Cuffless blood pressure estimation using only a smartphone

Cuffless blood pressure (BP) measurement is an all-inclusive term for a method that aims to measure BP without using a cuff. Recent cuffless technology has made it possible to estimate BP with reasonable accuracy. However, mainstream methods require an electrocardiogram and photoplethysmogram measurements, and frequent calibration procedures using a cuff sphygmomanometer. We therefore developed a far simpler cuffless method, using only heart rate (HR) and modified normalized pulse volume (mNPV) that can be measured using a smartphone, based on the knowledge that ln BP = ln cardiac output (CO) + ln total peripheral resistance (TPR), where CO and TPR are correlated with HR and mNPV, respectively. Here, we show that mean arterial pressure (MAP), systolic BP (SBP), and diastolic BP (DBP) could be estimated using the exponential transformation of linear polynomial equation, (a × ln HR) + (b × ln mNPV) + constant, using only a smartphone, with an accuracy of R > 0.70. This implies that our cuffless method could convert a large number of smartphones or smart watches into simplified sphygmomanometers.


Results
Three data points from the iPhone 6s smartphone were treated as missing values according to the iPhysioMeter SM 's built-in algorithm detecting outliers 34 . One data point from the dedicated photoplethysmograph could not be analysed because of an intense artefact, and was removed. This unanalysable data point was overlapped with one of three data points treated as missing values in the smartphone data set.  Table 1. HR and ln mNPV values derived using the smartphone. The agreement between HR and ln mNPV measurements derived from the iPhysioMeter SM run on an iPhone 6s and a laboratory photoplethysmograph with a 16-bit A/D converter are shown in Fig. 1.
Multiple linear regression analyses. The results of multiple linear regression analyses using brachial ln MAP, ln SBP, and ln DBP as independent variables and using ln HR and ln mNPV derived using only the smartphone as dependent variables are summarized in Table 2(A). The same analyses replacing data from the smartphone with data from the dedicated photoplethysmograph are shown in Table 2(B).
Accuracy of MAP, SBP, and DBP estimates. Scatterplots of paired MAP, SBP, and DBP values estimated using only the smartphone and measured using a brachial cuff sphygmomanometer, together with their Bland-Altman plots, are shown in Fig. 2(A). The same analyses using those from the dedicated photoplethysmograph are shown in Fig. 2(B). The estimated MAP, SBP, and DBP from the smartphone and the dedicated photoplethysmograph were not correlated with their residuals, respectively (all |r|s < 0.03). Shapiro-Wilk tests did not detect a strong violation of normality of the distribution in the residuals (all ps > 0.026).

