1. Introduction
Persons examined in a magnetic resonance imager (MRI) are exposed to noise and vibration, causing them stress that manifests mainly by heart rate (HR) and arterial blood pressure (ABP) changes [
1]. The HR changes can be detected from a photoplethysmography (PPG) signal, while the systolic or diastolic blood pressure (SBP or DBP) values are measured by a blood pressure monitor (BPM) [
2]. However, this type of a measurement arrangement is less comfortable for tested persons, and it causes problems with practical realization of experiments. It means that repeated squeezing of the arm by a pressure cuff during the whole experiment increases the negative stress and can induce a distortion of measured signals and subsequently determined parameters in the final effect. The practical difficulty of this type of measurement follows from the fact that the BPM device cannot be directly used for measurement inside the scanning area of a running MRI device due to interaction with a working magnetic field and a strong radiofrequency (RF) disturbance. The pressure cuff originally also includes a metal brace, which must be exchanged by another one from copper or brass material. Therefore, many researchers try to dispense with this type of direct blood pressure measurement approach and to use a cuffless method based on an estimation of ABP values from a PPG signal. It is well known that the second-derivative PPG wave (SD-PPG) [
3] consists of five areas corresponding to the time-domain features [
4]. These parameters can be calculated by a linear regression using the mean square method [
5,
6], nonlinear regression algorithms [
7], or least squares support vector machine algorithm [
8]. The most recent methods based on one-channel PPG signals use the pre-trained artificial neural networks for the continual estimation of the systolic blood pressure [
9]. This approach enables creating more compact wearable devices for the monitoring of sporting and other human activities or the control of different patients’ vital parameters in a hospital environment.
The precision of ABP estimation may be improved by the method based on heart pulse transmission parameters (HPTPs). Originally, the pulse transmission time (PTT) represented the time difference between the R peak of the electrocardiogram (ECG) and the systolic peak of the PPG signal. Due to a linear relationship between PTT and ABP [
10,
11,
12], the estimation of the blood pressure using the PTT is more accurate than using the PPG alone [
13,
14]. The PTT can also be determined from two or more PPG waves picked up by sensors placed at the defined distance
Dx [
7,
15]. Another parameter describing the current state of the cardiovascular system is the pulse wave velocity (PWV) [
10,
15,
16,
17]. In the current work, we also show the usefulness of the derived parameters: relative PTT (rPPT) and relative PWV (rPWV).
In general, an optical PPG sensor can work on the transmission or reflection principle. The transmission type of a sensor probe has usually the form of a finger ring or clip with a light source (one or more LED elements) and a photodetector placed on opposite sides of the sensed human tissue—this type is often used in oximeters. The reflection type of an optical sensor is placed mainly on fingers or a wrist, and it is typically fixed by an elastic/textile ribbon or integrated as a part of wearable devices—fitness bracelets, smart watches, etc. In this case, the photodetector measures the intensity of the light reflected from the skin, and it is placed on the same side of the skin surface as the light source transmitter. Multi-channel PPG signals picked up simultaneously by sensors mounted on different parts of hands or legs [
18,
19] are successfully used for this task. In continuous long-term screening and monitoring, sampling at lower frequencies may be used, e.g., 128 Hz in a wrist-based wearable device for atrial fibrillation detection [
20]. Special wearable smart devices used for the monitoring of multiple vital parameters contain sensors for the continual acquisition of ECG and PPG signals [
21,
22]. Commercial wearable PPG sensors use typical sampling frequencies between 50 and 100 Hz [
23]. The higher sampling rate (typically ≥ 500 Hz) is justified for analysis of the systolic pulse with higher accuracy, and it is also important for the precise determination of PTT and PWV parameters.
For evaluation and testing of the stability of the ABP estimation method, the PPG corpora are often used. The well-known PPG signal database MIMIC-II [
24] consists of collected multi-parameter records (PPG, ABP, RESP, and so on) taken from patients located in an Intensive Care Unit (ICU). The main problem of this database is that it includes only one-channel PPG signals, so the PTT as well as the PWV parameters cannot be determined. In addition, the PPG data are sampled at 125 Hz, and raw signals are partially corrupted. Due to manipulation and other operations performed in real conditions of the ICU, we cannot use this corpus to evaluate the performance of the currently developed ABP estimation method. Other open-access datasets, such as WESAD (Wearable Stress and Affect Detection) [
25] or PPG-BP [
26], have the same limitation of one-channel PPG signals only. Therefore, for purpose of this work, two small corpora of PPG records collected in the frame of our previous research [
27,
28] were finally used.
