Digital health provides the opportunity to combat pandemics, to deliver health care in remote regions, and to reduce the carbon footprint of health care delivery. Telehealth and remote monitoring, particularly in patients’ homes, has become an important option due to problems associated with keeping patients in hospitals and care centers for extended periods: the increase in the probability of acquiring nosocomial infections [1]; the deficiency in medical infrastructure to meet the demand of patients [2]; the increase in therapeutic dependence by older adult patients [3]; and the increase in hospitalization costs. These issues have led to a search for alternatives in medical care such as home health care.

This has led to an increase in the development of remote monitoring technologies of patients’ vital signs to improve the medical diagnosis [4,5]. Wearable devices for monitoring vital signs have become a powerful tool to improve health services and to implement home health care models [6,7]. These will allow to acquire vital signs of patients in daily environments while they carry out their activities in a normal way, allowing to obtain complementary information to improve the medical diagnosis, to constantly monitor the patient’s condition, and to improve their treatments [7,8].

A major challenge for the diffusion of digital health technologies lies in the signal quality of sensors for remote diagnosis and monitoring of patients, including electrocardiogram (ECG) signals of moving patients, which require the denoising of motion artifacts (MA). ECG is an important diagnostic technique for application in wearable devices, owing to the amount of information contained in the acquired waveform, distributed in different peaks and undulations called segments (P, Q, R, S, T, and U). The P segment represents atrial depolarization, the QRS complex represents ventricular depolarization and atrial repolarization, the isoelectric ST segment is the time when both ventricles are completely depolarized, the T segment represents ventricular repolarization, and the U segment represents papillary muscle repolarization [9,10]. Each segment is characterized by its unique shape, amplitude, duration, and time of occurrence, allowing to identify the way in which the electrical impulse is conducted through the heart muscle [10].

Newer devices have been developed to identify cardiovascular diseases in early stages (asymptomatic). Some of them are external loop recorders, implantable loop recorders, and Zio patches as well as wearable ECGs. Similar to traditional Holter, they share a sensitivity to artifacts [11], which leads to repeated ECG monitoring with cost increases in 11.1% of cases [9]. In the case of wearable ECGs, the information provided by these devices is not considered for clinical use owing to the contamination by different noise sources such as power line interference, baseline wander, and MA that have nonlinear, nonstationary, and unpredictable character, as well as ECG bandwidth overlaps [12-14]. In addition, the effects of daily life movements on the signal are difficult to predict, which makes the devices’ validation for medical use in home health care and outdoor conditions even more difficult. This motivates the development of techniques that allow the reduction of interference in the ECG signal.

Previously, research has been conducted to develop techniques that solve the problem of combined interference of MA, baseline wander, and power line interference in ECG signals [15]. Performance assessments of denoising techniques such as wavelet shrinkage (WS), empirical mode decomposition (EMD), wavelet independent component analysis (WICA), and EMD independent component analysis (ICA) have been performed. These methods present problems, although some of these work in the denoising of synthetic signals when working with signals from patients in movement. One of them is that their signal databases only consider a single source of information (ECG signal) to perform signal denoising, which makes it difficult to acquire the dynamics introduced by movement and significantly affects the performance of artifact reduction methods. However, it was found that depending on the segment of the ECG signal to be preserved, it is possible to use a specific denoising technique for that segment [16].

In recent years, significant advances in the development of techniques for feature extraction from cardiovascular signals in wearable monitoring have been made. The presence of MA has been identified as a significant source of noise in signal acquisition, masking information about the physiological process and leading to misdiagnosis. The MA reduction problem is still addressed in different ways as proposed by Yang and Tavassolian [17], where it is possible to obtain cardiovascular parameters from the seismocardiography signal analysis; they used the ICA on the inertial signals acquired from inertial measurement units (IMUs) and used the ECG and photoplethysmography (PPG) signals as reference signals. An and Stylios [18] evaluated conventional filtering methods using finite impulse response, infinite impulse response, moving average, moving median filters, and advanced decomposition methods such as wavelet, EMD, and adaptive filters to compare their performance. They found that all these methods have their limitations, but the best method was considered to be the adaptive filter. However, it depends on a good selection of the reference signal and still introduces distortion to the signal [18,19].

Other approaches use adaptive noise signal detection, EMD, or wavelet decomposition of the signal of interest and dynamic time warping (DTW) component selection to reduce baseline wandering and high-frequency noise, and they achieved signal improvements of up to 25% [11,20,21]. On the other hand, the electrode configuration and its interaction with the skin had been evaluated to determine the impedance variation and the noise introduction in the signal acquisition [22-24]. Many authors agree that the way to approach the problem is through the use of signal decomposition methods such as wavelets, EMD, and ICA, among others [25-27]. In addition, including multiple sources of information such as pressure signals, PPG and movement are essential to estimate the physiological parameters of interest [28].

