A new estimate technology of non-invasive continuous blood pressure measurement based on electrocardiograph

Abstract

Various physiological parameters have been widely used in the prevention and detection of diseases. In particular, the occurrence of cardiovascular diseases can be observed through daily measurement of blood pressure. Currently, the most common blood pressure measurement method records blood pressure on the upper arm. This can lead to the subject feeling uncomfortable and tension in the arm from the stress may lead to measurement errors. An electrocardiogram represents the electrical activity during heart function, but also contains blood pressure–related information. This study is an attempt to extract features related to blood pressure from the electrocardiogram signal using a new non-invasive blood pressure measurement technology that utilizes intelligent neural network algorithms to calculate blood pressure values from electrocardiogram parameters. In this study, the average error rate of the blood pressure measurement was lower than 5% compared to the common blood pressure machine. The proposed approach alleviates the errors caused by discomfort, which provides a more feasible method to continuously monitor blood pressure in less stressful conditions. This technology has significant potential for advancing healthcare.

Keywords

Electrocardiogram blood pressure neural network healthcare

Introduction

Various physiological parameters have been widely used in the prevention and detection of diseases. In particular, it is possible to observe the occurrence of cardiovascular diseases through daily blood pressure (BP) measurements. Currently, the most common BP measurement method records pressure in the upper arm. The subject might feel uncomfortable and the tension due to the pressure on the arm may lead to BP measurement errors. Previous research used a photoplethysmogram (PPG) and machine learning techniques to simultaneously and non-invasively estimate blood glucose level (BGL) and systolic blood pressure (SBP) and diastolic blood pressure (DBP).¹ Time–frequency (T–F) analysis is often used to study non-stationary cardiovascular oscillations, such as heart rate (HR) and variable BP in dynamic situations. One study previously used the T–F recursive autoregressive technique to investigate variability in pulse transit time (PTT), which is a cardiovascular parameter of emerging interest due to its potential to estimate BP non-invasively, continuously, and without a cuff. Recent studies suggest that PTT is not only related to SBP but also to HR.² An electrocardiogram (ECG) represents the electrical activity of heart function and also contains PPG- and BP-related information. The higher fidelity and more consistent ECG signal assists in identifying true cardiac pulses in an oscillometric signal that may be contaminated with noise and artifacts, thus lending greater confidence to the oscillometric BP estimation.³ For example, the use of an ECG signal was proposed for reconstructing an oscillometric signal contaminated with artifacts to more accurately assess BP.⁴ Similarly, synchronized ECG signals were employed for removing motion artifacts from oscillometric signals to increase the accuracy of BP measurements.⁵ The doctor can assess cardiovascular disease patients using parameters of the ECG signal, such as HR and heart rate variability (HRV), to avoid incidents of acute myocardial infarction or myocardial ischemia. Several methods for the estimation of BP based on the PPG signal have been proposed. They can be categorized into pure PPG-based methods^6,7 or hybrid PPG- and ECG-based methods.^8–11

In this study, we propose the use of a physiological signal to resolve a variety of physiological parameters to simplify the measurement process and time. This study extracts the features related to BP from the ECG signal and develops a new non-invasive BP measurement technology that utilizes intelligent neural network algorithms to estimate BP values from ECG parameters.

Methods and implementation

Design of the bio-signal measurement device

For this study, we created a bio-signal measurement device (Figure 1) that can measure the PPG and ECG simultaneously with a sampling rate of 100 Hz and synchronously saving the raw data to a personal computer (PC).

Figure 1.

The bio-signal measurement device.

Online peak detection of PPG and ECG

So and Chan’s¹² algorithm was used to find the locations of QRS complexes in ECG sequences and the peaks and valleys in PPG sequences. The fuzzy threshold algorithm adjusted the slope threshold in real time (Figure 2(a)). When the QRS complexes were detected in ECG sequences, we could detect the other peaks (P, Q, S, and T) of the ECG through time characteristics (Figure 3) between each wave and the R wave in ECG sequences. The peak finding of PPG was the same. The fuzzy peak detection algorithm is described below.

