Blood Pressure Estimation Using On-body Continuous Wave Radar and Photoplethysmogram in Various Posture and Exercise Conditions

The pulse arrival time (PAT), pre-ejection period (PEP) and pulse transit time (PTT) are calculated using on-body continuous wave radar (CWR), Photoplethysmogram (PPG) and Electrocardiogram (ECG) sensors for wearable continuous systolic blood pressure (SBP) measurements. The CWR and PPG sensors are placed on the sternum and left earlobe respectively. This paper presents a signal processing method based on wavelet transform and adaptive filtering to remove noise from CWR signals. Experimental data are collected from 43 subjects in various static postures and 26 subjects doing 6 different exercise tasks. Two mathematical models are used to calculate SBPs from PTTs/PATs. For 38 subjects participating in posture tasks, the best cumulative error percentage (CEP) is 92.28% and for 21 subjects participating in exercise tasks, the best CEP is 82.61%. The results show the proposed method is promising in estimating SBP using PTT. Additionally, removing PEP from PAT leads to improving results by around 9%. The CWR sensors present a low-power, continuous and potentially wearable system with minimal body contact to monitor aortic valve mechanical activities directly. Results of this study, of wearable radar sensors, demonstrate the potential superiority of CWR-based PEP extraction for various medical monitoring applications, including BP measurement.

continuous wave radar. The CWR system, as shown in Fig. 1a, includes two main transmitter and receiver parts. The transmitter, which is a local oscillator, emits a high-frequency wave signal T cw (t) through the body, described by (1), using the transmitter antenna as an on-body sensor. The carrier frequency of the setup is around 900 MHz.
The received signal R cw (t), defined by (2), includes the distance information d and periodic movement m(t) of a target organ in its phase, i.e. aorta or heart. The information is derived using a mixer and a demodulator as two in-phase (I) and quadrature (Q) base-band components given by (3) and (4) 41 . The base-band signals are amplified and low pass filtered before being recorded.
In Eqs 1-4, φ is the phase noise in the setup, θ is a mixture of setup, body movement and environment phase noise, and θ 0 is an estimation of the combination of all phase noises. Figure 1a shows the radar antennae placed at the sternum. Figure 1b shows the S parameter measurements for the antennae used. The S11 and S22 parameters are the ratio of the reflected power to the fed power at the transmitter and receiver antennae input which should be less than −10 dB for proper antennae operation 24 . The S11 and S22 measurements show that they are below −15 dB at 900 MHz, which means the antennae works properly around the system's carrier frequency.
For any antenna used on or near the body, specified safety threshold values of the Specific Absorption Rate (1.6 W/kg for 1 g of tissue or 2.0 W/kg averaged over 10 g of tissue) must be obeyed 24 . The emitted power fed from the antenna to the body is around 2 dBm (less than 2 mW) 24 . photoplethysmogram and electrocardiogram. PPG signal is recorded to measure PAT parameters using a sensor which is placed on the left earlobe. The reason that the earlobe is selected to measure PPG is that the signal quality is good in this area of the body, i.e. clear signals can be measured and, that the recorded signal from the earlobe is steady with most body movements.
The ECG signal is measured as a reference signal using two electrodes which are placed on the chest. ECGs are used to derive heart beats and each heart beat's R-peak to measure time parameters, i.e. PAT and PEP.
All three different setups i.e. PPG, ECG and CWR are placed on the same device, and their outputs are recorded at the same time leading to synchronized data collection. Note that, to consider the hardware delay of each signal, a comparison is done with a reference device ECHO (i.e. ultra-sound M-mode wave) and the results show no considerable time shift in recorded waveforms.
The subjects wear a calibrated cuffed sphygmomanometer during tests to record their SBPs simultaneously. To synchronize the recorded signals with a BP measurement cuffed device, all data are labelled based on the same reference time information.
Signal processing. The CWR, PPG and ECG signals are recorded simultaneously for all subjects. Figure 2a shows the block diagram of the proposed processing method to detect time parameters, i.e. PAT, PEP, and PTT. For the first step, a Chebyshev type II band-pass filter (BPF) is used to remove the 50 Hz components as well as noise and artefacts. To avoid the time shift caused by filtering, a zero-phase filtering algorithm is applied which helps to preserve the features of unfiltered signals exactly at the same time as the filtered wave-forms.
