. Some approaches that deal with EMG noise are briefly presented in Section 7.3. The influence of EMG noise can also be reduced by ensemble averaging when the recurrent property of the heartbeats can be exploited.

Respiratory activity influences electrocardiographic measurements not only through heart rate but also through beat morphology. Such beat-tobeat variations in morphology are caused by chest movements, changes in the position of the heart, and changes in lung conductivity [34, 35. During the respiratory cycle, the vector describing the dominant direction of the electrical wave propagation changes so that variations in beat morphology arise. Figure 6.17(d) presents an ECG with a pronounced variation in QRS amplitude being induced by respiration; in this example, the period length of a breath is approximately 5 s, suggesting that the subject is breathing at a rate of 12 breaths/minute. Although variations in QRS amplitude represent an undesirable signal characteristic, it may be exploited for estimation of the respiratory frequency [36−39].

3.3 Conclusions

In this chapter, we reviewed the theory concerned with the spectral content of the ECG signal and how the noise spectra impacts ECG content, important concepts as signal-to-noise ratio (SNR), magnitude, phase response, and power spectrum of a signal. We emphasized that ECG filtering covers three specific applications: baseline wander suppression, line frequency rejection, and muscle artifact reduction. Then we described and implemented, through computing simulations, several techniques based on linear and time-variant filtering, polynomial fitting, wavelet filtering and empirical mode decomposition. It is clear that some classic filtering techniques introduce significant distortion within the filtered ECG signal, are poorly inefficient, and unable to remove simple noise classes, such as the baseline wandering. Some emphasis should be given to Wavelet transform and empirical mode decomposition, which block the most common types of noise with a minimum distortion. However, depending on the final application, a certain level of distortion may be tolerated. Therefore the most appropriate filtering technique to be applied depends on what we want to extract from the ECG signal. Despite the considerable ECG signal denoising evolution, a high research demand for the analysis of the impact of different filtering techniques over specific tasks within ECG signal processing, such as wave segmentation, feature extraction, and arrhythmia classification is apparent.

2.2 Encoding (NRZ, NRZI, MANCHESTER, 4B/5B)

The first step in turning nodes and links into usable building blocks is to understand how to connect them in such a way that bits can be transmitted from one node to the other. As mentioned in the preceding section, signals propagate over physical links. The task, therefore, is to encode the binary data that the source node wants to send into the signals that the links are able to carry and then to decode the signal back into the corresponding binary data at the receiving node. We ignore the details of modulation and assume we are working with two discrete signals: high and low. In practice, these signals might correspond to two different voltages on a copper-based link or two different power levels on an optical link.

Most of the functions discussed in this chapter are performed by a network adaptor —a piece of hardware that connects a node to a link. The network adaptor contains a signalling component that actually encodes bits into signals at the sending node and decodes signals into bits at the receiving node. Thus, as illustrated in Figure 2.3, signals travel over a link between two signalling components, and bits flow between network adaptors.

Let's return to the problem of encoding bits onto signals. The obvious thing to do is to map the data value 1 onto the high signal and the data value 0 onto the low signal. This is exactly the mapping used by an encoding scheme called, cryptically enough, non-return to zero (NRZ). For example, Figure 2.4 schematically depicts the NRZ-encoded signal (bottom) that corresponds to the transmission of a particular sequence of bits (top).

The problem with NRZ is that a sequence of several consecutive 1s means that the signal stays high on the link for an extended period of time; similarly, several consecutive 0s means that the signal stays low for a long time. There are two fundamental problems caused by long strings of 1s or 0s. The first is that it leads to a situation known as baseline wander. Specifically, the receiver keeps an average of the signal it has seen so far and then uses this average to distinguish between low and high signals. Whenever the signal is significantly lower than this average, the receiver concludes that it has just seen a 0; likewise, a signal that is significantly higher than the average is interpreted to be a 1. The problem, of course, is that too many consecutive 1s or 0s cause this average to change, making it more difficult to detect a significant change in the signal.