Figure 2.

Fuzzy threshold algorithm: (a) the structure of the fuzzy threshold algorithm, (b) the membership function of the fuzzifier, and (c) the membership function of the defuzzifier.

Figure 3.

The wave segment parametric of a normal ECG.

Let X(k) represent the amplitude of the ECG data at discrete time k. The slope of the ECG wave can be obtained by equation (1)

slope (k) = - 2 X (k - 2) - X (k - 1) + X (k + l) + 2 X (k + 2)

(1)

slope__threshold = 1 / 2 \times \max p_{k}

(2)

The initial maxp is the maximum slope within the first 200 data points in the ECG file.

When two consecutive ECG data satisfy the condition that slope(k) is large than slope_threshold which is defined as equation (2), the onset of the QRS complex is detected. After the detection of the onset of the QRS complex, the maximum point (maxp) is searched and is taken as the R point. The maxp is then updated by the fuzzy threshold algorithm which is defined as equations (3) and (4)

\max p_{k} = fuzzy (peak - \max p_{k - 1}) + \max {p_{k}}_{- 1}

(3)

peak = height of R point - height of QRS onset

(4)

In the fuzzy threshold algorithm, the prediction error e_k is an input to the fuzzy algorithm created by the difference between peak and maxp_k₋₁ that can be represented as equation (5)

e_{k} = peak - \max {p_{k}}_{- 1}

(5)

In the fuzzy algorithm, a linguistic fuzzy inference rule is utilized to calculate the modified error $e'_{k}$ . The five linguistic parameters are as follows: “LN, negative large,”“SN, negative small,”“ZE, zero,”“SP, positive small,” and “LP, positive large” (Figure 2(b)). The example of a linguistic fuzzy inference rule describes as “If the e_k is LN then the $e'_{k}$ is LN.” In this study, the fuzzy inference rule table is Table 1.

Table 1.

Fuzzy inference rule table.

e_k	LN	SN	ZE	SP	LP
$e'_{k}$	LN	SN	ZE	SP	LP

LN: negative large; SN: negative small; ZE: zero; SP: positive small; LP: positive large.

The output variable $(e'_{k})$ of the fuzzy threshold is calculated by the center of gravity method as equation (6), where the variable S_i is the membership grade of the ith premise in the inference rule and B_i is the center value of the ith conclusion in the inference rule. The range of the defuzzifier is from B₁ to B₅ (Figure 2(c))

e'_{k} = \frac{\sum_{i = 1}^{n} S_{i} B_{i}}{\sum_{i = 1}^{n} S_{i}}

(6)

Analysis of the relationship between PPG and ECG

In order to transform the signal of the ECG into a value for BP, we need to calculate the correlation of every wave segment (Figure 4) with the arterial systolic and arterial diastolic to find the transformation parameters. In Figure 4, “TR” indicates the period between the T wave and the next cycle of the R wave, the same as TQ and TP. “ST” indicates the period between the S wave and the T wave in the same cycle, the same as RT and QT. “SB” indicates the period of systole and “DB” indicates the period of diastole. “ET” represents the ECG cycle, that is, the period of the R-R interval and “BT” indicates the PPG cycle.

Figure 4.

The relationship between ECG and PPG.

Non-invasive BP measurement technology

In this study, we developed an intelligent neural network algorithm (Figure 5) that transforms ECG signals into BP values. The algorithm uses the strongly correlated between ECG and PPG as input to the neural network and, through a large number of learning trials, enhances the correct of output. The neural network algorithm includes a multilayer perceptron (MLP) and error back propagation (EBP).

Figure 5.

The intelligent neural network structure.

