Abstract— Recent years have seen increased attention

being given to Blood Pressure (BP) monitoring. Among all

kinds of measurements, the monitors based on Pulse Transit

Time (PTT) have gain plenty of attention due to its

continuous and cuffless features. Additionally, several studies

proposed a fancy way to estimate photoplethysmography

(PPG) signal simply via a regular webcam. Nevertheless,

literatures on issues of integrating these two advanced

techniques have emerged on a slowly and scattered way.

Furthermore, accuracy of BP prediction model based on PTT

is often limited due to the lack of data. To address the

above-mentioned problems, we proposed an image based BP

measurement algorithm using k-nearest neighbor and

transfer learning results from MIMICII database to real task.

The study also introduces newly defined PTT features which

are especially suitable for image based PPG and domain

adaptation. Compared with the state-of-the-art algorithm,

root mean square error of SBP evaluation has been reduced

from 15.08 to 14.02.

I. INTRODUCTION

On the basis of the WHO statistics, hypertension and hypertensive renal disease is estimated to cause 7.5 million deaths, about 12.8% of all deaths per year globally [1]. Hypertension is also commonly regarded as a major causes of cardiovascular diseases (CVDs), which are prime cause of death in the world as well. According to Taiwan Hypertension Society, the number of people exposed to the risk of hypertension has reached up to 6.24 million. Among these patients, only one out of five measures blood pressure (BP) on a weekly basis [2]. Currently, studies pointed out that routinely self-measured BP at home is better than BP measured at hospitals in diagnosing BP-related diseases [3]. On top of that, many environmental factors (e.g. white coat effect) may lead to inaccurate results, as well as subsequent medical diagnosis. Consequently, home BP monitoring has gained increased attention from experts in order to control the progress of heart diseases and hypertension [4].

Resrach is supported by MOST 105-2221-E-009-041.

Po-Wei Huang and Chun-Hao Lin are with Institute of Electrical and

Control Engineering, National Chiao Tung University (NCTU), Hsinchu,

300 Taiwan (phone: +886-3-5712121#54428; e-mail: abc1199281@

sccp.cn.nctu.edu.tw, seanlin.eic02g@g2.nctu.edu.tw).

Bing-Fei Wu and Meng-Liang Chung are with the Department of

Electrical and Computer Engineering, NCTU (e-mail: {bwu, mlchung }

at cssp.cn.nctu.edu.tw).

Tzu-Min Lin is with Division of Rheumatology, Immunology and

Allergy, Department of Internal Medicine, Taipei Medical University

Hospital, Taipei, 110, Taiwan (e-mail: chateauxjolie@yahoo.com.tw).

For most patients, the difficulty in regular home BP measurement may result from the bulky and uncomfortable measuring process. Up to present, commonly used BP monitors (BPM) can be classified into three classes [5]: (a) invasive and continuous, (b) noninvasive and intermittent, and (c) noninvasive and continuous. First, with invasive arterial line, continuous arterial blood pressure (ABP) monitor can measure BP most accurately. Nonetheless, apart from the specific requirement for equipment, the stabbing pain of acupuncture makes this technique difficult to be accepted by patients. Second type of devices frequently utilize auscultation principles or oscillometric techniques (e.g., mercury sphygmomanometer and electronic blood pressure monitor). Although these devices are not invasive and easy to use, an inflatable cuff is necessary, which may give rise to discomfort during assessment. Additionally, devices based on these two principles can only provide intermittent measurement. Lastly, noninvasive and continuous BPMs are commonly developed using the volume clamp method [6] or pulse transit time (PTT) [7]. Volume clamp method was first developed by Penaz , a Czech physiologist, who integrated a finger cuff with built-in photoplethysmograph (PPG) sensor and a pneumatic servo system to estimate continuous BP [6]. Nevertheless, the requirement of a finger cuff may bring about discomfort and the accuracy may be controversial [8].

Recently, to achieve the noninvasive, continuous, and cuffless BP measurement, considerable attention has been paid to the research based on pulse wave velocity (PWV) or its reciprocal, PTT [9-13]. The velocity of arterial pulse propagating through the vessels depends on the arterial wall. Because the arterial wall varies with the arterial pressure, BP can be calculated from PWV or PTT. The original definition of PTT is the time taken by arterial pulse traveling from the heart to a peripheral site. Calculation of PTT conventionally requires R-peak of electrocardiogram (ECG) as the starting timestamp and maximum slope of PPG as the end point [5]. Previous studies found that PTT can as well be calculated from the time interval between two PPG signals measured at different peripheral sites [14, 15]. In spite of the noninvasive and cuffless properties of PTT, patients are required to wear finger clips, chest straps, or gel patches which might bring about skin irritation and discomfort.

