A chest-based continuous cuffless blood pressure method: Estimation and evaluation using multiple body sensors

A Chest-Based Continuous Cuffless Blood Pressure Method: Estimation and Evaluation Using Multiple Body Sensors

Accepted Manuscript

A Chest-Based Continuous Cuffless Blood Pressure Method: Estimation and Evaluation Using Multiple Body Sensors Fatemeh Heydari, Malikeh P. Ebrahim, Jean-Michel Redoute, Keith Joe, Katie Walker, Mehmet Rasit Yuce PII: DOI: Reference:

S1566-2535(19)30193-9 https://doi.org/10.1016/j.inffus.2019.07.001 INFFUS 1129

To appear in:

Information Fusion

Received date: Revised date: Accepted date:

8 March 2019 7 June 2019 7 July 2019

Please cite this article as: Fatemeh Heydari, Malikeh P. Ebrahim, Jean-Michel Redoute, Keith Joe, Katie Walker, Mehmet Rasit Yuce, A Chest-Based Continuous Cuffless Blood Pressure Method: Estimation and Evaluation Using Multiple Body Sensors, Information Fusion (2019), doi: https://doi.org/10.1016/j.inffus.2019.07.001

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

1

/ Information Fusion - - (2019) 1–15

Highlights • High accuracy of bio-impedance (BImp)-based PAT in BP calculation • Validations and measurements of BP from 41 participants’ in different conditions • BP estimation with five different PAT extraction methods • Evaluations of combinations of different BP calculation mathematical models

AC

CE

PT

ED

M

AN US

CR IP T

• Comparisons between BImp-based and commonly used PPG-based PATs in BP extraction

1

ACCEPTED MANUSCRIPT

Information Fusion Information Fusion - - (2019) 1–15

CR IP T

A Chest-Based Continuous Cuffless Blood Pressure Method: Estimation and Evaluation Using Multiple Body Sensors

AN US

Fatemeh Heydari1 , Malikeh P. Ebrahim1 , Jean-Michel Redoute1 , Keith Joe2 , Katie Walker2,3 , Mehmet Rasit Yuce1,∗

Abstract

PT

ED

M

Blood pressure (BP) is a critical vital sign in health, measured millions of times per day worldwide. Current BP measurement requires cumbersome tools, is painful and can be inconvenient. Non-invasive cuffless BP measurement based on pulse arrival time (PAT) techniques allow an alternative way of monitoring BP in healthcare settings with refined wearability and user-friendly features. PAT extraction requires at least two measurements, one as a time reference and another to obtain time delay; there are several approaches to calculate the PAT from various sensors placed on the body. Commonly used signals are electrocardiography (ECG) and photoplethysmography (PPG), which can be recorded from a patients body using more than two separate sensors attachment set-ups. In this work, cuffless BP calculation based on five different PAT readings using Bio-impedance (BImp) at the shoulder as an alternative to PPG, has been investigated. Sensor placement is on the patients chest; which hides them beneath the patient’s clothes making them more suitable for ambulatory monitoring systems. Technology performance was assessed using different postures, exercises and Glyceryl Trinitrate (GTN) spray doses; which provided stable, rising and falling BPs for evaluation. Data were collected from 41 participants who were sitting, standing and supine. Twenty-four of 41 participants undertook experiments including a handgrip task (isometric exercise), three periods of cycling on an exercise bike with light, moderate and heavy resistance settings and an observed rest period at the end. The remaining 17 of 41 subjects received GTN spray for predefined times with variable recovery periods afterwards. Different methods of PAT extraction from BImp data were compared for accuracy. Comparisons were made between PAT readings alone and PAT combined with Heart Rate and the combination model performed better when calculating BP. Simultaneously, data were collected using PPG-based PATs compared to BImp-based PATs. BImp-based PATs proved 3% more accurate than PPG-based PATs, demonstrating the potential superiority of BImp-based BP calculations.

CE

Keywords: Cuffless Blood Pressure, Pulse Arrival Time, On-body sensors, Bio-Impedance.

1. Introduction

AC

Blood pressure (BP) is a vital sign that is measured to detect cardiovascular disease or the impact of other diseases on the cardiovascular system [1]. Worldwide, cardiovascular diseases and stroke are the leading cause of death (15 million deaths in 2016) [2]. In both acute and chronic health settings, BP is measured in the majority of patient consultations, hence in billions of consultations per annum. Currently measurement of BP is usually undertaken using a sphygmomanometer [3], which uses a cuff worn on the upper arm or a Finapres device [4], where the cuff is worn on the finger. This is painful for patients when inflated and a proportion of patients can not tolerate the device. Having ∗ Corresponding

author Email address: [email protected] (Mehmet Rasit Yuce) 1 Department of Electrical and Computer Systems Engineering, Monash University, Melbourne, Australia. 2 Emergency Department, Cabrini Health, Melbourne, Australia. 3 Department of Epidemiology and Preventive Medicine, Monash University, Melbourne, Australia.