In the MLP processing, we set the learning rate η (1) and a tolerable error (0.001), the input nodes of input layer are i_j where j is from 1 to n, and n is the number of the high correlation coefficient (>0.7). The w_hi is the ith weight value of the hidden layer, the w_oh is the hth weight value of the output layer, and these initial weight values are random. The activity function (f) uses the log-sigmoid transfer function that is defined as equation (7), the y_o₁ and y_o₂ are the estimation values of SBP and DBP, respectively

f (x) = 1 / \exp (- 0.5 x)

(7)

The EBP is processing between two layers, the output of the n layer and jth neuron $(y_{j}^{n})$ can be obtained by equation (8) where f is the activity function, ${net}_{j}^{n}$ is the weight accumulated value of n − 1 layer. The error function (E) is defined as equation (9), where d_k is the kth target value and y_k is the output of the kth neuron in the output layer. To adjust the weight value by equations (10) and (11), where p is the training times of the pth

y_{j}^{n} = f ({net}_{j}^{n}), {net}_{j}^{n} = \sum_{i} w_{ji}^{n} y_{i}^{n - 1} + b_{j}^{n}

(8)

E = \frac{\sum_{k} {(d_{k} - y_{k})}^{2}}{2}

(9)

w_{ji} (p) = w_{ji} (p - 1) + Δ w_{ji}

(10)

Δ w_{ji} = - η \frac{\partial E}{\partial w_{ji}} = - η (d_{j} - y_{j}^{n}) f' ({net}_{j}^{n}) y_{i}^{n - 1}

(11)

Analysis of graphical user interface

This study used MATLAB’s graphical user interface (GUI) to establish an analysis platform for the ECG signal (Figure 6). The functions of this platform will show the detection of each peak, ECG rhythm analysis, and estimated value of BP.

Figure 6.

The analysis platform for the ECG signal.

Result

Measurement experiments

In this study, we input ECG, PPG, and BP measurements (Figure 7). The signal processing steps involved signal peak detection and correlation analysis and determined the best parametric relationship between these signals.

Figure 7.

Experimental measurements using (a) the PPG measurement device and (b) the ECG measurement device.

The peaks detected in the ECG and PPG by the fuzzy threshold algorithm are shown in Figure 8. The dotted line is the threshold of peak detection in Figure 8(b) and (d). Figure 8(c) displays the peaks from the PPG, with a threshold value of detection of a peak by 50% of the maximum amplitude of the PPG.

Figure 8.

Peak detection in ECGs and PPGs comparing raw and fuzzy data: (a) ECG raw data, (b) fuzzy ECG’s R wave peak detection, (c) PPG peak detection, and (d) Fuzzy PPG’s peak detection.

Analysis of ECG and PPG characteristics

Initially, we collected ECG and PPG data from five participants to find the characteristic between ECG and PPG. After signal processing, the characterizing correlation analysis between ECG and PPG signals is shown in Table 2. We selected from Table 2 the relevant characteristics higher than 68.2% (one standard deviation) of the interval parameters (ET, RT, QT, ST, TP, TQ, and TR) for the neural network input variables.

Table 2.

The characterizing correlation analysis between ECG and PPG signals.

PPG	BT	Bh	SB	SB	SB	DB	DB	DB
ECG	ET	Rh	RT	QT	ST	TP	TQ	TR
Subject 1	0.98	−0.2	0.8	0.88	0.8	0.94	0.96	0.96
Subject 2	0.97	0.26	0.68	0.75	0.74	0.95	0.96	0.96
Subject 3	0.92	−0.04	0.81	0.79	0.83	0.86	0.85	0.86
Subject 4	0.98	−0.1	0.67	0.7	0.56	0.82	0.91	0.9
Subject 5	0.99	−0.04	0.81	0.74	0.76	0.84	0.85	0.85
Average	0.97	−0.02	0.75	0.77	0.74	0.88	0.91	0.91

PPG: photoplethysmogram; ECG: electrocardiogram.