Discussion
The present findings offer a simple and moderately reliable cuffless method for estimating BP using only a smartphone, as well as a traditional finger photoplethysmograph. The estimated MAP, SBP, and DBP attained an accuracy of R > 0.70 (Fig. 2) when compared to a standard brachial cuff sphygmomanometer, across gender and across baseline and stress conditions, where pre-ejection period (PEP), one of the major error sources of the PWV-based method 12,14,16,17,39,40 , is expected to differ 41 , without relying on a complicated individual calibration procedure and on any other devices such as an ECG and dedicated PPG 11,16,21 . As predicted, judging from standardized β    Table 2, the estimated MAP depended both on HR and ln mNPV, the estimated SBP mainly depended on HR, and the DBP depended primarily on ln mNPV and secondly on HR. These findings imply that the present cuffless method could allow a huge number of smartphones all over the world 42 to be stand-alone, simplified sphygmomanometers. This could also be true for smart watches once they become capable of measuring mNPV. Considering the necessity and the benefits of frequent BP measurements in daily life around the world 43 , even such a simplified sphygmomanometer may be expected, in the not-too-distant future, to play a certain role in preventing people from developing serious cardiovascular disease. Having said that, the present method needs further development before it can be used for clinical assessments. The accuracy represented by R values > 0.70 for MAP, SBP, and DBP ( Fig. 2) were, firstly, sometimes inferior to those (|r|s > 0.74) from PWV-based cuffless BP with calibration 11,16,19 and sometimes superior to that (|r|s < 0.67) from a PWV-method where PEP is not constant 14 . However, the present method does not rely on calibration using a cuff sphygmomanometer or an ECG. Although incorporating an individual calibration process and/or PWV value into the multiple regression equation to reduce residual error would improve the estimation accuracy of our method to a certain degree, this would sacrifice its simplicity. Secondly, this accuracy is slightly superior to those (|r|s = 0.44-0.58) from PPG amplitude-based cuffless BP measurements 22,24 , and roughly equal to that (|r|s = 0.50-0.92) from PPG morphology-based cuffless BP measurements using multiple linear regression 23 , which only requires PPG. The method is the same as the present one in that both only require PPG, but the advantage of our method is that it only requires the use of a smartphone; there is no need for a data acquisition system or bio-amplifier. Thirdly, the accuracy of our method (|r|s = 0.48 and 0.59 for DBP and SBP, respectively) is marginally superior to that of a combined PWV-PPG-based method with machine learning using PPG feature extraction 21 . While this combined method does not require individual calibration, it still requires an ECG. Taken together, the present method is simple to use and relatively accurate.
All multiple correlation coefficients, R values, from a smartphone were slightly lower than those from a laboratory photoplethysmograph. In addition, all multiple regression coefficients, c, from the smartphone were higher than those from the dedicated photoplethysmograph. This might be due to the wavelength of the light used in the finger PPG; green for the smartphone, and near infrared for the laboratory photoplethysmograph. Near-infrared light has sufficient penetration depth to probe peripheral arterioles where the blood pressure is equal to that in the brachial artery 5,44 . In contrast, green light has a much shallower penetration depth compared to near-infrared light, and consequently probes more downstream arterioles near blood capillaries where the blood pressure decreases 44 . Thus, decrease in pressure is most likely linked to the higher c coefficients in the smartphone. Moreover, there is no guarantee that these decreases in BP are constant across participants or across MAP, SBP, and DBP. Thus, this variation plausibly results in lower R values.
The R values might be better by using higher order regression. Although "models with degree higher than 2 are rarely required in practice" 45 , inverse U-shape relationships are often observed between CO and HR e.g. 46 . So, we have tried testing quadratic models. However, due to quite high multicollinearity (in the current data set, r values between HR and HR-squared are over 0.99), regression analyses were not stable, such as resulted not in inverse U-shape as expected but in U-shape relationship and in non-significant β coefficients either in HR or HR-squared by cancelling out each other, in spite of virtually no increase in R values. Considering the fact that the literature sometimes failed to replicate such quadratic relationships e.g. 47 , and "likelihood surface is nearly flat near the maximum" 45 in regression models, situations where we can take advantage of U-shape curve or higher order regression might be very limited, contrary to expectations.
We removed three data points from the iPhone data according to its auto analysis algorithm, although there was no need to remove data in our previous study where a similar stress task was used 33,35 . This difference was most likely due, firstly, to the position of the finger photoplethysmograph sensors. In the previous studies, the sensors were attached on opposite sides of the finger, whereas they were attached on the same side of the neighbouring finger in the present study. Secondly, the difference may have been a result of the smartphone size. A larger smartphone was used in the present study. The placement of the sensors and the size of the device made it harder for the participants to hold the smartphone firmly. For the removed data, the root mean square errors of the delta difference in HR and ln mNPV between the laboratory photoplethysmograph and smartphone were 12.0 and 0.56, respectively. Without removing these values from the smartphone dataset, multiple regression analysis revealed that R values were decreased by approximately 0.05. Although these data indicate that the auto analysis algorithm works effectively to remove aberrant values, every effort should be paid to keep the fingertip firmly on the CMOS camera of the smartphone.
Our study has several limitations. Firstly, we used a standard brachial oscillometric sphygmomanometer as a reference, so the beat-by-beat estimated MAP, SBP, and DBP from the smartphone and dedicated photoplethysmograph were not validated. Although it is known that continuous BP monitors such as those that use the vascular unloading technique, sometimes produce inaccurate values 48 , further studies using an improved continuous monitor e.g. 5,49,50 , are needed. Secondly, our participants were limited to young Japanese men and women. Thus, further studies examining whether the multiple correlation coefficients and β coefficients differ depending on the population are needed. Thirdly, we have used resting BL and stressful MA, but have not used other situations, including other stressful tasks, changes in body posture, and exercise. Future studies incorporating these conditions are desirable, as only a few studies have investigated more than three conditions together e.g. 11,14 .
Despite these limitations, we offer a simple and moderately reliable method for estimating MAP, SBP, and DBP using only a smartphone without using a complicated calibration procedure requiring cuff sphygmomanometer measurements and any other devices, such as an ECG and dedicated PPG. The present method could potentially convert a huge number of smartphones into simplified sphygmomanometers.