The second problem is that frequent transitions from high to low and vice versa are necessary to enable clock recovery. Intuitively, the clock recovery problem is that both the encoding and the decoding processes are driven by a clock—every clock cycle the sender transmits a bit and the receiver recovers a bit. The sender's and the receiver's clocks have to be precisely synchronized in order for the receiver to recover the same bits the sender transmits. If the receiver's clock is even slightly faster or slower than the sender's clock, then it does not correctly decode the signal. You could imagine sending the clock to the receiver over a separate wire, but this is typically avoided because it makes the cost of cabling twice as high. So, instead, the receiver derives the clock from the received signal—the clock recovery process. Whenever the signal changes, such as on a transition from 1 to 0 or from 0 to 1, then the receiver knows it is at a clock cycle boundary, and it can resynchronize itself. However, a long period of time without such a transition leads to clock drift. Thus, clock recovery depends on having lots of transitions in the signal, no matter what data is being sent.

One approach that addresses this problem, called non-return to zero inverted (NRZI), has the sender make a transition from the current signal to encode a 1 and stay at the current signal to encode a 0. This solves the problem of consecutive 1s, but obviously does nothing for consecutive 0s. NRZI is illustrated in Figure 2.5. An alternative, called Manchester encoding, does a more explicit job of merging the clock with the signal by transmitting the exclusive OR of the NRZ-encoded data and the clock. (Think of the local clock as an internal signal that alternates from low to high; a low/high pair is considered one clock cycle.) The Manchester encoding is also illustrated in Figure 2.5. Observe that the Manchester encoding results in 0 being encoded as a low-to-high transition and 1 being encoded as a high-to-low transition. Because both 0s and 1s result in a transition to the signal, the clock can be effectively recovered at the receiver. (There is also a variant of the Manchester encoding, called Differential Manchester, in which a 1 is encoded with the first half of the signal equal to the last half of the previous bit's signal and a 0 is encoded with the first half of the signal opposite to the last half of the previous bit's signal.)

The problem with the Manchester encoding scheme is that it doubles the rate at which signal transitions are made on the link, which means that the receiver has half the time to detect each pulse of the signal. The rate at which the signal changes is called the link's baud rate. In the case of the Manchester encoding, the bit rate is half the baud rate, so the encoding is considered only 50% efficient. Keep in mind that if the receiver had been able to keep up with the faster baud rate required by the Manchester encoding in Figure 2.5, then both NRZ and NRZI could have been able to transmit twice as many bits in the same time period.

A final encoding that we consider, called 4B/5B, attempts to address the inefficiency of the Manchester encoding without suffering from the problem of having extended durations of high or low signals. The idea of 4B/5B is to insert extra bits into the bit stream so as to break up long sequences of 0s or 1s. Specifically, every 4 bits of actual data are encoded in a 5-bit code that is then transmitted to the receiver; hence, the name 4B/5B. The 5-bit codes are selected in such a way that each one has no more than one leading 0 and no more than two trailing 0s. Thus, when sent back-to-back, no pair of 5-bit codes results in more than three consecutive 0s being transmitted. The resulting 5-bit codes are then transmitted using the NRZI encoding, which explains why the code is only concerned about consecutive 0s—NRZI already solves the problem of consecutive 1s. Note that the 4B/5B encoding results in 80% efficiency.

Table 2.2 gives the 5-bit codes that correspond to each of the 16 possible 4-bit data symbols. Notice that since 5 bits are enough to encode 32 different codes, and we are using only 16 of these for data, there are 16 codes left over that we can use for other purposes. Of these, code 11111 is used when the line is idle, code 00000 corresponds to when the line is dead, and 00100 is interpreted to mean halt. Of the remaining 13 codes, 7 of them are not valid because they violate the “one leading 0, two trailing 0s,” rule, and the other 6 represent various control symbols. As we will see later in this chapter, some framing protocols make use of these control symbols.