To address above mentioned limitations, a novel contactless technique integrated with image-PPG (iPPG) and PTT has been proposed. The iPPG was first developed by Huelsbusch and Verkruijsse [16], who observed PPG signal simply by a regular webcam aimed at

Image Based Contactless Blood Pressure Assessment using Pulse

Transit Time

Po-Wei Huang, Chun-Hao Lin, Meng-Liang Chung, Tzu-Min Lin, Bing-Fei Wu. IEEE Fellow

mailto:seanlin.eic02g@g2.nctu.edu.tw

a human face. The method is based on the fact that blood volume pulse (BVP) during the cardiac cycle can result in the variation of optical properties on skin. Essentially, the principle employed by iPPG is the same as the one by PPG. Over the past few years, several advanced studies have been proposed to enhance the signal quality of iPPG [17-19]. On the basis of iPPG and PTT algorithms, an unobtrusive, continuous, and, most importantly, contactless BP monitor can be developed.

In 2015, the feasibility of integrating iPPG and PTT, also named as image-based PTT (iPTT), to estimate BP is first reported by Norihiro Sugita et al. [20]. In 2016, Cheol Jeong enhanced the performance using high speed camera at 420 frames per second [21]. Despite the enhanced accuracy, the requirement of high speed camera may make this technique difficult to promote. In this paper, we tried to evaluate BP via iPTT using a regular webcam with a relative lower frame rate (75fps).

With iPTT or PTT, previous studies exploit different algorithms to model the relationship between iPTT or PTT and BP. Cattivelli F. S. et al. [22] utilized pulse arrival time (PAT), heart rate (HR) and linear regression to evaluate SBP and DBP. Ruiping, Wang [23] et al, took

previous BP into consideration and then enhanced the accuracy. Nonetheless, the above-mentioned equation based methods require personal calibration before assessment. The lack of data limits the performance of model. X. R. Ding et al [24] took advantage of photoplethysmogram intensity ratio (PIR) and PTT to evaluate DBP and PP, showing enhancement upon using PTT only. Nevertheless, PIR is not available owing to varying amplitude of IPPG signals.

To achieve the goal of continuous, unobtrusive and contactless BP monitoring, the aim of this paper is therefore twofold: (a) to develop an unobtrusive and contactless BP monitor with newly designed iPTT which is suitable for iPPG. (b) to enhance accuracy of BP estimation with novel k-nearest neighbor (kNN) prediction model. Our results may help to preliminarily investigate the feasibility of transfer learning using database with abundant physiological data, MIMIC II [25]. Moreover, the iPTT method reported here could be beneficial to research attempting to evaluate BP with simply a regular webcam, improving the quality of BP monitoring and diagnosis about BP-related diseases.

Figure 1. (a) Operation principle of facial video-based BP estimation system. (b)Overall flowchart of HR/SBP/DBP estimation.

The operation principle of proposed system is shown in Figure 1a. A digital camera with 75 frames per second is focused on two regions of interest (ROIs), cheek and palm respectively. The ambient light is regarded as the light source. As described in Figure 1b, the proposed system consists of three signal processing stages, stage I – iPPG Signal Recovery, stage II – Feature point detection and HR/inter-beat interval (IBI)/iPTT Calculation, and stage III – BP estimation using Transfer Learning.

II. METHOD

A. iPPG Signal Recovery

The reason why pulse signal can be captured by a webcam aimed at skin is that a small fraction (

TABLE 1. DEFINITION OF IPTT.

Since iPPT only depends on consuming time, it will not

be influenced by DC noise like different illuminance or

distance. In addition, the extended iPTT not only enriches

the dimensions of input space but as well enable us to

maintain more information after transfer learning.

III. BP ESTIMATION BASED ON TRANSFER LEARNING

In this session, we will discuss how to estimate BP using transfer learning. First, the source domain database, MIMIC II, will be briefly introduced. Next, detail processes of domain adaptation from source to target is reported. Lastly comes to the building of kNN model.

A. MIMIC II Introduction

MIMIC II is one of the largest physiological databases in the world. The data are collected from wide variety of ICUs (surgical, neonatal, or coronary care). In the study, we exploit PPG, ABP, and ECG signals. Among the collected signals, ECG is simply utilized as timestamp for cardiovascular cycle. Because ABP, as well as PPG, varies with blood volume pulse, ABP and PPG signals are regarded as iPPG of face and palm respectively during transfer learning process.