The pass-band of the BPF is chosen individually and automatically for each subject using the heart rate frequency (HRF) extracted from the individual subjects' PPG signals, and is fed into the analyses of the ECG, PPG and CWR. The varying band-pass filtering cut-off frequency method is required because this type of signal is prone to interference from respiration and tissue vasomotion, and individual subjects have different best cut-offs. To calculate the HRF, the floor of the first harmonic of Fast Fourier transform (FFT) is used. Then the BPF cut-off frequencies are defined for each signal as follows: After filtering, a combination of I and Q, i.e. arctangent of Q/I, is also calculated as described in (6).
1 0 0 Wavelet. Sometimes using filtering alone may lead to a loss of some important timing information from a signal. In our study, utilizing BPFs may displace the foot of CWR signals, which leads to incorrect PEP measurements. Therefore, a wavelet transform is used to preserve considerable information and to remove the remained noise from the signals. The wavelet transform decomposes the signal into a number of scales having different frequency components, and analyses each scale with a certain resolution to obtain accurate features of the signal 33 . The sum of the overall time of the signal multiplied by a scaled and shifted version of the wavelet function, addressed as H a b ( , ), is given as (7) and (8): where, x(t) is the original signal, * denotes the complex conjugation, Ψ a,b (t) in the window function of the mother wavelet and Ψ * ((t − b)/a) is its shifted and scaled version 30 . The selection of a particular wavelet function that closely matches the morphology of the signal under consideration is the most important factor for signal decomposition 30 . Since the CWR waveform is different from the known wavelet functions, a special mother wavelet function which matches our signals is designed and used as shown in Fig. 2b. The blood pulse volume pumping from heart to the arterial tree leads to the aorta movements. The blood volume applies pressure to the walls of the aorta and depends on its elasticity and resistance, movements are generated. Hence, the movement pattern of the aorta is highly correlated (similar) to the aortic blood pressure (ABP) 42 . The aorta pressure waveform has two main peaks 42 which can be used as the basic pattern to design the wavelet function. An algorithm is used to adapt a new wavelet to the discussed pattern by extracting the similar waveform on the interval of [0,1], which is also quite close to each beat of the CWR signal. The algorithm uses a least squares polynomial approximation of degree 6 to create the wavelet function 43 . Using the designed wavelet function, the CWR signals are decomposed down to level 12. Comparing the decomposed elements shows that the wavelet coefficients at scales 8 to 12 are closely representative of the heart rate for the general population; therefore, only the mentioned scales are used for reconstruction. It also leads to elimination of high frequency and very low-frequency noise sources, including respiration.
Adaptive filter. By selecting the wavelet coefficients at special scales, unwanted noise is eliminated and smooths the signal, which is not negligible, in most cases. This selection may also lead to loss of some important time interval information; in other words, PEP measurements which are related to the foot of each beat of CWR signals. Therefore, adaptive filtering is used; the characteristics of which change in a way to obtain the best possible signal quality in spite of changing signal conditions 44 . An adaptive filter designs itself based on the characteristics of the input signal to the filter and a signal that represents the desired behavior of the filter on its input. Adaptive filters track the dynamic nature of a system and eliminate unwanted time-varying signals. Designing the filter does not require any other frequency response information or specification. An adaptive algorithm is used to reduce the error between the output signal and the desired signal 45 .
For CWR signal processing, the result of the wavelet transform is selected as the desired signal and the filtered signal is the input of the adaptive filter. In this case, those frequency components which are removed during wavelet transform remain; yet, since the wavelet-filtered signal is the desired signal, incorrect foot of signals can be removed. For the current study, an adaptive finite impulse response (FIR) filter uses a least mean squares (LMS) algorithm. The FIR filter weights are calculated using the LMS algorithm to minimize the mean square error (MSE) between the output signal and the desired signal 45 . The output of the adaptive filter is then used as a signal to derive PEP parameters. Figure 2c shows a sample of CWR signal after different levels of signal processing. Figure 2d shows a sample of signals which are measured with the three sensors during the recording of heart activities. The PAT is selected as the time difference between the ECG R-peak (t p ) and the maximum of the PPG signal (t m ). It also can be assumed as the time difference between the ECG R-peak and the rising slope of the PPG signal (t r ).

Blood pressure estimation.
As previously mentioned, a comparison between measured PEP from CWR and ICG sensors was undertaken to validate the precision of the PEPs [38][39][40] . In this study, the PEP measured from CWR signals after different levels of processing is compared with PEP measured simultaneously using ECHO i.e. ultra-sound M-mode wave to validate the accuracy and precision of the parameters, as shown in Fig. 2e. This figure shows a high correlation between two measured PEPs and that the foot of the CWR signal is close to the foot of M-mode signal. In literature 37 , it was shown that the QR interval can be assumed to be fixed; therefore, PEP is measured as the time interval between R-peak of ECG and the foot of CWR (t f ) signal.
