Exploiting moving slope features of PPG derivatives for estimation of mean arterial pressure

Abstract

Monitoring Mean Arterial Pressure (MAP) helps calculate the arteries’ flow, resistance, and pressure. It allows doctors to check how well the blood flows through our body and reaches all major organs. Photoplethysmogram technology is gaining momentum and popularity in smart wearable devices to monitor cuff-less blood pressure (BP). However, the performance reliability of the existing PPG-based BP estimation devices is still poor. Inaccuracy in estimating systolic and diastolic blood pressure leads to an overall imprecision in resultant MAP values. Hence, there is a need for robust and reliable MAP estimation algorithms. This work exploits the moving slope features of PPG contour in its first and second derivatives that directly correlate with MAP and does not require estimating systolic and diastolic blood pressure values. The proposed approach is evaluated using two different data sets (i.e., MIMIC-I and MIMIC-II) to demonstrate the robustness and reliability of the work for personalized non-invasive BP monitoring devices to estimate MAP directly. A mean absolute error of $1.28\mathrm{mmHg}$ and a standard deviation of $2.50\mathrm{mmHg}$ is obtained with MIMIC-II data-set using GridSearchCV random forest regressor that outperformed most of the existing related works.

Introduction

Mean Arterial Pressure (MAP) is a measure that describes blood pressure measurement in a person’s blood vessels during a single cardiac cycle. It is considered a better biomarker for perfusion than systolic blood pressure. It is essential to have a MAP of at least 60 mmHg to supply enough blood to the arteries, kidneys, and brain. The average MAP range is between 70 and 100 mmHg, and deviation from this range for a more extended period can cause severe pathophysiological consequences [1].

A MAP calculation is necessary in a variety of situations, including the estimation of peripheral vascular resistance, fractional pulse pressure, the calibration of devices that estimate central blood pressure (BP), and others, in addition to the pathophysiological and therapeutic implications of MAP [2, 3]. Accurate MAP value estimation is crucial in all clinical applications. The calculation of the area under the BP waveform, as determined by the time-averaged BP readings over the cardiac cycle, which is captured invasively by catheter-manometer devices, is the gold-standard approach for measuring the actual value of MAP. However, in standard clinical practice, MAP is calculated non-invasively using mathematical formulas or the internal, proprietary algorithms of automated oscillometric sphygmomanometers. In addition to the fact that many commercial devices do not display MAP but instead use it to determine systolic and diastolic blood pressure using proprietary algorithms, the accuracy of MAP estimate using oscillometric devices is infrequently documented in the literature. Several formulae have been proposed for the computation of MAP based on the idea of “form factor,.” Wezler and Bögerr [4] proposed the equation $MAP=0.42SBP+0.58DBP$ in 1939, where SBP stands for systolic and DBP for diastolic blood pressure. A different formulation of the formula mentioned above was put forth by Meaney et al. [5] as follows: $MAP=DBP+0.412PP$, where PP is the pulse pressure. The Gauer formula for MAP computation, which reads: $MAP=DBP+0.33PP$, is the most popular and extensively applied one [6]. Chemla et al. [7] presented $MAP=DBP+0.33PP+5mmHg$ as an improvement to the conventional formula in 1999. The heart rate was incorporated in the formula for calculating MAP by Razminia et al. [8] in 2004 as follows: $MAP=DBP+[0.33+(0.0012HR)]\times PP$. All these formulae for MAP are derived based on estimated SBP and DBP values applicable for invasive and non-invasive approaches for measuring BP in ambulatory settings.

Pressure is obtained during the pumping of the blood towards the left ventricle into the aorta and large arteries. The MAP is calculated using cardiac output (CO), systemic vascular resistance (SVR), and central venous pressure (CVP) analogous to the relationship between flow, pressure, and resistance [9] as:

$$\begin{array}{c}\hfill MAP=(CO\times SVR)+CVP\end{array}$$ 1

Figure 1 clarify this relationship by representing the flow of blood in the circulatory path and the resultant mean arterial pressure and other components like CO and CVP. The effect of CVP is negligible under average circulatory condition and hence can be ignored from Eq. 1. Different work has been done in this domain for non-invasive and cuff-less blood pressure estimation in terms of systolic blood pressure (SBP), diastolic blood pressure (DBP) and MAP using PPG signal (also with a combination of other physiological signal like ECG). The traces of PPG shows the volumetric changes of blood during circulation and are hence valid for measurement of blood pressure and assessment of hypertension [10]. The usage of PPG’s signal derivatives was created to make accurate recognition of the PPG’s points of interest and to enable the original PPG waveform easier to interpret. The earlier works were the calibration-free estimation of BP values by a significant feature called pulse transit time(PTT) whose derivation requires simultaneous measurement of ECG and PPG [11]. Further, the accuracies and performance of the estimation was accelerated by other reported works with feature like pulse intensity ratio [12] and also with additional temporal features with deep learning models like LSTM [13] and ANN-LSTM [14]. The requirement of simultaneous use of both physiological signals (PPG and ECG) was the weighty limitation since their continuous synchronization was necessary to keep PTT feature accurate at all times. The problem was addressed by using a single PPG sensor only for the estimation of BP. Morphological features from PPG [15] and temporal features from PPG derivatives [16, 17] demonstrated remarkable results with a good correlation with mean arterial pressure. Conversion of short duration PPG into visibility graph and estimation with transfer learning is one of the recent work [18]. All the aforementioned works rely on obtaining the target BP value from ground truth arterial blood pressure signal and map the PPG signal features in terms of detected key points, derived indices or 2-D images from PPG signal in both time and frequency domain. However, the MAP can be calculated using the estimated SBP and DBP values. Still, an erroneous SBP and DBP estimation yield an overall poor estimated MAP value. None of the associated works has any of the concrete features to directly estimate MAP with significant biological correlation. The usefulness of MAP lies in the fact that we can quickly conclude a clinical scenario with provided MAP values without needing SBP and DBP measurement for the subject diagnosed with hypertension or cerebral anomalies. For a normal patient, the value of MAP lies between 70-100 mmHg. A value below 60mmHg indicates that significant organs of our body are not getting sufficient blood, leading to malfunctioning of the vital organs. MAP greater than 100 mmHg reflects excessive pressure in the arteries, causing heart muscle damage and yielding blood clot. Several patients with sepsis and vasopressors are evaluated based on the values of MAP [19]. Hypotension, or chronically low blood pressure, is at the other extreme of the range and can also be fatal. When the MAP maintenance is insufficient, essential organs do not receive the necessary blood flow, which results in hypotensive shock and rapid organ failure. Intense bacteremia or hypovolemia frequently led to hypotension [20]. As a result, we can see how mean arterial pressure may be used to detect both hypertensive and hypotensive states and to provide doctors the diagnostic data they need to choose the best course of treatment.

Fig. 1.

a Simplified blood circulation model showing the path of mean arterial pressure with components cardiac output (CO), systemic vascular resistance (SVR), and central venous pressure (CVP) b Position of mean arterial pressure in a sample of arterial blood pressure waveform

This work intends to develop an automated framework for directly estimating mean arterial pressure using raw PPG. The proposed framework trains a regression model with the moving slope features of PPG from its first and second derivative, which directly relate to the cardiac output and can directly weigh up the estimation of target MAP with apparent biological correlation. The training is done using the database’s ground truth arterial blood pressure (ABP) signal. The novelty of the work lies in deriving a single feature from both VPPG ($1\mathrm{st}$ derivative) and APPG ($2\mathrm{nd}$ derivative) of PPG that are together capable of estimating MAP with adequate accuracy, unlike the existing approaches that had used a large number of temporal and frequency-domain PPG features. In this way, indicating blood pressure as a single MAP value might be helpful for inclusion in PPG-based continuous health tracker devices and diagnosis of hypertension and CVD patients.

Data source

The popular publicly available database called Multi-parameter Intelligent Monitoring in Intensive Care (MIMIC-I) has been used in this work for evaluation. This database contains the simultaneous recordings of physiological signals, i.e., PPG, ECG, and ABP data of 39 patients [21]. Further, we used another database as a subset of the Multi-parameter Intelligent Monitoring in Intensive Care (MIMIC)-II hosted by the University of California Irvine (UCI) Machine Learning Repository [21]. This database consists of 12000 signal parts of recorded physiological signals from about 1000 subjects and contains already clean and preprocessed signals meant to develop the algorithm for cuff-less blood pressure estimation. Both these databases have a sampling frequency of 125Hz for all signals. We discarded the records having discontinuities and having abrupt cycle duration.

Methodology

The complete framework of the proposed work is summarized in the block diagram illustrated in Fig. 2. Arterial blood pressure signal in the simultaneous recording with PPG acts as ground truth to compare and quantify the prediction of mean arterial pressure using proposed features. Preprocessing of raw PPG, key points detection, feature extraction, and estimation stages are further explained in the below subsections.