TABLE 2. IRREGULAR RANGE OF BP

To prune the outliers in collected data, two steps are required before training: (1) to smooth each signal with moving average (125 points). (2) to remove irregular and unacceptable BP value according to Table 2. After the trimming of data, an example set of signals we acquire is illustrated as Figure 5.

Figure 5. Sample data collect from MIMIC II

B. Domain Adaptation

As illustrated in Figure 6, domain adaptation is

competent at learning from the data in which domains of

sources and targets are different with same task [27]. There

are two steps to adapt and unify the training data from

different domains.

Figure 6. Transfer learning process.

1) Unifying the BP distributions via Multi-scale Entropy

(MSE): The mismatch of BP distribution between MIMIC

II and real world is resulted from the different groups of

subjects. The proportion of data related to unhealthy

people collected in MIMIC II is much higher than that in

real world, resulting in distorted BP and PTT relationship.

In light of this concern, MSE [28] is utilized to separate

healthy and pathologic groups in the MIMIC II database

(tolerance factor r=0.2 or 0.15). As depicted in Figure 7,

we can distinguish healthy signals (left) from unhealthy

signals (right).

Figure 7. If the sampling entropy is higher than 0.5, the signal is

determined as healthy signal (left) and vise versa. The vertical cyan line

indicates the optimal judging scale factor w.r.t the signal length.

2) Eliminating sample bias and variance via Z-Score:

While using MIMIC II database, we cannot obtain where

the ABP and PPG signals were measured. Moreover, the

PTT can vary widely owing to the different positions of

source signals. Although the absolute pulse transit time in

two domains are entirely different, shorter PTT still

indicates higher BP. That is to say, the consuming time is

still meaningful as long as we eliminate the bias of

different domains. In light of this, we exploit the Z-score

on each feature to eliminate the influence of bias and

variance.

[ ]

[ ]N

iPTT E iPTTiPTT

Var iPTT

(8)

where E[] is the expectation operator, Var[] is the

variance operator, and NiPTT is the Z-score of each

iPTT feature.

C. kNN Model

Taking NiPTT and IBI as input variables, we apply

kNN to learn the regression models of SBP and DBP respectively. In this model, k is determined as three and the Euclidean distance is used.

IV. EXPERIMENTAL RESULT

A. Experiment Setup

To our best knowledge, there is no open database which contains both BP signals and videos with palm and face. In consequence, a benchmark database called Camera DB has been built according to following setup. The facial videos were captured by Sony CEJH-15007 in bitmap format with VGA resolution, 75 frame rate and 8-bit depth. The ambient light is regular fluorescent lamp around 400 Lux. As shown in the first photo of Figure 1a, subjects were asked to sit still in front of the webcam with a distance about 70 cm. On the other hand, for each set of experiment, a pair of BP/HR values and a facial video with 40 sec were recorded. The detailed timeline is illustrated in Figure 8.

Figure 8.Timeline of a set of experiment.

B. Benchmark Dataset

Following the experiment setup mentioned above, we have collected data from 13 subjects (10 male and 3 female). Each subject conducted the experiment for ten times. As a result, there are total 1300 pairs of data collected. We select ten pairs of data in each 40sec randomly. Eighty percent of data are chosen randomly as training set and the rest are testing set. Detail statistic information of subjects are illustrated on the Table 3.

TABLE 3. STATISTIC OF SUBJECTS

*S.D. = Standard Deviation, BMI = Body Mass Index

Moreover, the distribution of the collected dataset is shown in Figure 9. In this dataset, subjects with ideal BP occupy the highest rate, 78.2%. Amount of pre-high BP accounts for 18.8%. Lastly comes the high BP and low BP which respectively take up 1.5%.

Figure 9. Distribution of the collected data. Areas with different color

(purple, green yellow and red) represents different level of BP (low, ideal,

pre-high, and high) respectively.

C. Result of HR Estimation

The Bland-Altman plot of HR estimation result is shown

as Figure 10. The mean absolute error (MAE) achieved

4.35 bpm and root mean square error (RMSE) could

achieve 5.15 bpm. This result verifies the effectiveness of

HR measurement based on iPPG.

Figure 10. Bland-Altoman Plot of HR Estimation

D. Result of BP Estimation

To our best knowledge, there is only one journal article

[21] which estimates BP based on video as well. The

similarities and differences between proposed algorithm

and Jeong’s algorithm are listed in Table 4. TABLE 4. COMPARE OF TWO ALGORITHMS

Although Jeong’s algorithm can only predict SBP and

trained with Camera DB, we have extended the method

with the same transfer learning technique to learn from

MIMIC II database. Tested with Camera DB, the

performances of BP evaluation are listed in Table 5. TABLE 5. PERFORMANCE COMPARE OF BP MEASUREMENT

*unit = mmHg

Jeong’s algorithm utilizes conventional PTT only;

therefore, the performance of benchmark algorithm

declines while trained with hybrid database. Using transfer

learning with kNN, both DBP and SBP estimation achieve

the best performance. In summary, the RMSE of SBP has

been reduced from 15.08 to 14.02 with proposed

algorithm.