Finally, the PTT-which is the difference between PAT and PEP-is measured to remove the effect of PEP from the PAT. After processing three ECG, PPG and CWR signals, the time parameters used for SBP estimation are calculated as following, and as highlighted in Fig. 2d: To measure the time parameters, the location of the ECG R-peak is detected using Algorithm 1; then the boundaries of each heartbeat are measured. www.nature.com/scientificreports www.nature.com/scientificreports/ To calculate PEP values, the Algorithm 2 is determined. As can be seen, a comparison is done between measured T f elements and predefined thresholds t f min, and t f max which are the reported ranges for PEP in 37 .
To calculate PAT parameters, the rising slope and maximum of corresponding PPGs are measured as the maximum of the first derivative and the first local maximum of the signal, respectively.
The SBP values are estimated from PTTs using different regressions as follows: where for Eq. 15, PTT can be PTT 1 or PTT 2 . To determine the coefficients of the equations, cuff-based SBPs were measured and used as the calibration points of the models. Then the measured SBPs and the whole PATs and PTTs recorded from subjects were used to calculate the coefficients in a way that the estimated SBPs have the minimum difference from the measured SBPs. The coefficients a, b and c are calculated using minimum least square (MLS) fitting method. Different mathematical models have been suggested in literature to calculate BP from PTT or PAT 46,47 . Equation 15 is one of the most common with very high accuracy and simplicity as reported by researchers 46,47 . Most suggested models are based on using a one-time parameter. This led us to employ two-time variables in a model as in Eq. 16 to improve accuracy of the results. Although the SBP estimation is done for each subject separately, in other words the equation coefficients are calculated separately, an algorithm is designed and used for regressions based on the two mentioned equations which do not need any calibration for coefficients or any tuning for an individual subject. We are using each three-minute cuffed sphygmomanometer measurements as calibration points to increase the calculation accuracy of a, b and c, so the errors will be reduced and the BP can be measured by Eqs 15, and 16 continuously. experimental protocol. Forty-three healthy volunteers participated in the experiments. Fifty two percent were male and 48% were female, between 40 and 65 years, (168 + 10) between 158 cm and 178 cm in height and weighed between 44 kg and 76 kg with no participants reported of any previous cardiovascular problems. A written informed consent was required from each volunteer who participated in this study before starting the experiments. A video displaying testing procedures was recorded during one participant's test. Two written informed consents were obtained from the subject for study participation and for sharing any identifying image/video in an online open-access publication, respectively. Data collection was undertaken at Cabrini Hospital, Melbourne, Australia, (Oct. to Nov. 2017) and an in-room emergency physician ensured participants' safety during testing. All experiments were performed in accordance with the relevant guidelines and regulations and the required Algorithm 1. R-peak detection. ECG, PPG and CWR signals were measured simultaneously and the data from 43 participants were included in the analysis. The experiments were undertaken in two sessions. All volunteers participated in Session 1, while 26 subjects from among them were selected randomly to take part in Session 2. All subjects were asked to breathe normally during tests. The CWR data from five subjects were corrupted during signal recording because of the hardware failure, hardware data loss, loose cable connection or interrupted data collection; therefore, they were omitted for this work. Figure 1c shows the placement of CWR and PPG sensors for a participant seated on the exercise bike. The figure also shows the cuff placement and the software interface used for data recording.
Session 1 consisted of measuring all signals for three different postures; six minutes of sitting, six minutes of standing and six minutes in the supine position. The procedures for Session 2 were as follows: • Holding handgrip for two minutes followed by a one-minute rest.
• Cycling with a fixed speed in three different bike resistant settings of light, moderate and heavy. Each cycle lasted for two minutes with a one-minute rest. • Two recovery stages after the cycling tasks, during which subjects sat on the bike without any activity.
All ECG, PPG and CWR signals were recorded continuously during tests. Also, the subjects were asked to wear a calibrated cuffed sphygmomanometer during tests and their BP values were recorded every three minutes. The timing of cuff measurements was set so that the BP measurement happened once for each test and started after the second minute. The error of the hydrostatic effect was removed by keeping the cuff around the arm at the same horizontal height of the heart during all measurements 48 .

