1. Introduction

Blood pressure is a vital health indicator and provides patients with significant health information since a major risk factor for the cardiovascular system is related to blood pressure ^[1]. In particular, continuous blood pressure monitoring is crucial for patients with hypertension ^[2], so several studies have suggested cuff-free blood pressure monitoring techniques based on the pulse transit time (PTT) ^[3]. PTT is the pulse-wave time interval between two artery locations measured from an electrocardiogram (ECG) peak to a photoplethysmogram (PPG) peak ^[4,5]. Extracted PTT features using the ECG and the PPG signals could be used to estimate the blood pressure ^[6].

Wang et al. applied a linear regression model to estimate blood pressure using ECG and PPG signals ^[7], and a Lavenberg-Marquardt artificial neural network (ANN) has been also applied ^[8]. Additionally, a ballistocardiogram (BCG) was also investigated to predict blood pressure together with ECG and PPG signals, and bidirectional long short-term memory neural networks were utilized ^[9].

In addition to PTT features, the combination of statistical features of physiological signals was studied ^[10]. Nevertheless, individual variation factors such as age, gender, medication, and diseases could affect estimation performed with a single cardiac cycle due to the higher blood wave velocity ^[11]. Investigation of blood pressure with lower wave velocity is crucial to improve the performance of blood pressure estimation.

In this study, we investigated the effects of the lower blood wave velocity on blood pressure estimation while adjusting the number of cardiac cycles. Analysis and comparison of blood pressure estimation while varying the cardiac cycle number could be used to confirm the optimal cardiac cycles. The remainder of this paper is structured as follows. Section 2 illustrates the data acquisition during the experiment. We describe the method in detail in section 3, including the preprocessing, feature extraction, estimation, and performance evaluation methods. In section 4, we present the experiment results and discuss them in section 5. In the final section, the conclusion of this final study is provided.

2. Data Acquisition

There were 16 people participated as subjects in this study (age: 24.7$\pm $ 3.6 years; height: 168.4$\pm $4 cm; weight: 69.6 $\pm $ 6 kg). Each subject conducted a continuous 30-minute blood pressure measurement using a Finometer device (Finometer Pro, Netherlands) on the middle finger of the left hand. The recorded finger pressure (FP) acquired with the Finometer was used as the reference blood pressure. The ECG signal was recorded according to Einthoven’s triangle ^[12,13], and the PPG signal on the index finger of the right hand was measured using a Biopac MP36 ^[14]. Two BCG signals were recorded from the back and the seat while the subjects were sitting on a chair. All the measured data were synchronized and sampled at 1 kHz. During the experiment, the right foot of the subject was submerged in cold water at 4 to 6$^{\mathrm{o}}$C to elevate their blood pressures ^[15], as can be seen in Fig. 1.

Fig. 2 illustrates the values of systolic and diastolic blood pressures and their distributions acquired using the Finometer device. These values were used as reference data for the estimation of blood pressure using ECG, PPG, and BCG signals. The mean and standard deviation of DBP and SBP were 62.96 $\pm $10.37 mmHg and 113.40 $\pm $13.12 mmHg, respectively.

Fig. 1. Experimental setup of blood pressure measurement.

Fig. 2. Continuous blood pressure acquired using a Finometer (top: continuous systolic and diastolic blood pressures, bottom: the distribution of blood pressures).

3. Method

3.1 Preprocessing

Fig. 3 shows the overall process of blood pressure estimation. In this study, the preprocessing includes bandpass filtering and peak detection. Four input signals were sampled at a sampling frequency of 1 kHz. A Butterworth bandpass filter was applied to each input signal to remove noise and baseline wandering with the various cutoff frequencies corresponding to the physiological signals (see Table 1) ^[16].