The main motivation of this work was to test whether these HPTPs are suitable for ABP estimation and whether they give sufficient estimation accuracy. This paper describes the procedure for HPTPs determination from the preprocessed two/three-channel SD-PPG signals. The linear regression method is used to perform ABP estimation from the HPTPs. At first, auxiliary investigations were performed with the aim of finding appropriate settings for PPG signal processing. They include an analysis of (1) the effect of different fS on the precision of the determined HPTPs; (2) the influence of the signal filtering effect on PPG wave properties; and (3) the simulation of consequences of the absolute measurement error of Dx distance on the precision of the PWV parameter. Next, the distribution analysis of the determined HPTPs depending on the type of the hand (left/right) and gender (males/females) was realized to map the mutual relationship between HPTPs and the measured SBP, DBP, and HR values. The main estimation experiment consists of testing multi-channel PPG records. The relative estimation error (REE) is evaluated to verify the estimation accuracy of SBP or DBP values. The partial results determined separately from the two used databases, types of hand, and genders of the testing subjects were compared numerically. Final estimation errors for both databases together were graphically evaluated by the scatter plots mapping correlations between the measured and estimated ABP values and using Bland–Altman plots.
4. Discussion
In practice, the sampling frequency
fS of about 150 Hz is sufficient for sensing of the PPG signals used for the determination of HR values [
31] and basic features of PPG waves (heart pulse amplitude and range, signal ripple, etc.). Our first auxiliary experiment confirmed that higher
fS values (at least 1 kHz) should be applied to obtain correct and precise HPTP values. Otherwise, the distances ∆
PSYS determined in samples can be too short (see values in the first line in
Table 1), which causes high errors in the time domain (for three-sample inaccuracy, it is up to 24 ms in the case of
fS = 125 Hz, while it is only 1.5 ms for
fS = 2 kHz). This effect has also a great influence on the accuracy stability of the determined PPT and other derived parameters.
The second auxiliary analysis demonstrates the “integration” effect of the MA filter applied on the smoothed PPG signal. The graphical form of the obtained results shows that increasing the length 2
NX +1 of the MA filter window has a direct influence on the recorded raw PPG signal: a too high value of the
NX parameter causes an undesirable effect on the amplitude
Ap and the width
Wp of systolic pulses—see
Figure 6. On the other hand, smoothing of the PPG wave by the MA filter has also some effect on the position and sharpness of the systolic pulse
PSYS, which may result in small fluctuations of the determined heart pulse period
THP and the subsequently determined distances ∆
PSYS. Therefore, to obtain a properly smoothed PPG signal without any secondary negative influence, the window duration and the sample rate are set by a compromise. The settings finally chosen for PPG signal processing in the main ABP estimation experiments were
NX = 32 and
fS = 1 kHz. In accordance with the results of the previously performed analysis of a limited precision of manually measured distances
Dx [
27], this effect on the inaccuracy of PWV and rPWV values can be omitted for the purpose of this work.
Prior to practical experiments with the PPG signal corpora, the distribution analysis of determined HPTPs in dependence on the finger (P1–P5), the type of the hand (left/right), and the gender (males/females) was realized. The graphical results in
Figure 8 are divided into two set of graphs, separately for male and female tested persons. In each of the graph pairs (
Figure 8b–d), there are visible differences between HPTPs determined from the PPG signals sensed on the left and the right hand. There are also detectable smaller differences between male and female subjects, which are primarily caused by the mentioned higher length of fingers (see the first graph pair in
Figure 8a). The next comparison analysis—the mapping of mutual positions of HPTPs with ABP and HR values—also shows a data grouping effect depending on the position of PPG signal sensing and type of tested subjects (male/female)—see two set graphs for SBP and DBP parameters in
Figure 9. For this reason, the results of the main estimation experiments should be selected and analyzed depending on the gender and the type of the hand. The last preliminary analysis was aimed at a comparison of the correlation between the HR values measured by the BPM device and those determined from the PPG signal—see
Figure 10. The distribution of HR values separately for left/right hands and summary results for both hands as well as the corresponding correlation coefficient
R show minimal differences. With respect to the determined minimum areas of 95% confidence bounds, the best result (maximum
R coefficient) was obtained for HR values from the left hands—compare plots in
Figure 9b–d. Based on this previous analysis of the range of SBP or DBP values determined from the PPG signal database, the minimum/maximum limits were set as follows: SBP
MIN = 80 mmHg, SBP
MAX = 150 mmHg; DBP
MIN = 50 mmHg, DBP
MAX = 105 mmHg. On the other hand, the conversion characteristic used for ABP values estimation is created by a linear regression approach (where the conversion line is defined by an origin in Cartesian coordinates and a direction); it can be also extrapolated for the testing of PPG signals with ABP values lying outside these chosen intervals. However, in this case, we may expect a significantly increased estimation error of SABP/DABP values.