Fig. 2.

Estimation framework for mean arterial pressure using raw PPG

Pre-processing and normalization

The raw PPG data during acquisition remains vulnerable to a variety of noise, including baseline wandering, certain low-frequency components in the form of respiratory noise, and electromyograph strays as higher frequency noise [22]. The preprocessing of the raw PPG follows the below steps:

• The raw PPG data is first re-sampled to 1KHz, 10-level wavelet decomposition is performed using the db8 mother wavelet.

• The wavelet coefficients, which represent the low-frequency components (baseline wandering), are eliminated by suppressing the coefficient to zero.

• Similarly, the wavelet coefficient representing the noisy high-frequency component like muscle activity artifacts and power line harmonics are removed by zeroing those coefficients.

• Finally, we apply the soft rigrsure method [24] over the remaining coefficients and reconstruct the PPG signal in time domain using the remaining coefficients. Rigrsure is an adaptive thresholding selection approach using the Stein’s criteria of unbiased risk prediction [25] given by :

  $$\begin{array}{c}\hfill Thres{h}_v=\sigma \sqrt{2log(N{log}_{2}N)}\end{array}$$ 2

  where, $\sigma$ represents the noise standard deviation, N refers to length of PPG signal.

This yields an artefact free clean PPG wave suitable for feature extraction. We have examined the effectiveness of a number of filtering and denoising techniques, including discrete wavelet decomposition (DWT) [26], infinite impulse response, empirical mode decomposition [27], and finite impulse response filters. Finally, wavelet denoising is chosen over other methods. The DWT is well known for its adaptive transformation property which results in gross segmentation of the associated signal with better energy compaction within wavelet subbands. These properties of DWT make it an effective tool for signal denoising. In addition, compared to other denoising methods, DWT provides a number of other merits such as a computational efficiency, adaptivity in different signal-to-noise ratio regimes and even helps dealing with non-stationary artifacts.

The corrupted PPG with noise and its corresponding clean version using the above described pre-processing approach is shown in Fig 3. We have also computed the average Percent Root Mean Square difference (PRD), Root Mean Square Error (RMSE) and Signal to Noise Ratio (SNR) between the raw PPG episode and its corresponding clean version each from MIMIC-I and MIMIC-II database and provided the results in Table 1. As we can observe that a lower PRD value signifies a better reconstructed PPG from the coefficients of interest that ensures a preserved morphology which retains clinical information. SNR value is calculated by the ratio of the power of clean PPG to the power of difference between the power of raw PPG and clean PPG. A significantly higher value of $SNR=23.95dB$ indicate the robustness of the proposed pre-processing approach to remove most of the artefacts from the raw PPG while maintaining its morphology.

Fig. 3.

Corrupted cycles of PPG with high frequency noise and its corresponding clean version by preprocessing with wavelet denoising scheme. black markers on clean PPG shows the detected systolic peak and diastolic peak from clean PPG. The corrupted parts are encircled

Table 1.

Percentage root mean squre difference (PRD), relative signal-to-noise ratio (SNR), and root mean squared difference (RMSE) calculated between raw PPG and clean PPG after preprocessing

Database	PRD	SNR (in dB)	RMSE
MIMIC-I	0.4663	23.92	1.4999
MIMIC-II	0.688	23.95	3.1513

After performing the pre-processing the amplitude values are normalized between +1 and -1 which helps to set an amplitude threshold to set eligibility for the fundamental peak of PPG to be regarded as systolic peak.

Key points detection from first and second derivative of PPG

Each cycle’s onset and the endpoint are detected to segment the clean and normalized PPG episode in a single cardiac cycle. After segmentation, each cycle’s first and second derivative is taken to see the fiducial points required to extract moving slope features. The first key point or maximum peak of PPG is referred to as the systolic peak, which is the result of a direct pressure wave that travels from the left ventricle to the periphery of the body. In contrast, the reflection observed from pressure waves of the arteries of the lower body is attributed to diastolic peak [23]. The key points ’a’ and ’b’ available in the second derivative of PPG is mainly used to calculate increased arterial stiffness of the body by the ratio b/a, which increases by aging [28]. Since it is challenging to recognize the key points from PPG in its natural form, its first derivative elevates its appearance as shown in Fig. 4.