Table 2.2. 4B/5B Encoding

4-Bit Data Symbol5-Bit Code000011110000101001001010100001110101010001010010101011011001110011101111100010010100110011101010110101110111110011010110111011111011100111111101

3.3.4 On-off keying modulation formats

3.3.4.1 NRZ modulated signal

The non-return-to-zero (NRZ) modulation consists in turning on the light during the total bit duration, when the symbol to be transmitted is “1” and to suppress it completely, when the symbol to be transmitted is “0”. The simplest and most easily modeled pulse shapes shown in Figure 3.3a have a rectangular profile in the form:

Figure 3.3. Coding of the binary sequence “0101101” by using (a) NRZ, (b) rectangular shaped RZ modulation formats.

(3.29)a(t)={1for0

with the square Fourier transform:

(3.30)|a˜(f)|2=(TsinπTfπTf)2

More sophisticated shapes of the pulse may be chosen for various purposes. A Gaussian pulse is very convenient for modeling and is also the result of a multi-filtering effect. It is characterized by time and frequency profiles:

(3.31)a(t)=exp(−t22τ2)and|a˜(f)|2=2πτ2exp[−(2πντ)2]

Raised cosine pulses, minimizing the frequency over shots of the spectral density, may be used for spectral shaping optimization and intersymbol interference reduction.

The general expression for the baseband NRZ modulation spectrum is obtained by reporting Eq. 3.31 in Eq. 3.29, which can be simplified to:

(3.32)SA(f)=T4(sinπTfπTf)2+14δ(f)

Figure 3.4a presents the normalized power spectrum, for NRZ modulation format. When an actual spectrum is observed, the relative height of the discrete part of the spectrum depends on the setting of the spectrum analyzer. For this reason the continuous part and the discrete part of the spectrum are represented separately. As a result of an average transmission over half of the time, the normalized optical power is half of its peak value.

Figure 3.4. (a) Continuous and (b) discrete parts of the normalized power spectrum for NRZ (ε=1) and RZ (ε=0.5) modulation formats with the same value of the pulse energy.

The two terms in Eq. 3.32 have equal contributions to the total power. In addition, half of the available power is wasted in the noninformative discrete component of the spectrum, which corresponds to the DC value of the modulated power. However, this value, obtained for NRZ coding, is the maximum value that can be achieved by ASK modulation.

The low spectral spread of NRZ coding makes it less sensitive to chromatic dispersion, but the time profile of the pulses makes this type of modulation very sensitive to system impairment, because of consecutive pulse overlap and intersymbol interference. A long sequence of identical symbols leads to a transmitted signal without any information on the digital period and phase, making the synchronization at the receiver difficult. Furthermore the mean value of the binary sequence changes as a function of the transmitted data, producing the so-called baseline wander, making the electronic process at the receiver more difficult. The very sharp leading and trailing edges of the pulse may be associated with high-frequency chirping, depending on the modulator type used.

3.2 Baseline wander removal

Pulse signals are sometimes accompanied by nonlinear baseline wanders because external disturbances, such as talking, respiration, and other involuntary wrist movements, are likely to be recorded by pulse-sensing units while implementing pulse acquisition. Zschocke et al. [70] found that respiration causes tiny periodic rotations of the wrist and extracted the respiratory signals.

Fig. 8(a) presents pulse signals with drifts. The split points of a single period are not horizontally linear, this typical drift is likely to be caused by respiration. The wanders caused by respiration are close to the Sin curve with a small amplitude [63]. However, the wanders caused by unconscious movement have a large amplitude and irregular rhythm in comparison with the true pulse signal [71].