2

ACCEPTED MANUSCRIPT

3

/ Information Fusion - - (2019) 1–15

AC

CE

PT

ED

M

AN US

CR IP T

continuous BP measurements (in the critically ill or when a patient is under longer term observation) is cumbersome and causes significant pain and trauma to the limb. The alternative is an invasive arterial catheter (arterial line) for patients in intensive care settings, which provides continuous readings at the cost of marked discomfort, a need to remain completely immobilised, a risk of local and disseminated bacterial infections and occasional arterial occlusion and limb loss [5, 6]. There is a great need for a comfortable, non-invasive, lightweight, wearable and accurate device to replace our current technology [3]. There has been a tremendous interest in new technologies recently. Currently, reliable vital wearables exist for pulse oximetry, telemetry, temperature, respiratory rate but not BP [7, 8]. BP measurement technology development has been focused on methods that capture the time it takes for a pulse of blood to travel from the heart to the organ where measurement is undertaken. This is measured as either pulse arrival time (PAT) or pulse transition time (PTT) [9, 10, 11, 12, 13], and BPs are calculated by using the Moens-Korteweg and Bramwell-Hill mathematical formulas [14, 15]. To date, the main way of measuring PAT and PTT is by using reflectance mode photoplethysmography (PPG) [16]. This has been the main focus of recent research [11, 12, 17, 18, 19, 20, 21]. There are multiple widely available sensors on the market and there is an emerging body of knowledge about how the sensors work and their limitations. Limitations include suboptimal accuracy compared to gold-standard devices and the requirement for at least two separate attachments on the patients body such as chest (for ECG), neck, ear and arm (for one or two PPG readings) which are externally visible when worn [19, 20]. Other less well-known ways of measuring PAT and PTT include electrical bio-impedance (BImp) [22], ballistocardiogram (BCG) [23], seismocardiogram (SCG) [24, 25] and ultrasonic technology [9, 10]. BCG and SCG research requires specialised technology; which is challenging to obtain. Despite the focus on PPG technologies, early pilot work on BImp shows promise as sensors can be placed across patients shoulders (measuring central elastic arterial changes), not peripherally, removing the need for peripheral arterial readings . Peripheral arterial readings are subject to vasomotion-induced inaccuracies, particularly in those patients who have long-term blood pressure problems and most need accurate measurements. In addition, sensors could potentially be worn discretely under clothing. Currently there is an incomplete understanding of how PAT (extracted from BImp) relates to BP and whether BImp is worth pursuing as a potential wearable technology; which can target central aortic BP [22, 26]. The BP regulation is a complex mechanism controlled by a large number of parameters (either constant or variable) and considering main influencing factors in invasive BP estimation algorithms may increase the accuracy. Some researchers combine measures of heart rate (HR) with PAT to increase the accuracy of the BP estimation [19, 17, 20]. When the blood flow (determined by heart-pumping action) is resisted, the BP is generated (BP = Cardiac output × Resistant). The HR along with the volume of the blood pumped out of the ventricles (stroke volume) define the cardiac output (cardiac output = stroke volume × HR) [27]. Thus, heart rate variability directly effects the cardiac output and BP. The aim of our investigation was to explore the utility of PAT extracted from BImp to monitor BP, compared with gold-standard oscillating sphygmomanometry (3-minutely readings) and/or Finipres (beat-to-beat readings). This required determination of the best way to extract PAT from BImp and whether accuracy was maintained under a variety of physiological conditions. We evaluated which mathematical method of calculating BP from PAT was most accurate. We also compared PAT readings alone with PAT combined with Heart Rate to determine which performed better in the calculation of BP. Finally, we compared BImp to PPG based PAT to determine whether BImp might potentially replace PPG in the future given the BImp sensor advantages. 2. Methods and Materials 2.1. Subjects and Experimental Details Measurements were recorded from 46 consenting, volunteer adult participants. The study was authorized by the Cabrini Human Research Ethics Committee (07-19-06-17) and registered with the ANZ clinical trial registry (ACTRN12617000774325). Some volunteers were healthy and others suffered from varying degrees of HTN. All subjects attended Cabrini Hospital, Melbourne, Australia, (Oct. to Nov. 2017) and Emergency Physicians monitored the subjects throughout data recording. Recordings commenced with participants sitting at rest. Five participant records were excluded from the data due to hardware errors or participant issues (thus 41 subjects remained). 3

ACCEPTED MANUSCRIPT

4

/ Information Fusion - - (2019) 1–15

Experiment Posture Exercise GTN

Records 41 24 17

Age [yrs] 52±11 50±11 53±12

weight [kg] 89±20 89±21 88±17

Height [cm] 171±11 167±13 172±7

Gender 51% male 50% male 53% male

Table 1: Participants characteristics.

CE

PT

ED

M

AN US

CR IP T

In order to verify different BP trends, the following steps were undertaken during data collection. First, all subjects assumed three different postures, two times each per posture. Then, subjects were divided into two groups and either undertook six exercise tasks (increasing BP) or received GTN spray (decreasing BP). Baseline characteristics of volunteers are presented in Table 1. All 41 subjects wore ECG, BImp and PPG sensors, a calibrated cuffed sphygmomanometer on the left arm and a cuffed beat-to-beat BP measurement tool on the right arm. Emergency Physicians supervised the sensors and BP measurement tools on-body attachments. During posture data collection, subjects sequentially stayed in one of three different poses for three minutes and each pose was repeated twice; thus, six postures occurred per participant. The sensors, cuffed sphygmomanometer and beat-to-beat BP measurement device placement is demonstrated in Figure 1a and 1b. A cuffed sphygmomanometer and beat-to-beat BP measurement devices were worn as comparators for mathematical model parameter extraction. Sensor signals (ECG, BImp, and PPG) and beat-to-beat BP values were recorded during the whole three minutes of each task. The cuffed sphygmomanometer commenced measurement at 120 seconds and returned SBP and DBP reading values after 60 seconds (the first two minutes of each task was considered as the subjects settling time). In the second stage of data collection, 12 males and 12 females (24 of 41 participants) undertook seated exercise on an exercise bike whilst the remaining 17 participants (53% male) rested supine for the GTN experiment. During exercise data collection, subjects undertook six tasks including four exercises followed by 6 minutes of rest and each exercise task lasted three minutes. Exercises included one handgrip (holding a hand grip for two minutes and a one minute rest) and three cycling periods (two minutes cycling and a one minute rest). During cycling tasks, participants aimed for 20-25 km/hr speed with the cycle resistance settings stable at light, moderate then heavy (changing with each new task). Glycerl Trinitrate (GTN) is a medicine used to relax blood vessel muscles, helpful when increased blood flow is required, for example during heart attacks. GTN can be sprayed under the tongue and usually lowers the BP [28]. For GTN trial data collection, dose of GTN and recovery times varied. The Emergency Physician administered GTN doses at 3-minutely intervals with the participant supine, until the BP reduced by 10mmHg or more (or they reached a maximum safe dose) and then subsequently participants remained supine until their BP recovered to baseline and the physician determined they were safe to sit up. At the beginning of every spraying task, the physician sprayed the GTN into each subjects mouth (this took less than 20 seconds). During both exercise and GTN experimental conditions, ECG, BImp, left ear PPG sensors signals and beat-tobeat BP values were recorded during the whole three minutes of each task. Similar to posture data collection, cuffed sphygmomanometer measurement commenced at 120 seconds of every task and recorded the BP value in one minute.