V. CONCLUSION

Integrated with iPPG and the concept of PTT, the paper

develops an unobtrusive and contactless BP monitor using

a webcam with relative lower frame rate (75 fps). On top

of that, newly designed iPTT is proposed for iPPG. With

iPTT, proposed method can preliminarily learn kNN

model from MIMIC II database using transfer learning.

The RMSE of SBP has been reduced to 14.02. Some

questions will be exploited in the future. Due to the

limitation of experiment, data set for testing is relative

small and remains to be built.

ACKNOWLEDGMENT

The work is supported by Grant no: MOST

105-2221-E-009-041.

REFERENCES

[1] WHO. (31, Aug). Global Health Observatory (GHO) data:

Raised blood pressure. Available:

http://www.who.int/gho/ncd/risk_factors/blood_pressure_pr

evalence_text/en/

[2] T. H. Society. Among 6.24 Million High Risk Group, Only

One out of five Measures BP weekly. Available:

https://groups.google.com/forum/print/msg/ths-announceme

nt/1u9Zd1fFHZY/_TkEv8cXBQAJ?ctz=4031061_24_24__

24_

[3] E. Dolan, A. Stanton, L. Thijs, K. Hinedi, N. Atkins, S.

McClory, et al., "Superiority of ambulatory over clinic blood

pressure measurement in predicting mortality," Hypertension,

vol. 46, pp. 156-161, 2005.

[4] Y. L. Zheng, X. R. Ding, C. C. Poon, B. P. Lo, H. Zhang, X. L.

Zhou, et al., "Unobtrusive sensing and wearable devices for

health informatics," IEEE Trans Biomed Eng, vol. 61, pp.

1538-54, May 2014.

[5] X. R. Ding, N. Zhao, G. Z. Yang, R. I. Pettigrew, B. Lo, F.

Miao, et al., "Continuous Blood Pressure Measurement From

Invasive to Unobtrusive: Celebration of 200th Birth

Anniversary of Carl Ludwig," IEEE Journal of Biomedical

and Health Informatics, vol. 20, pp. 1455-1465, 2016.

[6] J. Penaz, "Photoelectric measurement of blood pressure,

volume and flow in the finger," in Digest of the 10th

international conference on medical and biological

engineering-Dresden, 1973, 1973.

[7] P. A. Obrist, K. C. Light, J. A. McCubbin, J. Hutcheson, and

J. L. Hoffer, "Pulse transit time: Relationship to blood

pressure and myocardial performance," Psychophysiology,

vol. 16, pp. 292-301, 1979.

[8] S.-H. Kim, M. Lilot, K. S. Sidhu, J. Rinehart, Z. Yu, C.

Canales, et al., "Accuracy and Precision of Continuous

Noninvasive Arterial Pressure Monitoring Compared with

Invasive Arterial PressureA Systematic Review and

Meta-analysis," Anesthesiology: The Journal of the

American Society of Anesthesiologists, vol. 120, pp.

1080-1097, 2014.

[9] A. M. Carek and O. T. Inan, "Robust Sensing of Distal Pulse

Waveforms on a Modified Weighing Scale for Ubiquitous

Pulse Transit Time Measurement," IEEE Trans Biomed

Circuits Syst, May 23 2017.

[10] T. Seeberg, J. Orr, H. Austad, M. Roed, S. Dalgard, D.

Houghton, et al., "A novel method for continuous,

non-invasive, cuff-less measurement of blood pressure:

evaluation in patients with non-alcoholic fatty liver disease,"

IEEE Trans Biomed Eng, Sep 12 2016.

[11] F. Miao, N. Fu, Y. T. Zhang, X. R. Ding, X. Hong, Q. He, et

al., "A Novel Continuous Blood Pressure Estimation

Approach Based on Data Mining Techniques," IEEE J

Biomed Health Inform, Apr 28 2017.

[12] Q. Zhang, X. Zeng, W. Hu, and D. Zhou, "A Machine

Learning-empowered System for Long-term Motion-tolerant

Wearable Monitoring of Blood Pressure and Heart Rate with

Ear-ECG/PPG," IEEE Access, pp. 1-1, 2017.

[13] D. Buxi, J. M. Redout, and M. R. Yuce, "Blood Pressure

Estimation Using Pulse Transit Time From Bioimpedance

and Continuous Wave Radar," IEEE Trans Biomed Eng, vol.