From the point of view of the obtained ABP estimation results, the numerical comparison in
Table 2 shows negative
REE (ABP
EST < BP
BPM) and low standard deviation (less than 3%) for DB1. For DB2, REE is positive, with a higher standard deviation (more than 10%). In both databases, the estimation errors are higher for DBP values. In both cases, the HPTPs were taken from the joined left and right hands from all persons involved in each of the analyzed databases. Detailed results of the correlation between the measured and estimated SBP and DBP values depending on the used database (DB1 or/and DB2), the type of the hand, and the gender of the testing subject are enumerated in
Table 3. It shows a better correlation for HPTPs from the joined left and right hands and all tested subjects: the estimated SBP values have always lower
R values than the DBP ones, which is probably due to the larger range of pressure values. The higher correlation between the measured and estimated BPs for DB1 than for DB2 in
Table 3 is in correspondence with the lower SD of REE for DB1 (compare
Table 2).
Figure 10 and
Figure 11 represent the final results for data of all tested persons from both databases together. The scatter plot graphs confirm the partial analysis results—showing higher variance of the estimation error in the case of SBP values (the summary correlation coefficient
R is lower than for DBP values). Also, the Bland–Altman plots show lower standard deviation values for the estimated DBP values, but the situation with mean values is opposite: SBP values were finally estimated with the lower MAE. For both parameters, the mean ∆ SBP/DBP values have a negative character. It is in correspondence with the observed trend of REE values for the database DB1. The DB1 contains more PPG records, so its influence outweighed the DB2 having positive REE values (compare the numerical results in
Table 2). In summary, the MAE values obtained from our current work are principally comparable with the results of other cuffless BP estimation systems using PPT and/or PWV parameters as documented the
Table 6.
5. Conclusions
The performed experiments confirm the practical functionality of the proposed method for the estimation of SBP and DBP based on heart pulse transmission parameters determined from multi-channel SD-PPG waves. The final mean estimation errors achieved using the merged DB1 and DB2 databases were MAE
SBP = −2.6 ± 10.7, MAE
DBP = −3.5 ± 9.7 mmHg. Therefore, the MAE results practically fulfill the AAMI recommendation, but in the case of the required standard deviation, the limit of the maximum equal to 8 mmHg was exceeded. According to the BHS standard, the developed estimation algorithm works with an absolute error Δ
ABP lower than 5 mmHg in 65.1/47.8 percentage cases of SBP/DBP values, 93.3/100% were predicted with Δ
ABPs lower than 10 mmHg, and finally 100% were predicted with Δ
ABPs lower than 15 mmHg, as documented in
Table 5. Generally, it holds that the evaluated algorithm is considered as acceptable if the absolute difference ∆
ABP (difference between the estimated and the real ABP value) is less than 10 mmHg [
30]. Hence, these results can be regarded as promising for this first-step experiment as well as in comparison with other cuffless BP estimation systems in
Table 6. The second merit of this work is related with our long-term research aim. The currently performed experiments confirmed that multi-channel PPG signals sensed by special prototypes of optical sensors in the low magnetic field environment with a high RF disturbance present inside the scanning area of a running MRI device are fully usable for this purpose and can bring an applicable substitution of direct ABP measurement by a blood pressure device.
However, further improvements are necessary before practical usage of the proposed estimation method. We will try to eliminate the limited precision of manually measured distances
Dx with the aim of minimizing the possible effect on the inaccuracy of PWV and rPWV values. It could be achieved by using another (semi-automatic) approach based on the determination of distances from images. A similar technique is used for the tracing of human movement when the kinetic sensors placed on different parts of a body (arm or leg) are covered by small reflexing targets to be easily sensed and traced by cameras [
32]. Next, it is well known that the quality of the sensed PPG signals depends essentially on the actual state of the skin at the place of an optical sensor. It means the color of the skin surface together with the temperature, humidity, and the pressure effect of the worn sensor influence the features of the sensed PPG signals. Therefore, in the near future, we plan to perform more measurement experiments with the aim of collecting another multi-channel PPG signal database including also the attached temperature, humidity and pressure values measured by a contact method.
Finally, the basic limitation of the current work consists of a relatively small number of processed PPG signal records (taken from the databases DB1 and DB2). Next, the processed ranges of SBP and DBP values were relatively close: SBP ∊ <94~154> mmHg, and DBP ∊ <57~92> mmHg. These close ranges have a positive effect on the robustness (stability) and accuracy of the estimation procedure when the ABP values of the tested PPG signal lie in these intervals. Otherwise, if the SBP or DBP values of the tested PPG signals are outside these ranges, the estimation procedure also works, but it produces a higher estimation error. To eliminate this negative effect on the accuracy of the whole estimation process, we must collect one larger database accompanying the measured SBP and DBP values in a wider range. It is also necessary for the better practical applicability of this developed method to a broader population with more variable ABP values. Another way to solve this problem is to establish some cooperation with the nearest medical centers in Bratislava (Slovakia), Vienna (Austria), or Brno (Czech Republic).