Participants.
A total of 13 volunteers [six Japanese women and seven Japanese men; aged 20-24 years; body mass index = 20.4 ± 1.4 (mean ± standard deviation); at least 12 years of education; and living in Sapporo City], recruited via flyers placed around the university, participated in this study. According to the 2017 ACC/AHA classification 51 , six out of 13 participants were classified as normal, zero was elevated, four were hypertension stage 1, and three were hypertension stage 2. Because the present method is new and no direct effect size was available from previous studies, we reviewed studies examining PWV-based BP and found that a relatively small sample size was acceptable unless there was a special reason for a large sample size, such as using machine learning. For example, Payne et al. 14 used N = 12, Douniama et al. 39 used N = 14, and Patzak et al. 52 used N = 12. Based on these studies, we adopted a relatively small sample size N = 13. The criteria for inclusion in the study were being over the age of 20, having no current cardiovascular disease except for high blood pressure, and not taking any prescription medication. The participants received approximately US$20 for their participation. Written informed SCIEnTIFIC RepoRts | (2018) 8:7298 | DOI:10.1038/s41598-018-25681-5 consent was obtained from participants after we had provided them with a complete description of the study. This study was approved by the ethics committee of Hokkaido University and conducted according to the principles expressed in the Declaration of Helsinki. This is not a replicated study.
Apparatus and measurements. Smartphone. Finger PPG was measured from the left index finger using a smartphone (Apple, iPhone 6s), within which the iPhysioMeter SM app (version 2.0) was installed. iPhysioMeter SM is a software program that was designed to run on iPhone 5 s and later smartphones running with iOS 9.2 or later (Apple) and to allow a smartphone to be a reflectance mode photoplethysmograph using the built-in LED flash light and built-in CMOS camera as a light source and photodetector, respectively. The latest version of iPhys-ioMeter SM has a 60 Hz (frames per second; fps) sampling speed, which is twice as high as the previous version, and has been equipped with newly designed digital filters.
The beat-by-beat HR and ln mNPV were derived using an auto analysis function equipped with iPhysioMeter SM ab initio, and show high agreement with those derived from a dedicated photoplethysmograph. Values departing significantly from the preceding 10-s period, defined as those contributing to an increase in the SD of the period above 8.0 beats per minute (bpm) for HR or 0.25 arbitrary units (a.u.) for ln mNPV, respectively, are judged as outliers 33,34 .
Laboratory device. Finger PPG was measured using a transmittance mode photoplethysmograph with an 810-nm, near-infrared light-emitting diode (LED; Ushio Opto Semiconductors, SMC810) as a light source and a photodiode (OSRAM, BPW34FAS) as a photosensor, placed on opposite sides of the tip of the left middle finger, as previously reported 33,35 . By means of a bio-amplifier, finger PPG was recorded using an A/D converter (National Instruments, USB-6211) at a rate of 1 kHz with a resolution of 16 bits, and stored digitally in a virtual windows-based computer (Apple, MacBook Pro, Retina 15-inch Mid 2012).
The beat-by-beat HR was derived by dividing 60,000 ms by the inter-beat interval (ms) of the alternating-current (AC) component of the finger PPG. mNPV, an index of α-adrenalin-mediated sympathetic activity [30][31][32] , was calculated by dividing the AC amplitude (mV) of the finger PPG by their corresponding direct-current (DC) component (mV).
Reference. SBP and DBP were measured using a brachial cuff sphygmomanometer (NISSEI, DS-S10) attached to the right arm. MAP was calculated using the following formula: MAP = DBP + (SBP − DBP)/3.

Procedure.
The experiment was performed in a 4 × 5 m conference room maintained at a temperature of 24−26 °C. After the sensor of the finger photoplethysmograph and the brachial cuff of the sphygmomanometer were attached, the participants sat in a chair with both of their hands on a desk and holding a smartphone in their left hand while keeping their index finger on the CMOS camera. The brachial cuff and fingers were supported at heart level. Each participant was instructed to keep as still as possible until the end of the experiment in order to minimize movement artifacts.
The experiment began with a 7-min adaptation, which was followed by a 3-min baseline (BL), as shown in Fig. 3. Finger PPGs from the smartphone and the dedicated device were measured simultaneously during the BL period. Brachial BP measurements were carried out during the first and the third quarters of the BL (i.e., 0-45 s and 90-135 s, respectively). Next, the participants performed a 3-min mental arithmetic task (MA) where they were required to subtract 13 sequentially from 5,000 (thus, 4,987; 4,974; 4,961; …) as quickly and accurately as possible 35 . Our previous research 41 has shown that this MA causes a decrease in PEP. Similarly to BL, both finger PPGs were measured across MA, brachial BP measurements were made during the first and the third quarters of the MA. Data analysis. Analyses were conducted separately but in parallel for the smartphone and the dedicated photoplethysmograph.
The beat-by-beat HR and ln mNPV were averaged for each quarter period (i.e., 45 s). The averages derived from smartphone data including over 20% outliers judged by the auto analysis routine of the iPhysioMeter SM were treated as missing values. Because blood pressure was mainly measured within the first and the third quarters of BL and MA, only these averages were forwarded for further analysis. To gain a perspective of the data, physiological values from BL and MA were averaged for each period to produce single BL and MA values, and differences between conditions were compared using a paired two-sided t-test.
To examine the agreement of HR and mNPV measurements derived from the smartphone and from the dedicated photoplethysmograph, geometric mean regression analysis 53 and Bland-Altman analysis 54 were conducted.
Multiple linear regression analyses, applying brachial ln MAP, ln SBP, and ln DBP as each dependent variable and ln HR and ln mNPV as dependent variables, were conducted.