Results
The SBP values of a healthy person cannot vary more than +10 mmHg in two consecutive readings 49 , while this variation may reach +15 mmHg for anatomical abnormalities or measurement errors 49 . Since the probability of having anatomical abnormalities is very low, and the accuracy of the -cuffed sphygmomanometer used in this study is +3 mmHg, we can consider measurements with a higher than 20 mmHg (>15 + 3 mmHg) deviation were caused by measurement errors. Therefore, all the cuff measurements with an error of more than 20 mmHg, and their related PTT or PAT values, were removed. The collected data were processed and analyzed in two independent categories: postures (sitting, standing and supine) and exercises (handgrip, light, moderate and heavy exercise, recovery 1 and recovery 2) categories independently. The values of t f min and t f max were chosen as 27 ms and 78 ms respectively 37 .
pre-ejection period. Figure 3a,b show the box plots for two sessions of tests for all participants. Figure 3a shows the lowest PEP values, which were obtained in the supine posture. It can also be observed that the PEPs for sitting and standing positions are very close. A comparison of the results of the three different stages of cycling exercises, as shown in Fig. 3b, shows a negative trend between the strength of activities and measured PEPs.   Fig. 3a represent a slight difference for the different postures. Finding a clear trend between PAT and postures seems to be complicated. As can be seen, after calculating PTT (subtracting PEP from PAT) there is a similar trend for both PTT and SBP values due to changing the postures. The changes in PTTs for different postures show the noticeable effect of subtracting PEP from PATs to achieve a trustworthy pulse time delay parameter. Note that the relation between PTTs and SBPs should be considered in another approach, which is discussed in the next section. Figure 3a,b show both PAT and PTT have negative trends with respect to exercise stress during exercise tasks, while BPs have the same trend. While PATs and PTTs decrease when the bike exercise was intense, their related BPs increase. The changes for both PAT and PTT were quite clear, which means that the effect of exercise is sufficiently significant despite the influence of PEPs.  www.nature.com/scientificreports www.nature.com/scientificreports/ Blood pressure calculation. Figure 4a,b show the measured SBPs based on their related PTT and PAT values during three different posture tasks. As seen, there is not a clear relation between extracted PAT values with SBPs, while PTTs and SBPs remain negatively correlated despite having same trend due to posture changing.
Similarly, Figure 4c-f show the cuff-based measured SBPs against PAT 1 and PTT 1 for exercise tasks. Both PATs and PTTs illustrate an inverse relation to SBPs during exercise, which was also observed for their trends in the previous section. In Fig. 4e,f the estimated inverse relation between SBPs and PATs/PTTs using Eq. 15 are presented as curves. The curves are plotted (based on Eq. 15) for each subject independently, using the MATLAB curve-fitting tool. As can be seen, the fitted curves for different subjects based on PATs are more scattered than for those based on PTTs. The calculated curves for SBP-PTTs are more linear and coherence than SBP-PATs. This leads to a noticeable improvement in calculating SBPs based on PTTs due to eliminating the PEP effect from PATs. Figure 5a,b show the Bland-Altman (BA) plots with limits of agreement (LOA) for all estimated SBPs using Eq. 15, based on cuff-based measured SBPs, and time parameters including PAT 1 and PAT 2 , PTT 1 , and PTT 2 , for different posture and exercise tasks. As can be seen, there is a considerable improvement in calculating SBPs using PTTs compared to that of using the PATs.
In Fig. 5c the results of SBP estimation with Eq. 16, using the combination of PAT 1 and PAT 2 /PTT 1 and PTT 2 , are shown. Equation 16 cannot be applied for postures because the number of points are not enough (less than four for each participant) and this leads to over-fitting. The combination of PTTs compared with PATs provides significant improvement in terms of SBPs calculation accuracy. Table 2 presents the accuracy percentage for estimated SBPs using PAT 1 , PAT 2 , PTT 1 , PTT 2 , and cuff SBPs, which are based on Eqs 15 and 16 for different postures and exercise tasks. The accuracy is calculated using the cumulative error percentage (CEP) of less than 7 mmHg, less than 10 mmHg, less than 15 mmHg, between 10 mmHg and 15 mmHg, between 15 mmHg and 20 mmHg, and more than 20 mmHg. The CEP less than 7 mmHg is considered based on the newly established IEEE standard on continuous blood pressure measurement system 2 . By comparing the results of estimated SBP based on PATs and PTTs, it can be seen that eliminating PEP leads to an almost 9% improvement. Another observation is that the accuracy percentages for the posture session are higher than for the exercise tasks. This is because recorded signals during posture measurements are less noisy, due to the participants' steadiness, compared with exercise tasks where subjects' body movements increase due to cycling. Table 2 illustrate that the results of using Eq. 15 based on the first order of PTT 1 are slightly better than using PTT 2 for the same equation or using the second order Eq. 16 based on PTT 1 and PTT 2 .