64, pp. 917-927, Apr 2017.

[14] M. Nitzan, B. Khanokh, and Y. Slovik, "The difference in

pulse transit time to the toe and finger measured by

photoplethysmography," Physiological measurement, vol. 23,

p. 85, 2001.

[15] W.-C. Tsai, J.-Y. Chen, M.-C. Wang, H.-T. Wu, C.-K. Chi,

Y.-K. Chen, et al., "Association of risk factors with increased

pulse wave velocity detected by a novel method using

dual-channel photoplethysmography," American journal of

hypertension, vol. 18, pp. 1118-1122, 2005.

[16] W. Verkruysse, L. O. Svaasand, and J. S. Nelson, "Remote

plethysmographic imaging using ambient light," Optics

express, vol. 16, pp. 21434-21445, 2008.

[17] G. de Haan and V. Jeanne, "Robust pulse rate from

chrominance-based rPPG," IEEE Trans Biomed Eng, vol. 60,

pp. 2878-86, Oct 2013.

[18] W. Wang, A. den Brinker, S. Stuijk, and G. de Haan,

"Algorithmic Principles of Remote-PPG," IEEE Trans

Biomed Eng, Sep 13 2016.

[19] F. Litong, P. Lai-Man, X. Xuyuan, L. Yuming, and M. Ruiyi,

"Motion-Resistant Remote Imaging Photoplethysmography

Based on the Optical Properties of Skin," IEEE Transactions

on Circuits and Systems for Video Technology, vol. 25, pp.

879-891, 2015.

[20] N. Sugita, K. Obara, M. Yoshizawa, M. Abe, A. Tanaka, and

N. Homma, "Techniques for estimating blood pressure

variation using video images," in Engineering in Medicine

and Biology Society (EMBC), 2015 37th Annual

International Conference of the IEEE, 2015, pp. 4218-4221.

[21] I. C. Jeong and J. Finkelstein, "Introducing contactless blood

pressure assessment using a high speed video camera,"

Journal of medical systems, vol. 40, p. 77, 2016.

[22] F. S. Cattivelli and H. Garudadri, "Noninvasive cuffless

estimation of blood pressure from pulse arrival time and heart

rate with adaptive calibration," in Wearable and Implantable

Body Sensor Networks, 2009. BSN 2009. Sixth International

Workshop on, 2009, pp. 114-119.

[23] R. Wang, W. Jia, Z.-H. Mao, R. J. Sclabassi, and M. Sun,

"Cuff-free blood pressure estimation using pulse transit time

and heart rate," in Signal Processing (ICSP), 2014 12th

International Conference on, 2014, pp. 115-118.

[24] X. R. Ding, Y. T. Zhang, J. Liu, W. X. Dai, and H. K. Tsang,

"Continuous Cuffless Blood Pressure Estimation Using Pulse

Transit Time and Photoplethysmogram Intensity Ratio,"

IEEE Transactions on Biomedical Engineering, vol. 63, pp.

964-972, 2016.

[25] A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P.

C. Ivanov, R. G. Mark, et al., "Physiobank, physiotoolkit, and

physionet," Circulation, vol. 101, pp. e215-e220, 2000.

[26] J. Cheng, X. Chen, L. Xu, and Z. J. Wang, "Illumination

Variation-Resistant Video-Based Heart Rate Measurement

Using Joint Blind Source Separation and Ensemble Empirical

Mode Decomposition," IEEE J Biomed Health Inform, Oct

06 2016.

[27] S. J. Pan and Q. Yang, "A survey on transfer learning," IEEE

Transactions on knowledge and data engineering, vol. 22, pp.

1345-1359, 2010.

[28] M. Costa, A. L. Goldberger, and C. K. Peng, "Multiscale

entropy analysis of complex physiologic time series," Phys

Rev Lett, vol. 89, p. 068102, Aug 05 2002.

http://www.who.int/gho/ncd/risk_factors/blood_pressure_prevalence_text/en/http://www.who.int/gho/ncd/risk_factors/blood_pressure_prevalence_text/en/https://groups.google.com/forum/print/msg/ths-announcement/1u9Zd1fFHZY/_TkEv8cXBQAJ?ctz=4031061_24_24__24_https://groups.google.com/forum/print/msg/ths-announcement/1u9Zd1fFHZY/_TkEv8cXBQAJ?ctz=4031061_24_24__24_https://groups.google.com/forum/print/msg/ths-announcement/1u9Zd1fFHZY/_TkEv8cXBQAJ?ctz=4031061_24_24__24_