Fig. 4.

a PPG cycles with missing diastolic peaks b First derivative of PPG (VPPG) with detected diastolic peak encircled c second derivative of PPG (APPG) with encircled ’a’ and ’b’ as its major key points

Extraction of moving slope features

After the detection of critical points from VPPG and APPG, the following steps are used to compute moving slope features:

• The section of contour between systolic and diastolic peak from VPPG is selected. Let this section is denoted as $C_v$.

• Moving slope values are calculated for $C_v$ using a sliding window of 3-points [29]. This will generate a vector of local slope values for $C_v$.

• The time duration between systolic and diastolic peak is taken denoted as $T_v$.

Now the moving slope feature can be formulated as:

$$\begin{array}{c}\hfill M{S}_v=\frac{Max[Movingslope(C_v)]}{T_v}\end{array}$$ 3

where, $M{S}_v$ is the movingslope feature for VPPG, $Movingslope(C_v)$ is the slope vector containing the computed slope values of the fragment. Similarly, we take the contour section between points ’a’ and ’b’ from APPG denoted as $C_a$ and the computed time duration between them denoted as $T_a$. The feature can be formulated as:

$$\begin{array}{c}\hfill M{S}_a=\frac{Max[Movingslope(C_a)]}{T_a}\end{array}$$ 4

where, $M{S}_a$ is the moving slope feature for second derivative of PPG or APPG. Both these features are jointly used after normalization in regression model to train with true target MAP obtained from ground truth invasive ABP signal to yield estimation results. After feature extraction a total sample size of input data in MIMIC-II and I contains 21290742 and 4594474 samples respectively. The feature set has input size of $21290742\times 6$ extracted from MIMIC-II and $4594474\times 6$ from MIMIC-I respectively. Our feature set contains attributes as amplitude of systolic peak and diastolic peak, maximum absolute moving slope derived from VPPG and APPG, and amplitude of point a and b from APPG respectively.

Estimation using different regressors with GridSearchCV

We utilized three popular regressor i.e. random forest, XGBoost and Elastic net regressor to obtain the estimation of MAP. In order to automatically tune the hyper-parameters of the model we employed grid search CV mechanism that aids in fitting our estimator (model) to our training data by looping through predefined hyperparameters range. We divided our training and testing data in the ratio of 80:20 and used ten-fold cross validation to trace accuracies for each fold and the average results of each regressor are reported. Since both the data-set has different number of samples thus the obtained best hyperparameters of random forest for regression against both datasets after GridSearchCV is provided in Table 2.

Table 2.

Optimized hyperparameter using gridCV for prediction with random forest

Data	Optimized hyper-parameters with GridSearchCV
	Estimators	Max features	Min. sample split	bootstrap	Max depth
MIMIC-II	600	log2	10	True	70
MIMIC-I	500	log2	4	True	55

Results and analysis

To demonstrate the performance of the regression model for the proposed features to estimate MAP, we utilized three metrics to compare the closeness of estimation with the reference MAP values and pulse pressure (difference between SBP and DBP, $SBP-DBP$). We computed mean absolute error (MAE), standard deviation (SD), and coefficient of determination($r^2$) after getting the predicted MAP from the test data.

Estimation results

The estimation results are summarized in Table 3. The best results are obtained with MIMIC-II data-set with $MAE=1.28$ with standard deviation of 2.50 and a strong coefficient of determination $r^2=0.92$ as it contains higher number of subjects and large data compared to MIMIC-I for training of the model. Thus, we can treat our random forest regressor trained with MIMIC-II database as generalized model for the proposed frameowrk. We also combined the data of both MIMIC-I and MIMIC-II (UCI) and obtained the minimum and maximum MAE of 3.79 and 7.06 respectively with STD of 6.66 and 10.53 respectively.

Table 3.

Estimation results in terms of MAE, STD and $r^2$ analyzed for both MIMIC-I and MIMIC-II database using different regressors with gridsearchCV

Regression Model	Dataset	MAP estimation				PP estimation
		MAE	STD	R2	RMSE	MAE	STD	R2	RMSE
Random forest	MIMIC-I	2.93	4.94	0.82	4.98	3.03	5.95	0.82	6.18
	MIMIC-II	1.28	2.50	0.92	2.67	3.47	6.42	0.78	6.99
XGBoost	MIMIC-I	3.76	4.42	0.79	4.95	4.27	5.95	0.75	6.52
	MIMIC-II	1.91	3.06	0.88	4.03	2.59	4.42	0.86	4.91
Elastic net	MIMIC-I	2.94	4.65	0.85	5.99	3.33	5.27	0.76	5.67
	MIMIC-II	2.90	3.88	0.86	4.11	3.06	4.82	0.77	5.17