Estimated value of BP

Using the ET, RT, QT, ST, TP, TQ, and TR time segments as an input signal to the neural network, this study established two back-propagation neural network architectures with 4 input nodes, 40 hidden nodes, and 1 output node. One block estimated the SBP (ET, QT, RT, and ST as input), and the other estimated DBP (ET, TP, TQ, and TR as input). The 50 participants in this study were 20–26 years old, and we collected two datasets per subject for a total of 100 datasets. Of these, 50 datasets were used for training and another 50 datasets for testing. After learning 50 datasets (Figure 9), the subjects direct 1-min ECG measurements, and calculated BP values, we compared the estimated value of BP (ECG-BP, PTT-BP) with the value of the mercury sphygmomanometer (KENLU model K-300), as shown in Appendix 1. The average error rate of the estimated SBP was 1.96% and the average error rate of the DBP was 2.14% in ECG-BP model. The average error rate of the estimated SBP was 6.23% and the average error rate of the DBP was 6.23% in PTT-BP model.

Figure 9.

The training performance of the neural network.

Conclusion

In this study, the average error rate of the BP measurement was lower than 5% (1.96%/2.14% for SBP/DBP) compared with the KENLU model K-300 mercury sphygmomanometer. BP can be affected by extrinsic environmental and behavioral factors, as well as intrinsic cardiovascular regulatory mechanisms. The back-propagation neural network utilizes a supervised learning algorithm that can trace variable HR with user behavioral change and adapt to estimate the BP value in real time. In summary, this study can be used to observe the dynamic changes in BP, and the proposed approach alleviates the errors caused by discomfort to provide a feasible method to continuously monitor BP under less stressful conditions. This study measured the ECG bio-signal to estimate the BP value using a signal measurement device that reduces interference from high frequency noise and motion aliasing, rather than simultaneously measuring the ECG and PPG. We can observe that the ECG-BP model proposed in this article is more accurate than the PTT-BP model¹¹ in Appendix 1. This technology has significant potential to advance healthcare.

Footnotes

Appendix

Appendix 1

The estimated values of blood pressure using the intelligent neural network algorithm.

Subjects	Mercury manometer		ECG-BP		PTT-BP
Subjects	SBP	DBP	SBP	DBP	SBP	DBP
01	118	72	123.29	74.80	108.54	67.65
02	138	88	130.77	86.38	128.34	80.73
03	138	86	136.57	86.51	134.34	83.73
04	130	76	134.76	76.53	134.34	83.73
05	108	72	105.80	72.27	118.24	70.64
06	112	70	114.01	70.89	113.90	78.80
07	128	84	128.43	84.70	125.70	85.60
08	130	90	128.02	90.36	140.80	87.76
09	118	76	116.70	76.57	120.43	75.57
10	120	80	118.32	80.81	124.16	81.20
11	130	90	124.50	90.24	126.52	89.94
12	138	98	137.71	98.80	136.29	98.80
13	108	68	106.08	68.34	105.92	72.72
14	114	72	122.53	72.10	125.53	69.10
15	126	78	123.96	78.06	128.09	75.06
16	108	72	106.72	73.36	136.90	85.30
17	106	64	108.40	67.27	121.82	63.58
18	122	72	129.11	78.27	129.57	66.20
19	138	86	134.92	84.93	130.02	70.35
20	118	78	119.61	81.09	126.37	69.12
21	122	80	125.51	80.75	127.82	69.61
22	108	68	120.31	68.84	125.66	64.88
23	116	78	117.33	78.11	124.85	68.61
24	128	88	124.58	83.83	127.92	82.64
25	118	84	116.93	84.44	123.46	68.13
26	130	88	131.65	89.83	129.36	70.13
27	108	68	109.26	69.68	127.80	69.60
28	128	70	128.11	70.58	123.98	68.31
29	130	96	129.85	96.65	130.13	90.39
30	102	66	101.59	66.22	109.86	66.30
31	110	68	109.59	68.11	120.12	66.38
32	110	70	109.94	42.57	120.05	66.36
33	114	64	112.18	62.97	121.24	65.75
34	118	70	117.37	71.36	127.76	65.59
35	128	74	128.28	75.62	130.11	70.38
36	126	76	126.23	76.20	128.28	69.76
37	124	68	124.56	69.91	128.33	69.78
38	108	66	107.85	66.70	106.38	65.12
39	128	80	125.02	80.88	127.49	77.49
40	140	82	129.01	82.52	127.00	79.33
41	120	86	118.09	86.33	127.01	89.33
42	138	78	137.45	79.38	126.06	69.01
43	120	86	120.92	86.44	130.13	70.39
44	136	86	136.21	86.79	130.04	78.36
45	118	86	117.38	85.96	129.73	80.25
46	86	54	85.64	54.06	104.40	58.45
47	106	70	107.00	69.68	110.14	70.39
48	118	84	124.29	82.89	110.04	80.36
49	128	68	128.20	68.56	129.89	66.31
50	130	66	131.67	66.99	129.89	66.30