Although some of these methods use the acquisition of multiple signals as other sources of information for the determination of cardiovascular parameters, few of them are focused on evaluating the recovery of the waveform of the ECG signal. Similarly, dealing with the acquisition of multiple physical magnitudes such as pressure or PPG through a single channel does not allow confirmation of the correct denoising of the signal of interest. The method proposed in this paper allows the redundant and simultaneous acquisition of ECG signals and movement signals to obtain more complete information of the physiological process masked by the movement artifacts.

It has been observed that the correct location of the electrodes on the volume conductor of the patient has a great influence on the result of the ECG, to the point that a bad location of these can lead to diagnostic mistakes [29,30]. However, it has been identified that certain modifications in the signal acquisition hardware, such as adding additional leads in the back and taking information of the movement of the person, allow obtaining additional information of the signal of interest [16,31]. In this study, additional electrodes were added to the acquisition hardware of the ECG signal in the chest and back of the volunteers. This represents a novel method for the acquisition of the signals in a redundant and simultaneous way to improve the denoising, decreasing the distortion introduced in the MA reduction process.

The distortion concept in the context of biomedical signal processing refers to the change in the natural shape of the signal due to external processes or disturbances, which include changes in the amplitude, duration, or time of occurrence of the segments that compose the signal and lead to loss of information. The concept of redundant measurement refers to the acquisition of information from the same interest source from ≥2 different measurement points, with the purpose of preventing the loss of information or increasing the sources of information of interest, as is true in this study. Redundant measurements are performed simultaneously to ensure that the information acquired from the different measurement points is synchronized, which is defined as multichannel synchrony.

Further improvement of noise and MA reduction is critical to achieve an optimal diagnostic yield with wearable health care monitoring systems. The diagnostic yield will be an important determinant of the devices’ cost-effectiveness [32]. Only with a reliable diagnosis of specific cardiac arrythmias such as atrial fibrillation will the devices’ cost-effectiveness allow widespread application, even in resource-constrained environments [33].

This paper presents a novel method for the reduction of noise and MA in ECG signals from walking individuals in ambulatory vital signs monitoring applications. We have introduced a new method called Redundant denoising Independent Component Analysis (Rd-ICA). It is based on (1) redundant and simultaneous measurement of ECG signals in the chest and back; (2) acquisition of triaxial movement signals from IMUs; (3) a reference signal synthesized from an autoregressive model, which considers the features of a resting ECG signal obtained through ensemble average (EA) ECG [9,16]; and (4) separation of interest sources and noise sources through multichannel ICA. After the separation of the signals, the identification of the ECG signal is made through the comparison of the components with the synthesized reference ECG signal.

The performance of this method is tested with a database composed of data sets of movement signals and ECG signals acquired in the chest and the back from healthy volunteers in conditions of rest and movement. In addition, the performance of the Rd-ICA method is compared with the performance presented by state-of-the-art denoising methods such as the WS method and the WICA method. The calculation of performance indexes is performed with indexes such as the signal-to-noise ratio (SNR), the DTW, and a proposed index defined as weighted distortion assessment (WDA). It measures the characteristics of the shape of wave found with the EA ECG method.

This section shows the protocol for recording ECG signals, presents some previous state-of-the-art methods for ECG signal denoising, and finally shows the proposed Rd-ICA method. In this work, the comparison of the methods’ performance was also carried out. A new index based on the signal distortion characterization through the EA ECG has been proposed.

A database of ECG signals acquired from a population of 20 healthy volunteers aged, on average, 26.3 (SD 5.7) years with an average BMI of 24.4 (SD 4.8) kg/m² was registered. Database registration was performed for bipolar leads DI, DII, and DIII in the chest and the back and the triaxial movement signal of the volunteer. The experiment was divided into 3 stages: (1) rest before movement, (2) controlled movement in laboratory conditions, and (3) rest after movement. Each volunteer was asked in the first stage to remain at rest for 5 minutes, which led to the acquisition of reference ECG and motion signals. In the second stage, each volunteer was asked to perform a walk at a normal travel speed of 4.2 (SD 0.8) km/h for 5 minutes, which produces contamination that masks and distorts the ECG signal significantly. In the third stage, each volunteer was asked again to be at rest for 5 minutes. This protocol was performed to obtain a database of ECG signals contaminated with MA composed of redundant and simultaneous ECG signals acquired in the chest and the back, also with the movement of the volunteer registered through IMUs [24]. ECG signals were acquired at a sampling frequency of 250 Hz and 24-bit resolution [34]. Informed consent was obtained from all participants involved in the study.