AC

2.2. Signal Processing

In Figure 1c, the block diagram of signal processing is presented. Firstly, all recorded raw signals (ECG, BImp, and PPG) are filtered, using a bandpass filter (BPF) to eliminate respiration, 50 Hz power line noise and motion artifacts. Most of the motion artifacts were from respiration or sudden movements during the exercised experiment. The sudden movements and respiration were rejected using low pass and high pass filtering, respectively. The remaining artifacts (passed through the filter) created an error in the characteristic extraction. This error was corrected with moving average filtering on PATs (section 2.3). A Chebyshev type II BPF was designed with the Matlab filter design toolbox including calculating pass and stop frequencies separately, based on heart rate frequency (HRF) and signals frequency range around it. The effect of motion and respiration on the PPG signal was less than that of on BImp. Therefore HRF is obtained using the fast 4

ACCEPTED MANUSCRIPT

5

/ Information Fusion - - (2019) 1–15

Raw

BImp

ECG

1

1

0

0 1

Filtered

1 0

0 118

120

77

CR IP T

116

Time [s]

(a)

Time [s]

(d) Sitting

Heavy Exercise

PPG BImp ECG

1

1 0 1

0 1 0

0 1

0

0

90

HR

AN US

1

120 100

85 0

5

Time [s]

(b)

0

5

Time [s]

(e)

0.5

PAT D1

ECG wave

PAT Max

BImp wave

PAT Int.Tan.

ED

0

R pk

M

0.2 0

PAT ft

1st Derivative

0.5 0

PT

0.5 0

(c)

81

2nd Derivative

PAT D2 0.5

Time [s]

0.7

(f)

AC

CE

Figure 1: (a) Sensors, cuffed sphygmomanometer and Finapres (finger cuffed beat-to-beat BP measurement device) placement on the body; (b) Placement of the sensors and BP measurement devices on the body for a participant on an exercise bike; (c) Signal processing block diagram; (d) recorded (raw) and filtered ECG and BImp signals for subject 1, during first sitting posture; (e) Filtered ECG, BImp and PPG signals, and extracted HR samples captured during first sitting posture and heavy exercise for subject 1; (f) PAT feature selection.

Fourier transform (FFT) of the PPG signal. The floor of the first FFT’s harmonic in the domain of PPG is almost equal to HRF. The frequency ranges are determined below: f passlow EGC = HRF, f passhigh ECG = 30Hz , f passlow BImp = HRF, f passhigh BImp = 2.5. × HRF

(1)

For PPG, frequency ranges are the same as BImp. In Figure 1d, samples of both ECG and BImp signals before and after filtering for subject one, during the first sitting position is shown, respectively. Next, R-peaks are extracted using Algorithm .1. The recorded signals of experimental conditions contain movement artifacts causing large amplitude changes in ECG signal (which even could pass through the bandpass filter). 5

ACCEPTED MANUSCRIPT

6

/ Information Fusion - - (2019) 1–15

Algorithm 1 R-peaks extraction procedure Inputs: ECG Outputs: R-peaks Initial R-peaks: the local maximum peaks of the ECG signal which have the 1/ f passlow .ECG minimum distance from each other with 0.5 (very small value to consider almost all peaks) minimum voltage drop from each side (peak prominence). To calculate, first find all local peaks using first derivative of the signal. Exclude the peaks with less than 0.5 peak prominence. Sort the peaks based on their amplitude from high to low. In order, measure the time distance of each peak with the next peak. If the distance is less than minimum distance (1/ f passlow .ECG ) exclude the related peak. Store the remaining peaks initial R-peaks. 5: R-peaks amplitude mean values: the mean value of ECG at extracted initial R-peaks in the previous step. 6: Movement artifice clearing: Eliminate parts of the ECG signal which have a voltage higher than ECG mean value + R-peaks amplitude mean values or less than ECG mean value - R-peaks amplitude mean values. 7: Final R-peaks: the local maximum peaks of movement cleared ECG signal which has the 0.5/ f passlow .ECG minimum distance from each other with 0.5×R-peaks amplitude mean values peak prominence (minimum voltage drop from each side). Use the same algorithm presented in state 4.

AN US

CR IP T

1: 2: 3: 4:

This problem may make R-peak detection algorithms unable to detect some peaks. The algorithm presented above is simple and adjusts the detection with varying parameters based on input’s HR and amplitude. It is easy to develop for whole data and works quite well for collected signals in this study. 2.3. Feature Selection

In this paper, the time delay between the R-Peak of ECG signal and five landmarks on the BImp pulse wave are defined as PATs. These five landmarks are extracted using the following techniques:

PT

ED

M

1. Foot value (PAT f t ), the maximum value of pulse wave. 2. First derivative maximum value (PATD1 ), the maximum value in first derivative (maximum positive gradient) of pulse wave. The maximum of the pulse wave signals upward trend rate, related to the peak velocity of the vessels wall. 3. Foot value (PATD2 ), the maximum value in second derivative of pulse wave (point of inflection). This value is related to the maximum velocity of the vessels wall. 4. Intersecting tangents value (PATInt.T an. ), the intersection point of the tangent to the maximum gradient and tangent of the foot. Note that, since by definition the gradient at the foot is zero, the tangent to the foot is horizontal. 5. Maximum value (PAT Max ), the maximum value of pulse wave.