Further, the scatter plot after regression is shown in Fig. 5. Maximum data for MAP prediction illustrated in scatter is found to be close to best fit line ($y=7.02\pm 0.92X$) showing a significant correlation of proposed features with mean arterial pressure. To visualize the frequency error after estimation, The error histogram for MAP and PP estimation is shown in Fig. 6. The maximum error counts are found to be in the span of $-2$ to $+2$, reflecting the preciseness of the model in regression. We compared 1000 predicted values with the reference MAP and PP value to investigate the outlier positions. We plotted the bland-altman plot shown in Fig. 7. We can see that out of 1000 difference points (between true and predicted values), only few predicted samples deviate above the agreement line under 95% confidence interval with little mean difference of 0.04 and upper and residuals as 5.04 and $-4.97$ respectively showing little outliers which reflects the robustness of regression against outliers.

Fig. 5.

Scatter plot after regression with random forest for MAP and PP.(With obtained best fit line $Y=7.02\pm 0.92X$ and $Y=15.93\pm 0.78X$ for MAP and PP respectively)

Fig. 6.

Histogram of errors obtained by difference between predicted and true value after regression for target MAP(left) and PP(right)

Fig. 7.

Bland-altman plot for the difference between predicted and true value of 1000 predictions of MAP

Biological correlation of proposed features with mean arterial pressure

The overall PPG waveform exhibits a slowly varying morphology, however it is not necessarily informative. The purpose of PPG signal derivatives is to make it simpler to analyse the original PPG waveform and to accurately identify the PPG’s points of interest. It is generally regarded that VPPG may be used to investigate vascular ageing, arterial stiffness, atherosclerosis, endothelial dysfunction, and erectile dysfunction [20]. In other words, while the first derivative gives data on blood volume, it is challenging to determine the gradient’s direction. Given that the PPG monitors variations in blood volume, the first derivative refers to changes in blood volume at a measurement site, as a result, the first derivative’s positive and negative values represent, respectively, a rise and a decrease in blood volume. The rate of blood volume changes associated to the presence of opposing pressure is provided by the second derivative. The maximum positive value in the $2\mathit{nd}$ derivative indicates the maximal rate of blood volume increase and the lowest negative peak indicates the maximal rate of volume decrease while the slowest change of volume rate change is always zero value [30].

The calculated elapsed time between the systolic and diastolic peak ($T_v$ in Equation 3) leads to an essential index known as large artery stiffness index [31] given by :

$$\begin{array}{c}\hfill SI=\frac{h}{T_v}\end{array}$$ 5

where SI is the stiffness index and h is the subject height. Since subject height information is not available in the analyzed database, it is replaced by the maximum slope value of the PPG fragment between this transit time. An increase in arterial stiffness causes an increment in cardiac output that will reflect in the rise in transit time between systole and diastole and growth in the PPG contour slope [32]. Thus, this increment elevates the mean arterial pressure according to Equation 1. Thus, the feature directly impacts MAP and correlates with cardiac output. The transmission velocity of both forward and reflected waves increases with arterial stiffness, thus the reflected wave arrives sooner in the central aorta [33]. The early arrival of reflected wave indicates the modification in steepness of the PPG wave and hence in its slope.

Time-complexity analysis

we have done time complexity analysis for the proposed approach in the model we used for estimation with MIMIC-II dataset (as this dataset has significantly large number of samples compared to MIMIC-I) illustrated in Table 4. Here, obs/sec refers to observations processed per second. Amongst all the three regressors training time in XGBoost is lowest since in this boosting each tree can be built only after the previous one and each tree is built using all cores. This makes XGBoost a very fast algorithm. The random forest regressor has moderate training time a little higher than XGBoost. Hence, by taking all parameters in consideration like estimation accuracy, prediction speed and training time it can be concluded that random forest fits well for real time execution when deployed in a BP monitoring device.

Table 4.

Prediction speed and training time of the model used for estimation of MAP in the proposed method

Model	Dataset	Prediction speed (obs/sec)	Training time (Sec)
Random forest	MIMIC-II	80000	97.041
Elastic net	MIMIC-II	630000	302.34
XGBoost	MIMIC-II	240000	57.04

Comparative analysis

Since the regression is also performed for the target as pulse pressure (SBP-DBP) that yields remarkable results given in Table 3 which proves the leverage of the proposed features for estimation of SBP and DBP but the correlation of features with MAP is 0.92 which is far better as compared with PP which is only 0.78 also the MAE and STD for MAP estimation are less compared to PP. This shows a strong correlation of moving slope features with MAP.