ECG: electrocardiogram; BP: blood pressure; PTT: pulse transit time; SBP: systolic blood pressure; DBP: diastolic blood pressure.

Academic Editor: Stephen D Prior

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors gratefully acknowledge the support provided to this study by the Ministry of Science and Technology, Taiwan, Republic of China, under Contract MOST 104-2627-E-006-001.

References

Enric

MM.

Non-invasive estimate of blood glucose and blood pressure from a photoplethysmograph by means of machine learning techniques. Artif Intell Med2011; 53: 127–138.

Liu

Poon

CCY

Zhang

YT.

Time–frequency analysis of variabilities of heart rate, systolic blood pressure and pulse transit time before and after exercise using the recursive autoregressive model. Biomed Signal Proces2011; 6: 364–369.

Ahmad

Chen

Soueidan

. Electrocardiogram-assisted blood pressure estimation. IEEE T Biomed Eng2012; 59: 608–618.

Walloch

RA.

Automatic blood pressure monitor employing artifact rejection method and apparatus. US Patent 5,337,750, 1994.

Non invasive blood pressure measurement and motion artifact: a comparative study. Beaverton, OR: Revision Labs Inc., 1998.

Teng

Zhang

. Continuous and noninvasive estimation of arterial blood pressure using a photoplethysmographic approach. In: Proceedings of the 25th annual international conference of the IEEE engineering in medicine and biology society, Cancun, Mexico, 17–21 September 2003, pp.3153–3156. New York: IEEE.

Csordás

Mersich

Jobbágy

Á.

Advanced indirect method for measuring blood pressure. Period Polytech Elec2009; 53: 115–121.

Ilango

Sridhar

A non-invasive blood pressure measurement using android smart phones. IOSR J Dent Med Sci2014; 13: 28–31.

Fung

Dumont

Ries

. Continuous noninvasive blood pressure measurement by pulse transit time. In: 26th annual international conference of the IEEE engineering in medicine and biology society IEMBS ’04, San Francisco, CA, 1–5 September 2004. New York: IEEE.

10.

Poon

CCY

Zhang

. Cuff-less and noninvasive measurements of arterial blood pressure by pulse transit time. In: 27th annual international conference engineering in medicine and biology society (IEEE-EMBS), Shanghai, China, 17–18 January 2005. New York: IEEE.

11.

Kim

Jung

. Estimation of systolic and diastolic pressure using the pulse transit time. Proc World Acad Sci Eng Tech2010; 43: 726–731.

12.

Chan

. Development of QRS detection method for real-time ambulatory cardiac monitor. In: Proceedings of the 19th annual international conference of the IEEE engineering in medicine and biology society, Chicago, IL, 30 October–2 November 1997, vol. 1, pp.289–292. New York: IEEE.