The removal of baseline drifts is another preprocessing step required for the correct measurement of diagnostic pulse characteristics [72]. Simple baseline drifts were removed by subtracting the estimated drifts from the original pulse signal [68]. To removal complex wanders caused by certain factors, the Biorthogonal 3.1 wavelet was employed to stabilize the baseline [5]. Wang et al. [73] corrected the baseline wanders using a dual-tree complex wavelet transform and cubic spline interpolation.

To adaptively eliminate the different types of wanders, Xu et al. [74] denoised the pulse signal using a discrete Meyer wavelet filter and cubic spline estimation based on an energy ratio threshold. Considering the valuable information, Zhang et al. [75] reconstructed the clean pulse signal with empirical mode decomposition (EMD), which decomposes the raw signal into several components and then removes unrelated drifts. The comparisons reported in Ref. [76] suggest the wavelet transform technique is more effective for drift removal, as it gives better results with less computation.

9 Conclusions

The ECG contains different noise components like baseline wander, power line interference, muscle and movement artefacts, and electrode contact noise. But in most of the articles which are reviewed in the survey, the linear methods have provided a good performance. However there are not much work performed to study the classification performance in the presence of natural noise processes. In summary there is a need for testing both linear and non-linear methods on noisy data. This is a gap in the literature which needs to be supported in the future. However the authors intuition is that these non-linear methods may provide a good performance on noisy data than linear methods. There are studies to support some nonlinear methods such as higher order spectra are noise immune and robust in the presence of noise. Unique bispectrum, bicoherence and RP plots for each ECG class can be proposed. These plots can be used to find the efficacy of drugs and cardiac health. The signatures extracted from these plots will be unique and characterize the particular class. These features coupled with robust classifier can yield an accurate CAD system. The performance of the system can be improved by using a combination of linear and nonlinear domain features coupled with robust classifiers.

2.3.2 Indirect measurement of respiration

Breathing is physiologically connected with cardiovascular activities. Decades of research have revealed the indirect modulation of respiration in cardiac measurements such as ECG and PPG. These modulations can be divided into three types, namely baseline wander (BW), amplitude modulation (AM), and frequency modulation (FM) [57,72]. Baseline wander (BW) refers to the slow changes of signal baseline, which is usually discarded for heart analysis. AM and FM indicate the amplitude and frequency of cardiac-related peaks (QRS for ECG, systolic peaks for PPG), showing a high correlation with respiratory waveforms. While AM stems from changes in ECG recording condition (i.e., lower conductivity during inflation), FM originates from a common control of breathing and cardiac rhythms.

The standard process of acquiring breathing information comprises the extraction of respiratory signals (waveforms) and estimation of RR based on the pre-processing data after AM and FM. In this process, the critical point is to obtain robust observation of continuous breathing waveforms. Although accurate extraction can be achieved with single-mode measurement using PPG [73], the research consensus suggests that information confusion is required at both the waveform extraction and RR estimation stages, using multi-mode measurements and multiple modulations [57].

A significant advantage of using indirect measurement for respiration is that these techniques can be easily and immediately deployed into commercial devices (PPG or ECG function enabled) with only tiny firmware modifications. Therefore, PPG and ECG based respiratory monitoring are especially suitable for health monitoring of patients with chronic cardiopulmonary diseases. However, the inherent problem of low signal quality is a significant challenge in most cases. Recent efforts have introduced signal quality evaluation indexes to deal with related issues [74–76]. Orphanidou et al. have used heartbeat features and template matching to assess signal quality, automatically labeling the ECG/PPG signal as acceptable or unacceptable [77]. Such methods can significantly reduce the false alarms resulting from low quality signals. In turn, these methods can improve the battery life of wearable devices by reducing energy consumption. The development of assessment methods is necessary and has great potential in wearable healthcare devices.

2.1 Filtering

The preprocessing stage uses a filtering block to delete artifact signals from an ECG signal [20]. Usually, an ECG signal is initially bandpass filtered with different frequency ranges before analyzing it. Bandpass filtering is widely used to delete muscle noise, baseline wander, power line interference, and low- and high-frequency noise components and to limit ADC saturation and address antialiasing.