AC

CE

PAT selection characteristics using BImp wave are shown in Figure 1f. The features, B-PAT f t , B-PATInt.T an. , B-PATD1 and B-PAT Max are defined as the time interval between the R-Peak and the the landmarks points of the BImp signal. In addition, P-PAT f t , P-PATInt.T an. , P-PATD1 and P-PAT Max are defined as the time interval between the R-Peak and the landmarks points of the PPG signal. For each data collection group, two sets of procedures are performed in two cases: continuous (beat-to-beat) and cuff-based. In the continuous case, BP is extracted from beat-to-beat recorded values using a Finapres device [4, 29] and its corresponding PATs are derived from each beat of the ECG and BImp signals. The PAT values of each recording set (each tasks records) in experimental conditions are extracted by applying zero-crossing detection to both BImp and PPG signals first and second derivatives. Out of range values are omitted with adding minimum and maximum thresholds to the results. Then PATs are passed through a moving average filter with a size of five for postures and 10 for exercise and GTN. In cuff-based data, for each task of each experiment (posture, exercise, and GTN), one cuff measurement represents the BP value and one PAT value is calculated from the average of the beat-to-beat PATs obtained 30 seconds before to 30 seconds after cuffed sphygmomanometer commencement.

6

ACCEPTED MANUSCRIPT

7

/ Information Fusion - - (2019) 1–15

3. Results and Discussion It is implausible that the SBP value of a normal human body fluctuates more than ±10mmHg in two continuous readings. Variations more than ±15mmHg may demonstrate a measurement fault or anatomical abnormality [30]. The accuracy of the cuffed sphygmomanometer device used in this study was ±3mmHg [31]. Therefore, there are no cuff values with more than ±20mmHg (>15+3mmHg) changes comparing to their previous and next records along with their relevant PATs in data. In practice, 36 of 475 records (7.5%) have been removed, as this was a considered a measurement error.

CR IP T

3.1. Data Fusion

AC

CE

PT

ED

M

AN US

For each participant, posture data were combined into three groups: sitting, standing and supine. All six exercise tasks (handgrip, light exercise, medium exercise, heavy exercise, rest1, and rest2) were examined separately. Regarding GTN data, BP responses were inconsistent (cuffed sphygmomanometer results and beat-to-beat device continuous analysis). Some participants had the expected BP dose-response (each spray lowered systolic BP a little more). Others had no response, or a delayed response (or even a lower BP during recovery compared to during medicated periods). We assumed, that if a GTN cuff-based systolic BP was lower than the preceding supine cuff-based systolic BP, (just prior to commencing GTN), then the drop was due to the medication. We created two groups of records; medicated and normal. The normal group data includes participant records following GTN exposure, where the systolic BP failed to drop. Firstly, the pattern of each of the five different landmarks extracted from BImp wave is investigated within cuffbased data. The trend of all B-PATs due to posture changing, increasing exercise stress and GTN usage are compared in Figures 2a, c and e, respectively. Overall, all B-PATs follow a similar inclination during each experiment. As demonstrated, each type of B-PAT has a related reaction to posture, changing with a shift from sitting to standing (overall less than 11ms decrease) and a noticeable drop, around 40ms, from standing to supine. During exercise, handgrip progresses towards heavy exercise and B-PAT values decrease around 29% and return towards normal values during resting periods with an overall elevation of 49%. For data extracted from GTN exposed participants (normal and medicated), PAT values increase by 10% (overall 6ms) from normal to medicated. For all five PATs, similar trends occurred during each experiment. Therefore, only the B-PATD1 trend was compared to the trends extracted from PPG signals and the cuffed BP (SBP and DBP) values. Figure 2b compares the change trend of all participants BImp-based PATD1 , PPG-based PATD1 and Cuff BP values during changing posture. Both B-PATD1 and P-PATD1 after a short drop from sitting to standing, drop significantly during the supine posture (with a 34ms and 13ms decrease in B-PATD1 and P-PATD1 , respectively). From sitting to standing, the SBP changes by less than 0.0001% and DBP rises by 0.07%, while, similar to PATs, they decrease on an average of 2mmHg and 14mmHg after undertaking supine posture, respectively. Note that the similar trend in responding to positions, should not conflict with the negative relation of the PAT and BP. Figure 3a shows the SBP and DBP records by a cuffed device, are based on their corresponding extracted B-PATD1 values, for all participants. There is a clear difference between the supine position in comparison with the other two postures, because of both B-PATD1 and BP reduction. As it can be seen, although BP and B-PATD1 have a comparable reaction to posture switching, both of them stay negatively correlated with each other. Similarly, the comparison of the B-PATD1 with P-PATD1 against cuff-based BP values using participant data throughout exercise experiments, is shown in Figure 2d. It can be seen that, as exercise stress increases from handgrip to heavy exercise, both B-PATD1 and P-PATD1 decrease from 172ms to 117ms and 234ms to 193ms, respectively, in contrast to the cuff-based SBP and DBP (rise from 136mmHg to 154mmHg in SBP and 74mmHg to 86mmHg in DBP) values. After completing exercise and during the two rest periods, the B-PATD1 and P-PATD1 rise around 48.6% and 21.46%, respectively, while cuff SBP and DBP drop back by almost 0.1% to lower values. In Figures 3c and 3d, the similarity of the cuff-based and continuous B-PATD1 values in exercise data are shown with their related BP values. It can be seen that the highest BP values (and the lowest B-PATD1 values) are recorded during exercise (mostly heavy) and the lowest BP values (and the highest B-PATD1 values) are obtained during the two finishing rests. It is quite clear, that B-PATD1 and BP, in rising trends of the BP, are fully reversed. Figures 2f, 3e and 3f, represent the properties of the subjects who undertook GTN throughout data collection. As with the other two experiments, during GTN (BP falling from normal to medicated) the B-PATD1 and P-PATD1 are rising (an overall rise from 128ms to 205 ms for B-PATD1 and from 134ms to 207ms for P-PATD1 ). There are 7

ACCEPTED MANUSCRIPT

8

CR IP T

/ Information Fusion - - (2019) 1–15

(b)

AN US

(a)

(d)

PT

ED

M

(c)

(e)

(f)

CE

Figure 2: The trend of B-PATs in different postures (a), exercises (c) and GTN experiment (e). The comparison between the trends of cuff-based B-PATs, P-PATs, SBP and DBP for postures (b), exercises (d) , and GTN experimental conditions (f).