To acquire signals, a custom wearable device was used that performs the acquisition of the signals of ECG; PPG; and noninvasive blood pressure that is redundant, simultaneous, and synchronized [31,35]. The ECG is acquired on the volunteer’s chest and back, and an IMU is included on each ECG lead to record the inertial activity and movement. It similarly occurs for the PPG signal that is acquired both in the left wrist and the right wrist and for the noninvasive blood pressure signal that is recorded in both arms through the oscillometric method [8]. The IMUs allow acquiring the inertial activity and movement to analyze the nonlinear dynamics of artifact contamination on the signal.

All the signals are recorded simultaneously and synchronized in time to guarantee signal redundancy and the possibility of applying the Rd-ICA method. For the ECG signal, Ag/AgCl electrodes were used, which have proven to be the ones with the least interference in recording of the signals due to their correct coupling with the skin.

In clinical ECG, it is common to place the electrodes on the arms such as the left and right wrists and the left foot under the Einthoven triangle model. In long-term examinations such as Holters, it is common to use the positioning of the electrodes according to the Einthoven triangle under the Mason-Likar [36] method in which the electrodes are located on the person’s chest. This model is the most frequently used in ECG wearable monitoring [36,37]. Signals were acquired through the connection of an electrophysiological signal recording equipment to acquire the biopotentials in the torso and back of 20 volunteers as previously validated [24,31]. A group of sensors were placed in the location proposed by the Einthoven triangle in the Mason-Likar [36] method in the chest. These locations were interpolated on the back of the volunteer, considering anthropometric locations [38]. In addition, the acquisition of triaxial movement signals was performed through an IMU located in the electrophysiological sensors. It should be noted that the acquisition of electrophysiological signals in the chest and back was performed redundantly and simultaneously, synchronized with the movement signals. Figure 1 shows the distribution of electrodes in the chest and back of the volunteer.

The most frequently used techniques to reduce MA have previously been described in detail, including WS, ICA, and WICA [17].

The method proposed in this paper is based on the simultaneous and redundant acquisition of the ECG signal, the movement of the person, the separation of multichannel components, and the selection of the improved signal through the comparison with a modeled signal from previous information. The processing scheme of the Rd-ICA method is presented in Figure 2.

The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Human Studies Institutional Ethics Committee of Universidad de Antioquia (protocol code 16-59-711 of May 19, 2016). Informed consent was obtained from all participants involved in the study.

The EA ECG method was applied to ECG signals of volunteers acquired in resting conditions (Figures 3A and 3B) and to ECG signals acquired with MA (Figures 3C and 3D).

Figure 5A and Figure 5B show the EA ECG of the signals acquired in resting conditions in the chest and back of a volunteer, respectively. The solid line represents the average signal and the dashed lines represent the SD signals. Figures 5C and 5D present the acquired signals in movement conditions; in the same way it presents the average signal by means of the continuous line and the SDs by means of the dashed lines. The dispersion of the different waves that compose the EA is observed during the movement.

Some state-of-the-art denoising methods were applied to ECG signals contaminated with MA. These signals were selected only from the chest of the volunteers while they performed movement. The methods applied were the WS and WICA method.

The Rd-ICA method was applied to the redundant and simultaneous ECG signals contaminated with MA and acquired in the volunteer’s chest and back. The comparison was made with a reference ECG signal synthesized from an autoregressive model. The result of applying denoising methods on the ECG signal of a volunteer is presented (Figure 6). In addition, the EA ECG of the denoising result of the ECG signal through the WS, WICA, and Rd-ICA methods is shown (Figure 7).

A graphic comparison between the results of the denoising presented by each of the evaluated methods is presented (Figure 6). In addition, the result of applying the EA ECG over the improved signals with each of the methods, including average and SD signals is presented (Figure 7).

This paper presents a novel method for MA reduction in ECG signals through the acquisition of redundant and simultaneous signals, acquisition of the person’s movement, modeling of the reference signal from prior knowledge of the ECG signals at rest, and processing of this set of signals through multichannel processing techniques. This method was called Rd-ICA. This was applied to a data set composed of ECG signals acquired from healthy volunteers at rest and in movement conditions.

The performance of the proposed Rd-ICA method was compared with state-of-the-art methods for MA reduction such as WS and WICA. To compare the performance of the different methods, each was applied to the data set, and the performance was evaluated from the measurement of indexes such as SNR, DTW, cross-correlation, and the proposed WDA that consider the morphological features of the signal [16].

The use of a single denoising method does not guarantee the reduction of noise in all features. Sometimes it is necessary to use specific denoising methods to improve the signal and maintain some feature with fidelity. Despite this, the Rd-ICA method presents a good alternative for the reduction of MA in contaminated ECG signals, as shown by the results obtained in this paper.