The frequency range of 0.1–100 Hz for the bandpass filtering is most often used [4,21–23]. Other frequency ranges used in bandpass filtering are 1–40 Hz [24–26], 0.5–40 Hz [12,27,28], 1–30 Hz [29,30], 0.4–40 Hz [31], 0.05–40 Hz [32], 0.5–50 Hz [33], 1–120 Hz [34] and 1–100 Hz [35]. Refs. [36] and [37] also used a bandpass filter to remove noise, but they did not specify the frequency ranges of the filter. The output of the bandpass filter proceeds through a moving average filter to smooth the signal [38–41].

Analog low-pass filtering has a noticeable effect on the QRS complex, epsilon, and J-waves but does not alter the repolarization signals [42]. A good low-pass filter can filter out the noise and still leave a large amount of data for further processing [43]. A low-pass filter is designed to remove the high-frequency component in the ECG waveform. Low-pass filters with the cut-off frequency of 11 Hz [36,44], 90 Hz [45], 30 Hz [46,47], 35 Hz [48], 50 Hz [49], 100 Hz [50] and 70 Hz [51] were used to delete high-frequency noise and power line interference. References [52–55] also used a low-pass filter, but they did not specify the cut-off frequency.

Unlike low-pass filters, analog high-pass filters do not attenuate much of the signal. However, analog high-pass filters suffer from phase shifts that affect the first 5–10 harmonics of the signal [42]. The main intention of a high-pass filter in ECG work is to remove the DC offset, which in turn is largely caused by the electrode/gel/body interface [56]. High-pass filters with cut-off frequencies of 0.5 Hz [45,49,51,57], 1 Hz [46,47], 2.2 Hz [55] and 5 Hz [44] were also used to remove baseline wander and for drift suppression. Reference [53] used a high-pass filter to determine the level of high-frequency noise that is available in any beat.

The motivation behind a notch filter is to attenuate several singular frequencies while preserving the others [43]. Notch filters combine both high- and low-pass filters to create a small region of frequencies to be removed. High “quality” notch filters can be created in software that target only 50 or 60 Hz, but the drawback of these filters is that they can create unusual ringing, especially for waveforms with high rates of change [56]. Notch filters centered approximately 50 Hz [45,51] and 60 Hz [32] were used to remove power line interference and suppress DC components. References [46,55,58,59] also used a notch filter for the same purpose.

ECG signals are also filtered with two median filters that have 200 and 600 ms widths to remove the baseline wander [58,59] and the P and T waves [48]. A series of three median filters was used to remove the ECG isoelectric line [60]. The median filters usually have the order of two [54] and fifty [61]. A local median filter is used in [62,63] to decrease the especial effects and arbitrary noise, and two steps of median filtering is used to delete the baseline wandering [4]. An averaging filter helps to assess the polarity of the P and T waves [36]. In that study, the average value of each six-adjacent points was used. Zero-phase filtering provides sharp peaks near QRS complex regions and smooths out fake spikes [39]. Savitsky-Golay filtering or digital smoothing polynomial filtering was also utilized for smoothing the ECG signals at the beginning of the preprocessing stage [64,65]. An adaptive filter was used to reduce power line interference [52]. The adaptive filter achieves performance close to a fixed Kalman filter with an optimally selected noise covariance method [45]. Rahman et al. proposed various adaptive filters based on the normalized signed regressor LMS algorithm, normalized signed LMS algorithm and normalized sign–sign algorithm [66]. The adaptive filter ensures that the signal waveforms are not distorted while the noise is being removed [67]. The performances of abovementioned filters are given comparatively in Table 1. Since only two papers include an SNR improvement of the raw ECG signals, different performance criteria are given in the following table, although the respective values do not indicate the actual contributions of the filters.

Table 1. Performances of the filters used in the preprocessing step.