AC

simultaneous cuff-based drops of 5mmHg (134 to 129mmHg SBP) and 7mmHg (89 to 82mmHg DBP) when GTN is administered. It is clear that, as BP decreases, both PATs increase and the values of B-PATD1 and both cuff-based and continuous BP are negatively correlated. Figure 4a shows a sample of extracted beat-to-beat B-PATD1 values with their corresponding SBP and DBP values (obtained using a calibrated Finapres device) using posture data for subject 1. Likewise, in Figure 4b a sample of B-PATD1 and continuous Finepres BP shows exercise data from subject 15. As it can be seen, during exercise tasks that the BP increases and during the rest period, BP starts to fall to normal values. Transient BP drops within exercises tasks are demonstrated and occur during cuff device measurements (subjects were asked to stop cycling in this period). In Figures 4c and 4d, two samples using the GTN data of subjects 30 and 35, are shown. The reaction of subjects to the GTN is clearly different. In subject 30 there are local drops and rises which occur just after each spray. Simultaneously, the DC level of the BP decreases gradually until the end of the second recovery (Figure 4c). 8

ACCEPTED MANUSCRIPT

9

/ Information Fusion - - (2019) 1–15

Cuff DBP [mmHg]

120

0.1

80

40

0.2

0.1

B-PATD1 [s]

0.2

B-PATD1 [s]

(a)

100 0.1

0.15

100 80 60 40

0.2

0.1

0.15

0.2

B-PATD1 [s]

B-PATD1 [s]

(c)

0.1

0.2

90

M

120

70

50

200

100 0.1

(e)

0.2

Beat-to-Beat PATD1 [s]

0.1

0.2

100

0.1

Medium Exercise Rest2

120

80

40

0.1

0.2

Beat-to-Beat PATD1 [s]

(d)

200

B-PAT1 [s]

B-PAT1 [s]

Light Exercise Rest1

Normal

ED

160

80

Medicated

Cuff DBP [mmHg]

Cuff SBP [mmHg]

Normal

Beat-to-Beat SBP [mmHg]

150

120

Handgrip heavy Exercise

300

AN US

200

Medium Exercise Rest2

Beat-to-Beat SBP [mmHg]

Light Exercise Rest1

Cuff DBP [mmHg]

Cuff SBP [mmHg]

Handgrip heavy Exercise

(b)

Beat-to-Beat DBP [mmHg]

160

Supine

120

CR IP T

Cuff SBP [mmHg]

Standing

0.2

Beat-to-Beat DBP [mmHg]

Sitting

200

Beat-to-Beat PATD1 [s]

Medicated

80

40

0.1

0.2

Beat-to-Beat PATD1 [s]

(f)

CE

PT

Figure 3: All participants cuff-based (a) and beat-to-beat (continuous) (b) SBP and DBP values based on their B-PATD1 values during postures experiment. All exercise experiment’s cuff-based (c) and beat-to-beat (d) SBP and DBP values based on their related B-PATD1 . The GTN data cuff-based (e) and beat-to-beat (f) SBP and DBP values based on their related B-PATD1 .

AC

In contrast, in subject 35 the BP starts to fall at the end of the first spray, comes back to normal values during the third spray and stays around that area until the middle of the first recovery. Then it reduces again and returns to normal at the end of the second recovery. 3.2. BP estimation The BP calculation based on PAT requires a predefined mapping model between BP and PAT values. Also, different PATs (landmarks) can be combined with other parameters, such as HR, to increase the accuracy of the BP estimation. In this study, the first stage of BP calculation and the accuracy of the BP estimation are compared between the five different landmarks using the following mathematical models. Inverse models: a + b, PAT a BP = + b, PAT 2 9 BP =

(2a) (2b)

ACCEPTED MANUSCRIPT

10

/ Information Fusion - - (2019) 1–15

DBP [mmHg] SBP [mmHg] B-PATD1 [s]

DBP [mmHg] SBP [mmHg] B-PATD1 [s]

0.15

0.1 3

6

9

12

15

120 100 2

60

6

4

6

13

50 40

0

3

6

9

Time [min]

12

15

0.1 250 200 150 80 70 0

18

120 110

M

100 65 60 15

18

ED

12

Time [min]

(c)

9

Time [min]

12

15

Recovery 1,2

AN US

0.12

9

6

Beat-to-Beat Spray 1,2,3

DBP [mmHg] SBP [mmHg] B-PAT [s] D1

[s] D1

DBP [mmHg] SBP [mmHg] B-PAT

0.13

6

3

(b)

Beat-to-Beat Spray 1,2,3,4 Recovery 1,2,3

3

Rest 1,2

0.15

(a)

0

Beat -to-Beat Light, Medium & Heavy Exercise

Hand grip

Supine1,2

CR IP T

Beat-to-Beat Standing1,2

Sitting1,2

0.16

140 130 120

70 60 0

3

6

9

Time [min]

12

(d)

PT

Figure 4: Beat-to-beat B-PATD1 , SBP and DBP sample (a) during posture tasks of subject 1, (b) during exercise tasks of subject 15, (c) during GTN tasks of subject 30, and (d) during GTN tasks of subject 35.

AC

CE

where the a and b coefficients are obtained using the least squares data fitting method for each subject. These mathematical models indicate a subject-specified inverse relationship between BP and PAT. In Table 2a, the accuracy percentage (≤10mmHg error) of SBP and DBP estimation using Eq. 2 for cuff-based values is compared between five different PATs extraction methods. As it can be seen, the overall estimation accuracy for DBP is higher than SBP (almost 6% in each estimation) and B-PATD1 is the most correlated PAT to both the SBP and DBP values. For both SBP and DBP estimation Eq. 2a presents a higher performance compared to that of using Eq. 2b. The highest accuracy percentage of cuff SBP and DBP estimation with ≤10mmHg error is recorded using BPATD1 and Eq. 2a model. For posture, SBP and DBP are calculated with 89.43% and 89.02% accuracy, respectively. The accuracy percentage for exercise is 83.33% in SBP and 90.28% in DBP, and for GTN, 80.72% and 86.75% are obtained to calculate SBP and DBP, respectively. The combination of two inverse equations (second order inverse model) can increase the flexibility of the BP calculation and may lead to more accurate performances. The three-dimensional models investigated in this study are as below.