The ECG signals acquired redundantly in the chest and back of the volunteers showed similar information from the same source; this information belongs to the ECG signal that is of interest in this study [59]. In addition, it contains information from other sources among which are MA, which are measured as information from independent sources. This allowed the separation of the signal of interest and the signal of artifacts through the proposed multichannel signal processing technique.

Most of the methods reported are based on the separation by components, the extraction of the noise components and their elimination, then the reconstruction of the signal of interest. The works at the frontier of knowledge make use of methods such as EMD, wavelet, and adaptive filters with some modifications and improvements, and these are the ones that have presented the best performance, but according to those reports, they also introduce a large amount of distortion in the signal of interest.

The method proposed in this paper is advantageous and novel from the point of view that it acquires the signal redundantly, thus providing a way to validate the processing applied to the signal of interest. Other finding was that both components of the ECG signal acquired on chest and back were significantly similar, while the artifacts contamination showed differences between the 2 signals acquired on the chest and back of volunteers [24]. In addition, the motion component helps to determine the dynamics of the signal, evaluate the nonlinearity of contamination by artifacts and perform a better estimation of the components of interest through the proposed Rd-ICA method.

This technique presents its application with wearable vital signs monitoring devices, which have their main field of application in outpatient vital signs monitoring. This technique requires a modification in the signal acquisition hardware as it requires redundant and simultaneous ECG signal acquisition from the chest and back of patients. In addition, it requires the measurement of movement through IMUs. These modifications are possible to implement in wearable devices [31]; therefore, the technical feasibility can be affirmed for implementation in outpatient monitoring.

The proposed technique has important potential in the processing of physiological signals from different sources with the use of redundant acquisition of the interest signals. It shows its application in the identification of signals from cardiac arrhythmias in conjunction with the EA ECG method and the proposed WDA index. Some of the ECG monitoring devices that potentially may include this technique to reduce artifacts and to improve their diagnostic potential are external loop recorders, implantable loop recorders, traditional Holter, and wearable ECGs. This kind of improvement will reduce the associated costs with repeated tests.

The method proposed in this paper presents a considerable advance in the reduction of MA in ECG signals as the results showed an improvement in the denoised signal. Its advantage is not only in the evaluation of signal indexes such as SNR but also in the preservation of the signal waveform, low distortion introduction, and the potential use in medical diagnosis. It was observed that the Rd-ICA method presents a significant improvement in the conservation of the characteristics of the signal compared with the WS and WICA when it was evaluated using the EA ECG method.

In the same way, the possibility to have the redundant ECG signal and the motion signals from the inertial sensors provides the proposed method with the capability to separate the component of interest from the components of artifacts. At the same time, it allows the method to determine these components autonomously by comparing this signal with a reference signal built from an autoregressive model that considers the cardiovascular characteristics of the volunteer. This allows the proposed method to be implemented in wearable monitoring systems and in autonomous monitoring systems.

On the other hand, there is evidence of the need to carry out more extensive research to evaluate the reliability and clinical validity of this method in patients with relevant diseases. Although this method presents the possibility to differentiate events coming only from the cardiovascular system and separate them from external events such as the volunteer’s movement. This is due to the possibility of obtaining the ECG signals redundantly and simultaneously. In addition, the need to add new sensors for the redundant measurement of the signals and the obtaining of the movement signal through IMUs implies significant modifications in the hardware of the existing monitoring systems. Despite this, the possibility of making these modifications is evident and opens the possibility to generate new designs, enabling the growth of the market for wearable medical devices.

Currently, the research and development of biosensors have a great boom thanks to the advantages they show, such as easy application and portability. Their easy implementation in wearable devices for the continuous measurement of different vital sign signals or physiological variables in everyday environments makes these devices a great option in vital signs monitoring [60]. Despite its great applicability, the information acquired by these biosensors is distorted by different sources of noise and artifacts, such as MA. The method proposed in this work is not limited to the denoising of the ECG signal, but it can be used for other physiological signals that can be redundantly acquired and are susceptible to be affected by MA. Some of these sensors are PPG biosensors, enzymatic biosensors for glucose measurement, intraocular pressure, and hydration percentage [60,61].

The technique’s ability to improve the quality of the signal is critical for diagnosing specific cardiac arrhythmias in real-world use. The diagnostic yield has been shown to be a major determinant in a technique’s economic assessment; for example, in diagnosis after palpitations [62] or syncopes [63,64], in screening of athletes [65,66], or in identifying asymptomatic atrial fibrillation [33,67]. To explore these applications, the acquisition of a database that considers more extreme movements and patients with common cardiac pathologies is required, which will provide information about the effect of artifact promotion techniques in the correct identification of arrhythmias or the malfunction of heart. Such a database would allow future work on the proposed method and a benchmark with existing methods to evaluate their performance in MA reduction as well as its benefits in the identification of waveforms modified by specific cardiac arrhythmias.