Considering the complexity of the calculation for Eq. 16, using two variables from the same signals i.e. PTT 1 and PTT 2 does not necessarily leading to the optimum solution.

Discussion and conclusion
The proposed method in this study investigates continuous BP estimation based on PTT measurements for the first time using on-body radar sensors on the chest and PPG signals from the ear. By positioning the CWR sensor on the sternum and evaluating its signal, PEP is measured as the difference between the foot of CWR signal and ECG R-peak. Comparisons between the effects of changing posture and different exercise tasks on PTT, PAT and their related SBP were conducted. The results show that the removal of PEP values from PATs of PPG signals to estimate the SBP based on PTT, leads to more accuracy. Two different regression methods (first order and second order equations using one variable and two variables respectively), based on Eqs 15 and 16, are implied to calculate SBPs from PTTs and PATs.
The proposed CWR sensors have the ability to observe mechanical movements of the aortic valve (the aortic arch opening) similar to ECHO, which is considered as a reference method to observe aortic activities with the highest accuracy in PEP extraction 36 . Despite the focus on ICG technologies, CWR presents a low-power, continuous and potentially wearable system with minimal body contact to monitor aortic valve mechanical activities directly. In addition, compared to other pulse wave monitoring methods, such as PPG, CWR can provide central elastic aorta monitoring, removing the peripheral vasomotion-induced inaccuracies, which requires higher quality in the design of CWR sensors for recordings and processing in future work. This shows the potential of the CWR-based system to provide a simple wearable device, which can monitor both PEPs and PTTs (in results, beat-to-beat SBPs).
Data are collected during different posture and exercise tasks considering their effects on time parameters and BPs separately. This leads to better understanding of changing posture without having activity impacts and vice versa. Removing the effect of PEPs from PATs leads to a clear negative relation between PTTs and SBPs for various postures, and a more coherent negative relation (as proved in other research studies 7,27 ) for exercise tasks. It should be considered that the variations of time intervals for different postures significantly relate to each subject's specifications. In addition, the BP variations are very low for different postures 12,26 . The experimental conditions are designed to observe higher than normal BP values, while there is a need to study the decreasing trends of BP as well. In addition, investigating continuous data (beat-to-beat) can provide a better understanding of the system and its time parameters.
The feet of CWR signals using the maximum of the second derivative, compared with 37 presents the PEP values 37 . The output of CWR contains three main signals as I, Q and arctangent Q/I. Instead of using these CWR signals separately, the combination of I, Q and arctangent-based extracted PEPs is presented to obtain more accurate values. To derive PTT parameters, the PEPs are subtracted from PATs. The PATs are measured as the rising slope and maximum of corresponding PPGs by calculating the maximum of the first derivative and the first local maximum of the PPG signal. The proposed signal processing methods based on wavelet and adaptive filtering techniques are employed to clear radar signals by removing unwanted noise and artefacts without causing any time shift. A BPF is designed to omit unwanted frequencies considering each subject's heart rate which allows more accurate respiration/movement artefact's filtering.
In some cases, the radar signals are not as reproducible in shape, especially during exercise tasks. There may be some other factors such as the heart size, position within the thorax and orientation, and respiration which may affect the signal and the place of its foot, subject by subject 24 . One other challenge to achieve a clear radar signal is the shape of antennae as an on-body sensor, and also the way that they are attached to the body. Redesigning the current antennae to be small and flexible, targeting more reliable signal detection along with applying more advanced signal processing methods to extract PEP values with higher accuracy, will be investigated in future works.
The SBP calculations based on PATs and PTTs are investigated separately using two different mathematical models. The results show a noticeable effect of PEP-correction in the SBP calculation accuracy. For both experimental conditions, an overall improvement of 10% is achieved using PTTs compared to PATs. The highest improvement is obtained using Eq. 15 based on PTT 1 s which also has the advantage of simplicity. Based on the results represented in this study, the potential superiority of CWR-based PEP extraction for different purposes such as SBP calculations is demonstrated.
Sometimes the PEP can change in an opposite direction to the PAT during BP changes due to the variations in preload and afterload 34 . Note that the PEP (in parallel PTT) can be influenced, not only by position or activity of the subject, but also by other parameters such as heart rate, vessel elasticity, age, body fat and gender. Adding the effect of these parameters to the mathematical regression methods could provide more accurate SBPs estimation, which will be studied in future works.