10

ACCEPTED MANUSCRIPT

11

/ Information Fusion - - (2019) 1–15 100

-100 50

mean= 0.002

100

150

200

250

50 30 10 0 -10 -30 -50

mean= -0.0005

-100

100

Continuous SBP [mmHg]

200

300

Continuous SBP [mmHg]

(a)

(b)

GTN

30 10 mean= 0 -0.0014 -10 -30 -50 50

CR IP T

50 30 10 0 -10 -30 -50

50

Exercise Difference [mmHg]

Postures Difference [mmHg]

Difference [mmHg]

100

100

150

200

Continuous SBP [mmHg]

(c)

Figure 5: The errors of continuous SBP estimation using Eq. 5a with B-PATD1 and B-PAT Max in (a) different postures, (b) exercises, and (c) GTN tasks.

AN US

Second order models: a b + + c, PAT 1 PAT 2 b a + + c. BP = 2 PAT 1 PAT 22 BP =

(3a) (3b)

CE

PT

ED

M

The HR is calculated from the R-peak to R-peak interval in the ECG signal and adding HR to the BP calculation, accuracy is improved. The following equations were developed, using the combination of the previous models with the addition of HR. Additional HR models: a + b × HR + c, (4a) BP = PAT a BP = + b × HR + c, (4b) PAT 2 a b + + c × HR + d, PAT 1 PAT 2 a b BP = + + c × HR + d, 2 PAT 1 PAT 22 BP =

(5a) (5b)

AC

where the a, b, c and d coefficients are obtained using a least square data fitting method. The results of BP estimation using Eq. 3, Eq. 4 and Eq. 5 are compared to that obtained using PPG signal and reported in Table 2b. For Eq. 4 PATD1 and for second order models, PATD1 and PAT Max are used to obtain BP. The accuracy increases using two PATs with Eq. 3a presenting an equal improvement to that of using Eq. 3b with an overall rise of around 2% for each experimental data. Adding HR to the mathematical models increases the accuracy of each model by around 9% and the most accurate model for BP calculation using BImp is Eq. 5. The highest accuracy for SBP calculations is 99.12% (posture), 94.53% (exercise) and 94.44% (GTN). The highest accuracy for DBP is 100% (posture), 98.4% (exercise) and 100% (GTN) less than 10mmHg error. The table clearly demonstrates that the performance of the B-PATs, on average, is 2% higher than that of P-PATs (PPG based) when compared to cuff-based data. For continuous BP calculation, according to the outcomes of the cuff-based data, the performance of the Eq. 5 is investigated and the estimation results are reported in Table 2c. As it can be seen, similar to cuff-based data, the performance of the BImp-based PATs continuous BP calculation is more accurate when compared to that based on PPG by 3% overall in each experiment. Eq. 5 can provide the highest accuracies of 86.88%, 77.97% and 92.61% in continuous SBP, and 98.07%, 96.21% and 99.45% in continuous DBP estimation for posture, exercise, and GTN 11

ACCEPTED MANUSCRIPT

12

120 100 100

120

140

160

180

Cuff-based measured SBP [mmHg]

Data Fit X=Y

200

150

100 100

150

(b)

200 150 100 50 50

100

150

200

250

200 150 100 50 50

100

150

(d)

200

250

300

Cuff-based estimated SBP [mmHg]

M

Cuff-based measured SBP [mmHg]

Data Fit x=y

250

140 120 100

100

(e)

120

140

160

Cuff-based measured SBP [mmHg]

(c)

Exercise r = 0.94957

300

160

200

GTN r = 0.97033

Data Fit x=y

150

AN US

Data Fit x=y

Cuff-based estimated SBP [mmHg]

Cuff-based estimated SBP [mmHg]

(a) Posture r = 0.94679 250

200

Cuff-based measured SBP [mmHg]

GTN r = 0.96744 Data Fit x=y

CR IP T

140

Exercise r = 0.96216

Cuff-based estimated SBP [mmHg]

160

Posture: r = 0.98158 Data Fit X=Y

Cuff-based estimated SBP [mmHg]

180

Cuff-based estimated SBP [mmHg]

Cuff-based estimated SBP [mmHg]

/ Information Fusion - - (2019) 1–15

100

50 50

100

150

200

Cuff-based estimated SBP [mmHg]

(f)

ED

Figure 6: The results of linear regression analysis between measured SBP and estimated SBP values (using Eq. 5a with B-PATD1 and B-PAT Max ) with extracted r-squared values. (a), (b) and (c) for cuff-based data and (d), (e) , and (f) for continuous data.

CE

PT

experimental conditions, respectively. The beat-to-beat SBPs estimation errors using the Eq. 5a for every experiment, are shown in Figures 5a, b and c. The results of the linear regression analysis between measured SBP and estimated SBP values (using Eq. 5a with B-PATD1 and B-PAT Max ) with their extracted r-squared values for both cuffed-based and continuous SBPs are shown in Figure 6. Overall we observe higher than r = 0.94 value for all experimental conditions. This illustrates high similarity between estimated SBP values and measured values in both cuff-based and continuous cases. 4. Conclusions

AC

This paper describes how cuffless systolic and diastolic BP calculations were evaluated using bio-impedancebased PAT extraction, which was calculated from five different landmarks on the rising slope of the pulse wave. The BImp signals were obtained from the subclavian and carotid arteries (sensors placed at the shoulder). Three experiments were performed to validate the estimation using a variety of conditions to change participant BP: varied resting postures (sitting, standing, supine); increasing exercise levels (handgrip, cycling at light, moderate, high intensity, resting) and variable doses of GTN medicine, whilst capturing continuous (beat-to-beat) and cuffed-based BP values for comparison. Forty-one participants contributed posture records, 24 exercised and 17 received GTN spray. We investigated five different BImp-based PATs trends, where every PAT showed similar variations during different experiments. Also, the responses of the cuff-based BImp PATD1 (the maximum value in the first derivative of pulse wave) were compared data from the PPG signal, along with cuffed recorded SBP and DBP values. Both the BImp 12