We also compared our best-obtained results with previously reported results summarized in Table 5. All the mentioned approaches have the same database and metric for a fair comparison with the proposed approach. It can be seen that our mean absolute error and correlation coefficient with MAP founds to be better then others due to proportional consistency of our slope features with mean arterial pressure. Previously reported works have derived numerous temporal, frequency, and whole-based features of PPG to target only systolic and diastolic blood pressure. In contrast, our proposed features have a direct association with MAP values.

Table 5.

Comparison with other approaches in terms of MAE(mmHg), STD(mmHg) and r

Methods	Database	MAE	STD	r
Morphological features[15]	MIMIC-II	4.58	5.53	0.75
Perfusion index[34]	.	6	.	0.84
Federated learning[35]	MIMIC-II	2.95	19.33
Whole based approach[36]	MIMIC-II	2.61	4.16	0.91
Temporal features[17]	MIMIC-II	1.48	2.36	0.96
Proposed work	MIMIC-II	1.28	2.50	0.92
Proposed work	MIMIC-I	2.93	4.94	0.82

Discussion

The proposed features have demonstrated a notable accuracy for estimating MAP. The approaches in previous literature had utilized either the large number of morphological features or features derived from deep network to map distinct SBP, DBP, and MAP values extracted from ABP signal [34, 36]. However, too many features for estimation leads to a significant over-fitting problem and can raise computation complexity for miniature health wearables. The merit of the proposed work lies in the derivation of rational pathological feature extracted from VPPG and APPG that clinically correlates with MAP. Thus, without exploring the variety of other temporal or frequency features of PPG, our proposed moving slope features can sufficiently estimate MAP values with adequate accuracy. In Sect. 4.2 it was explained that the variation of MAP solely depends upon the variation in cardiac output, which in turn is related to arterial stiffness. Arterial stiffness is one of the common phenomena of thickening and stiffening of the arterial wall. It is associated with high blood pressure that increases with age [37, 38]. Hence apart from taking part in the estimation of mean arterial pressure, the statistics of the proposed features can be helpful in the early diagnosis of stiffness, CVD, and hypertension. Since the PPG signals are more prone to noise during acquisition, careful pre-processing is the primary requirement to spot valid fiducial points in PPG contour. The limitation of proposed feature lies in the severe deviation of PPG from its natural shape that can affect the analysis and extraction of local slope values. However our proposed pre-processing scheme is robust to remove most of the artefact present in PPG to make it suitable for key-point detection and feature extraction. Further the derivatives of PPG also helps to elevate the detection of fiducial points. Hence, the proposed work can hold a good estimation accuracy under moderate acquisition condition of PPG signals via wearable and portable non-invasive BP monitoring devices. In a clinical domain like haemodialysis environment, MAP has been used as a guide to monitor critical intravascular changes in patients. It is indicated that MAP variance can be used to measure changes in patient condition in different health care settings, from acute post-operatively to rehabilitation or chronic areas, when compared to a patient’s customary MAP at baseline on admission to hospital. Although a patient’s blood pressure may appear to be normal, the MAP may reveal a major change in their intravascular state. This is especially noticeable when the systolic reading is substantially lower than normal. A patient with a BP of $105/75(MAP=84)$ may have a very different condition than a patient with a BP of $105/45(MAP=64)$, especially if this is lower than their typical MAP, even if they are asymptomatic at the time of BP recording. Indeed, a proper BP assessment must consider more than just the systolic figure; the diastolic number, which is often disregarded, is an important component of blood pressure value.

Conclusion

The true driving pressure for peripheral blood flow is mean arterial blood pressure, which is physiologically a better indication of perfusion to important organs than systolic blood pressure. Peripheral resistance can be estimated using MAP readings and cardiac output measurements. The proposed features for this estimation hold a good biological correlation by capturing the maximum slope of the PPG curve between the diastole and systole phase. Future work in this area can include the derivation of some other pathological indices independent of critical points of PPG and can associate with MAP. Analysis with multiple data-set and use of only PPG signal yielding consistent results indicate the suitability of this framework in personalized health monitoring devices and can be helpful for cardiac practitioners and physicians for early diagnosis of hypertension and cardiac anomalies.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declarations

Conflict of interest

All authors declare that they have no conflicts of interest.

Consent to participate

Not applicable

Consent to publish

Not applicable

Ethical approval

Not applicable