Peak detection was then applied to each filtered signal to find the peak points of the signals. Next, the distance between ECG and PPG peaks was calculated. In our proposed scenario, we estimated blood pressure using one to three cardiac cycles for comparison. In this study, PTT in one cardiac cycle was calculated from the ECG peak to the peak of the other signals, which is denoted as RB1 in Fig. 4. The peak distance in two cardiac cycles was calculated from the ECG R-peak in the first cardiac cycle to the PPG B-peak in the following cardiac cycle, which is denoted as RB2. RB3 was calculated using the distance between the first ECG peak and the second following cardiac cycle. In the same way, the distance between ECG peak and BCG peak was also calculated.

Fig. 3. Block diagram of blood pressure estimation.

Fig. 4. The distance between ECG and PPG peaks in the time domain.

Table 1. Cutoff frequencies of the bandpass filters for the physiological signals.

Signal	HPF (Hz)	LPF (Hz)	Order
ECG	0.5	35	2
PPG	0.5	15	2
BCG	4	15	2

3.2 Feature Extraction

Using the distances of ECG-PPG peaks and ECG-BCG peaks, features were extracted. There were seven features in total: RB, RJ1, RJ2, ECG, PPG, BCG1, and BCG2 peak amplitude, as can be seen in Fig. 5. RB is the feature extracted from the distance between ECG and PPG peaks and defined as PTT ^[17]. PTT is calculated with Eq. (1).

(1)

$ PTT=\frac{xBpeak-xRpeak}{f_{s}} $

where $xBpeak$ and $xRpeak~ $are the time instances of the PPG and ECG peaks, and $f_{s}$ denotes the sampling frequency of 1 kHz.

Fig. 5. 1000 samples of nine features (1000 out of 37,044 data points) are used for blood pressure estimation: (a) RB; (b) RJ1; (c) RJ2; (d) ECG peak amplitude; (e) PPG peak amplitude; (f) BCG1 peak amplitude; (g) BCG2 peak amplitude.

4. Results

Blood pressure estimation was performed using three scenarios: one cardiac cycle, two cardiac cycles, and three cardiac cycles. The results in the three scenarios are then compared. Table 2 shows the model performance with respect to the number of cardiac cycles in estimating diastolic blood pressure. It is demonstrated that three cardiac cycles yield a better result compared with two cardiac cycles and one cardiac cycle.

Table 3 provides the model performance of systolic blood pressure estimation with respect to the number of cardiac cycles. The lowest value of RMSE was obtained using two cardiac cycles, but the MAE, $R^{2}$ and MAPE values were the best when using three cardiac cycles. This demonstrates the significance of the three cardiac cycles for the estimation of systolic blood pressure.

Fig. 6 illustrates the error between the estimates and ground truth of blood pressures using three cardiac cycles. For the error of diastolic blood pressure estimation, the 95% confidence interval was calculated as [-6.3, 4.7], and [-11, 12] was used for the error of systolic blood pressure estimation. The mean differences between the estimation and the ground truth of diastolic/systolic blood pressure were 0.77 and 0.6, indicating a small bias of the proposed model.

Fig. 6. Bland-Altman plot of blood pressure estimation using three cardiac cycles (top: the estimated diastolic blood pressure, bottom: the estimated systolic blood pressure).

5. Discussion

In this work, we investigated the effects of multiple cardiac cycles to estimate blood pressure. In particular, significant features were extracted between ECG and PPG peaks and between ECG and BCG peaks. Distance features (PTT) have been reported to have high correlation with blood pressure ^[3]. Numerous algorithms have used these features to estimate blood pressures in a single cardiac cycle. Our main contribution in this study is the feature extraction from multiple cardiac cycles, which has more information than a single cardiac cycle for estimating blood pressure. We also investigated the effects of these features on the performance of blood pressure estimation.

where PTT is the amount of time for a blood wave to move across two body locations, $L$ denotes the distance across which the wave propagates and $PWV$ denotes pulse wave velocity.

6. Conclusion

In this study, we proposed a cuffless blood pressure estimation model using multiple cardiac cycles, which have more information than a single cardiac cycle. The results showed that three cardiac cycles yield more accurate blood pressure estimation than two cardiac cycles or one cardiac cycle.