Filter type (Ref.)Frequency/DurationPerformance (Criteria)Bandpass [4]0.1–100 Hz99.7% (total accuracy)Bandpass [21]0.1–100 Hz98.3% (average accuracy)Bandpass [23]0.1–100 Hz95.24% (average accuracy)Bandpass [24]1–40 Hz96.42% (identification accuracy)Bandpass [25]1–40 Hz100% (subject identification accuracy)Bandpass [26]1–40 Hz94.8% (correct detection rate)Bandpass [12]0.5–40 Hz2.57% (error rate)Bandpass [27]0.5–40 Hz98.3% (average accuracy)Bandpass [28]0.5–40 Hz100% (subject identification accuracy)Bandpass [29]1–30 Hz99% (accuracy)Bandpass [30]1–30 Hz99.94% (identification accuracy)Bandpass [31]0.4–40 Hz0% (error rate)Bandpass [32]0.05–40 Hz100% (accuracy)Bandpass [34]1–120 Hz95.8% (average accuracy)Bandpass [35]1–100 Hz100% (average accuracy)Lowpass [36]11 Hz98% (average prediction accuracy)Lowpass [44]11 Hz97.01% (overall accuracy)Lowpass [46]30 Hz96.2% (accuracy)Lowpass [47]30 Hz100% (accuracy)Lowpass [48]35 Hz86.6% (average accuracy)Lowpass [49]50 Hz99.6% (average accuracy)Lowpass [50]100 Hz80% (accuracy)Lowpass [51]70 Hz88.84% (global accuracy)Highpass [49]0.5 Hz99.6% (average accuracy)Highpass [51]0.5 Hz88.84% (global accuracy)Highpass [57]0.5 Hz95.3% (detection accuracy)Highpass [46]1 Hz96.2% (accuracy)Highpass [47]1 Hz100% (accuracy)Highpass [55]2.2 Hz92.5% (classification accuracy)Highpass [44]5 Hz97.01% (overall accuracy)Notch [51]50 Hz88.84% (global accuracy)Notch [32]60 Hz100% (accuracy)Median [58]200 and 600 ms93.59% (overall accuracy)Median [59]200 and 600 ms90% (sensitivity)Median [60]200 and 600 ms89.22% (accuracy)Median [54]200 and 600 ms94% (multiway accuracy)Median [61]200 and 600 ms100% (sensitivity)Median [62]200 and 600 ms97.1% (accuracy)Median [63]200 and 600 ms97.41% (average accuracy)Median [4]200 and 600 ms99.7% (total accuracy)Savitsky-Golay [64]N/A96.02% (overall classification accuracy)Savitsky-Golay [65]N/A96.31% (overall classification accuracy)Adaptive [52]N/A100% (positive predictive accuracy)Sign based normalized adaptive [66]N/A25.8473 dB (average SNR improvement)Adaptive morphological [67]N/A65.5% (SNR improvement)

The median filter removes the outliers and shot noise that are independent of the magnitude. The median filtering is less sensitive to the outliers than the mean filter. The median filters are used most often for noise suppression or smoothing, while high-pass filters are typically used for signal enhancement.

One could use a static notch filter, but you would have to reject a wider range of frequencies to accommodate the variability in the main frequency. The adaptive filter follows the main frequency, and thus, the stop band can be much narrower, which retains more of the useful ECG information. The mean amplitude values for the notch-filtered signals were less than those for the raw and adaptive-filtered signals. Adaptive filtering can be a powerful tool for the rejection of narrowband interference in a direct sequence spread spectrum receiver. This finding is exactly the difference between normal and adaptive filters. In a normal filter, the filter coefficient is static, while it dynamically changes in an adaptive filter.

What can be done to prevent a wandering baseline in ECG?

How does wandering baseline become corrected?

What can cause a wandering baseline artifact to occur on an ECG?

What is wandering baseline on ECG?