ACCEPTED MANUSCRIPT

13

/ Information Fusion - - (2019) 1–15

DBP SBP Exercise DBP SBP GTN DBP

ft 78.45 78.05 80.08 80.08 69.44 68.75 87.5 86.81 68.42 66.67 75.42 75.44

D2 87.40 87.40 89.83 89.84 79.17 79.86 89.58 90.28 75.90 75.90 83.13 83.13

(a)

PT

ED

DBP

CE

SBP

DBP 1

AC

2

Algorithm Eq. (5a) Eq. (5b) Eq. (5a) Eq. (5b)

BImp Exr2 84.72 84.72 83.33 83.33 94.53 94.53 90.97 90.97 91.67 91.67 98.44 98.44

Pst1 86.88 86.88 98.07 98.07

D1 89.43 89.02 89.02 89.02 83.33 81.25 90.28 90.28 80.72 80.72 86.75 86.75

Max 89.02 89.02 88.21 88.21 80.56 79.86 90.28 90.28 77.11 77.11 86.75 86.75

GTN 84.34 84.34 84.34 84.34 94.44 94.44 86.74 86.75 86.75 86.75 100 100

Pst 91.46 91.47 90.65 90.65 99.12 99.12 91.06 91.06 91.46 91.46 100 100

PPG Exr 84.03 83.33 79.17 79.17 83.33 94.53 90.97 90.97 91.67 91.67 90.97 98.44

GTN 83.13 83.13 83.13 83.13 93.30 93.30 86.74 86.74 86.75 86.75 100 100

GTN 92.61 92.57 99.45 99.45

Pst 83.33 83.35 93.67 93.67

PPG Exr 71.95 71.93 85.61 85.62

GTN 90.81 90.74 96.34 96.34

AN US

Pst1 91.87 91.87 90.65 90.65 99.12 99.12 92.28 92.28 92.27 92.27 100 100

M

SBP

Algorithm Eq. (4a) Eq. (4b) Eq. (3a) Eq. (3b) Eq. (5a) Eq. (5b) Eq. (4a) Eq. (4b) Eq. (3a) Eq. (3b) Eq. (5a) Eq. (5b)

Int. T an. 89.02 88.62 90.65 90.65 78.47 77.08 87.5 87.5 75.90 75.90 86.75 86.75

CR IP T

SBP Postures

Algorithm Eq. (2a) Eq. (2b) Eq. (2a) Eq. (2b) Eq. (2a) Eq. (2b) Eq. (2a) Eq. (2b) Eq. (2a) Eq. (2b) Eq. (2a) Eq. (2b)

(b)

BImp Exr2 77.97 77.92 96.21 96.20

Posture data. Exercise data. (c)

Table 2: BP estimation accuracy percentages in ≤10mmHg for posture, exercise and GTN data. (a) using Eq. (2a) and Eq. (2b) for cuff-based five different landmarks. (b) using Eq. (3), Eq. (4) and Eq. (5) with cuff-based PATD1 and PAT Max for BImp and PPG signals. (c) Continuous BP estimation using Eq. (5) and continuous PATD1 and PAT Max .

PATD1 and the PPG-based PATs behaved similarly and were negatively correlated to BP values in every task. We observed continuous and cuff-based BPs and simultaneously recorded B-PAT values, for each experiment. Measured data obtained from the GTN medicated and normal groups was analyzed considering cuff-based recorded BP values and a clear difference between the two groups was visible. However, in the beat-to-beat case, BP had dif13

ACCEPTED MANUSCRIPT

14

/ Information Fusion - - (2019) 1–15

CR IP T

ferent local rises and falls from subject to subject and grouping them similar to cuff-based values, made it difficult to classifying medicated and normal data from each other. This requires additional study to understand the impacts of GTN on BP results in various conditions in future works. We investigated BP extraction using different mathematical models combined with HR for each experiment and compared all the results obtained from BImp with the same implementation on PPG signal. When using only PAT to calculate BP, PATD1 was the most accurate PAT extraction method. Using two different PAT landmarks and by adding HR, accuracy improved compared to both cuff-based and continuous BP estimation. Overall, the results of BImp-based calculations were more accurate than PPG-based calculations. Whilst not formally tested, our subjects reported that wearing the more accurate BImp electrode dots was more comfortable than wearing the PPG sensor. Acknowledgments

AN US

The authors would like to thank T. Wu and previous members of the Biomedical Integrated Circuits and Sensors (BICS) Laboratory, Monash University for their help in the sensor design of prototype. The authors would like to thank P. Howley and M. Hebblewhite from Planet Innovation for their helpful discussions, and all volunteers for assistance as human subjects. This work is supported by the Victorian Government through the Future Industry Fund Sector Growth Program Stream 1. The work is partially also supported by a ARC linkage (LP160101823). The authors gratefully acknowledge the support of the Monash Institute of Medical Engineering for this project. References

AC

CE

PT

ED

M

[1] A. L. Valderrama, S. C. King, M. G. George, Y. Hong, E. Gregg, Vital signs: Awareness and treatment of uncontrolled hypertension among adultsunited states, Centers of Disease Control and Prevention, Morb. Mortal. Wkly. Rep. 61 (2012) 703709. [2] Top 10 causes of death, Global Health Observatory (GHO)data, World Health Organization,Geneva, 24 May 2018. [3] J. E. Sharman, T. H. Marwick, Accuracy of blood pressure monitoring devices: a critical need for improvement that could resolve discrepancy in hypertension guidelines, Journal of human hypertension (2018) 1. [4] J. Penaz, Photoelectric measurement of blood pressure, volume and flow in the finger, Digest of the 10th International Conference on Medical and Biological Engineering, Dresden: International Federation for Medical and Biological Engineering (1973) 104. [5] N. H. Petersen, S. Ortega-Gutierrez, A. Reccius, A. Masurkar, A. Huang, R. S. Marshall, Comparison of non-invasive and invasive arterial blood pressure measurement for assessment of dynamic cerebral autoregulation, Neurocritical Care 20 (1) (2014) 60–68. [6] C. Kornbau, K. C. Lee, G. D. Hughes, M. S. Firstenberg, Central line complications, International Journal of Critical Illness and Injury Science 5 (3) (2015) 170178. [7] G. Fortino, R. Giannantonio, R. Gravina, P. Kuryloski, R. Jafari, Enabling effective programming and flexible management of efficient body sensor network applications, IEEE Transactions on Human-Machine Systems 43 (1) (2013) 115–133. [8] R. Gravina, P. Alinia, H. Ghasemzadeh, G. Fortino, Multi-sensor fusion in body sensor networks: State-of-the-art and research challenges, Information Fusion 35 (2017) 68–80. [9] S. R. Steinhubl, E. J. Topol, A skin patch for sensing blood pressures, Nature Biomedical Engineering 2 (2018) 633–634. [10] C. Wang, X. Li, H. Hu, L. Zhang, Z. Huang, M. Lin, Z. Zhang, Z. Yin, B. Huang, H. Gong, et al., Monitoring of the central blood pressure waveform via a conformal ultrasonic device, Nature Biomedical Engineering 2 (9) (2018) 687. [11] K. Matsumura, P. Rolfe, S. Toda, T. Yamakoshi, Cuffless blood pressure estimation using only a smartphone, Scientific reports 8 (1) (2018) 7298. [12] X. Ding, B. P. Yan, Y.-T. Zhang, J. Liu, N. Zhao, H. K. Tsang, Pulse transit time based continuous cuffless blood pressure estimation: A new extension and a comprehensive evaluation, Scientific reports 7 (1) (2017) 11554. [13] N. Boubouchairopoulou, A. Kollias, B. Chiu, B. Chen, S. Lagou, P. Anestis, G. Stergiou, A novel cuffless device for self-measurement of blood pressure: concept, performance and clinical validation, Journal of human hypertension 31 (7) (2017) 479. [14] J. C. Bramwell, M.B., M.R.C.P., A. Hill, F.R.S, The velocity of pulse wave in man, Proceedings of the Royal Society of London B: Biological Sciences 93 (652) (1922) 298–306. [15] C. Vlachopoulos, M. O’Rourke, W. W. Nichols, McDonald’s Blood Flow in Arteries: Theoretical, Experimental and Clinical Principles, CRC Press, 29 July, 2011. [16] T. Tamura, Y. Maeda, M. Sekine, M. Yoshida, Wearable photoplethysmographic sensorspast and present, Electronics 3 (2) (2014) 282–302. [17] Q. Zhang, D. Zhou, X. Zeng, Highly wearable cuff-less blood pressure and heart rate monitoring with single-arm electrocardiogram and photoplethysmogram signals, Biomed Eng Online. [18] X. Ding, Y. Zhang, J. Liu, W. Dai, H. K. Tsang, Continuous cuffless blood pressure estimation using pulse transit time and photoplethysmogram intensity ratio, IEEE Transactions on Biomedical Engineering 63 (5) (2016) 964–972. [19] D. Buxi, J.-M. Redout, M. R. Yuce, A survey on signals and systems in ambulatory blood pressure monitoring using pulse transit time, Physiological Measurement 36 (3) (2015) R1. [20] M. Sharma, K. Barbosa, V. Ho, D. Griggs, T. Ghirmai, S. K. Krishnan, T. K. Hsiai, J.-C. Chiao, H. Cao, Cuff-less and continuous blood pressure monitoring: A methodological review, Technologies 5 (2). [21] G. Fortino, V. Giamp`a, Ppg-based methods for non invasive and continuous blood pressure measurement: an overview and development issues in body sensor networks, in: 2010 IEEE International Workshop on Medical Measurements and Applications, IEEE, 2010, pp. 10–13.

14

ACCEPTED MANUSCRIPT

15

/ Information Fusion - - (2019) 1–15

AC

CE

PT

ED

M

AN US

CR IP T

[22] D. Buxi, J. M. Redout, M. R. Yuce, Blood pressure estimation using pulse transit time from bioimpedance and continuous wave radar, IEEE Transactions on Biomedical Engineering 64 (4) (2017) 917–927. [23] G. Fierro, F. Silveira, R. Armentano, Central blood pressure monitoring method oriented to wearable devices, Health and Technology 6 (3) (2016) 197–204. [24] C. Yang, N. Tavassolian, Pulse transit time measurement using seismocardiogram, photoplethysmogram, and acoustic recordings: Evaluation and comparison, IEEE J Biomed Health Inform. [25] A. K. Verma, R. Fazel-Rezai, A. Blaber, K. Tavakolian, Pulse transit time extraction from seismocardiogram and its relationship with pulse pressure, computing in Cardiology 42 (2015) 37–40. [26] F. Heydari, M. P. Ebrahim, T. Wu, K. Walker, K. Joe, J. Redoute, M. R. Yuce, Continuous cuffless blood pressure measurement using body sensors, in: 2018 IEEE SENSORS, 2018, pp. 1–4. [27] D. E. Mohrman, L. J. Heller, Chapter 9. Regulation of Arterial Pressure, The McGraw-Hill Companies, New York, NY, 2014. [28] E. Ravina, The Evolution of Drug Discovery: From Traditional Medicines to Modern Drugs, John Wiley and Sons., March, 2011. [29] The Finometer MIDI datasheet, non-invasive hemodynamics, Finapres Medical Systems (FMS) in combination with BeatScope Easy software. [30] W. A. Brzezinski, Blood Pressure. In H. K. Walker, W. D. Hall, J. W. Hurst editors. Clinical Methods: The History, Physical, and Laboratory Examinations. 3rd edition, Boston: Butterworths, 1990, Ch. 16. [31] NBP-24NG datasheet, Blood Pressure Monitor, NORAV medical, Available online: www.norav.com.

15