Optimized Cuffless Blood Pressure Measurement Using ECG, PPG and Linear Regression

1. Introduction

Blood pressure (BP) monitoring is one of the crucial aspects of cardiovascular health management [1], [2]. Nevertheless, traditional cuff-based methods can prove inconvenient and uncomfort [3]. In response, the SINTEC European project [4] has developed an in- novative blood pressure estimation algorithm specifically designed for wearable cuffless devices. This algorithm utilizes SINTEC’s Electrocardiogram (ECG) and Photoplethysmo-gram (PPG) cuffless sensors, along with state-of-the-art Shimmer sensors [5], demonstrating promising accuracy even with low-frequency sampled signals. In this work we present a BP monitoring device allowing frequent BP measurements (ideally continuous BP monitoring), with minimal discomfort.

This article explores the meticulous development and testing of the SINTEC algorithm in both laboratory and hospital settings, underscoring its potential to transform continuous and non-invasive blood pressure monitoring through wearable technology. The study involved a 20-minute recording to evaluate the machine learning regression coefficients crucial for algorithm calibration. SINTEC’s expertise played a key role in fine-tuning the algorithm to accommodate low-frequency sampled signals and different signal sampling frequencies for ECG ( fsE) and PPG ( fsP), all integrated into a unified algorithm. This adapt- ability not only ensures accuracy, but also positions the algorithm for future applications on wearable devices. This opens up prospects for measuring blood pressure in a convenientand dependable way, enhancing user comfort and overall reliability.

1.1. Key measurements relevant to Blood Pressure estimation: Heart Rate and and Pulse Transit Time

The BP algorithm relies on key measurements—Heart Rate (HR) and Pulse Transit Time (PTT)—from ECG and PPG signals to estimate BP [6,7]. The ECG captures the heart rhythm, allowing for the derivation of HR (Eq. (1)), [8]. This parameter, given in beats perminute, is calculated from the time interval between R-peaks, offering a precise measure of cardiac activity [9]:

(1)

PPG, an optical signal, measures pulse patterns, revealing PTT: the duration it takes for blood to travel between two points in the circulatory system. PTT is assessed as the time interval between the R-peaks of ECG and the subsequent S-peak identified in PPG signal [10], as illustrated in Figure 1.

1.2. Estimation of Blood Pressure using Bramwell-Hills and Moens-Kortweg Equations

The correlation between BP and the HR and the PTT is established through the Bramwell-Hills and Moens-Kortweg equations, complemented by the pioneering findings of Leslie A. Geddes [12,13,14]. The synthesis of these studies results in a linear approximation of BP, expressed in Eq. (2):

(2)

The coefficients a, b and c are subject-specific parameters that must be determinedthrough a calibration process. These coefficients depend on various factors, including vessel wall thickness, blood density, vessel diameter, length, and elastic modulus. Eq. (2) serves as the linear regression model employed in this study [15], [16]. To estimate maximum and minimum BP values, the assessment of Arterial Blood Pressure (ABP) is divided into Diastolic Blood Pressure (DBP) and Systolic Blood Pressure (SBP):

(3)

2. Material

2.1. Devices

Shimmer and SINTEC devices and their respective setups, utilized in this study, are shown in Figure 2 and Figure 3. In Table 1, the SINTEC devices are compared to the state-of- the-art Shimmer devices. The algorithm was trained while wearing devices with cuffs during the calibration phase. Laboratory recordings were conducted using the Omron HeartGuide [17] to record BP values , whereas in the hospital environment, a professional sphygmomanometer was employed for accurate measurements.

2.2. Data collection

This study comprises 87 acquisitions, 18 conducted in a hospital and the remainder in a laboratory environment. Among them, 12 were acquired exclusively with SINTEC devices in the hospital setting. Additionally, in the hospital, it was possible to obtain 6 simultaneous acquisitions using both Shimmer and SINTEC devices on the same subjects. This study aims to understand the algorithm performance in various situations, considering different settings and signal characteristics.

3. The algorithm

In the section, the major steps of the BP algorithm are illustrated. The algorithm was implemented in Python; however, the code can be easily adapted for use in other coding environments like MATLAB.

3.1. Signal loading

ECG and PPG signals are provided in .mat files, each containing the signal data and the associated timestamp in Unix format. Additionally, a separate .csv file is loaded, containing 20 blood pressure recordings obtained from the reference device required for calibration. The .csv file is organized into three columns: Unix timestamp, SBP and DBP. To ensure consistency, the timestamps from the reference device were adjusted by subtracting 60 seconds from each value. This adjustment accommodates the fact that the measurements in the .csv file denote the time of the readings from the reference device, typically allowing for one minute per measurement. Then, it necessary to specify the sampling frequency for both ECG and PPG signals ( f_sE, f_sP).

3.2. Signal filtering

Signals are filtered after the loading phase. The delay introduced by the filtering techniques has minimal impact on the overall performance of the algorithm and it remains uncompensated. The delay introduced by the baseline removal function is identical in both signals, while the delay introduced by the low-pass Butterworth filter is negligible.

ECG filtering

An initial signal quality assessment, done during SINTEC project, reveals that ECG signals can be affected by baseline wander, complicating the determination of R-peaks detection. In particular, the Shimmer ECG signals were more susceptible to baseline wander than the SINTEC ECG signals. To mitigate this, a function is introduced to flatten the input ECG signal and apply a moving average filter using a specified window size. The window size is experimentally set at 0.5 seconds. The smoothed signal is then subtracted from the original, eliminating the baseline noise, Figure 4. This approach has demonstrated superior results compared to the baseline removal function based on envelopes adopted in [16], as it minimizes peak distortion, especially in low-sampled signals.

PPG filtering

Due to the narrow bandwidth of PPG signals (with a low-frequency range of PPG Power Spectral Density from 1 Hz to 8 Hz) [18], a non-equalized low-pass Butterworth filter is applied to the PPG signal to eliminate high-frequency noise. This filter is specifically configured with a 10 Hz cut-off frequency and a 8 Hz attenuation frequency. The chosen filter order is 4, as recommended in [19]. The passband ripple and stopband attenuation are set at 0.1 dB and 15 dB, respectively. Following the filtering process, the signal undergoes further processing through the previously described baseline removal function, whose window size is set at 0.5 seconds. An example of the obtained result is shown in Figure 5.

3.2. Signal preparing

Filtered signals are then aligned and cut in the following steps. Aligning signals is essential for optimal HR and PTT measurements. Ensuring simultaneous starts for both ECG and PPG signals is necessary, and this requires identifying corresponding start samples at the same time point. To achieve this alignment, the timestamp of the first measurement from the reference device is utilized. It helps determine the initial samples of both ECG and PPG by identifying the indices with the least positive differences between the reference timestamp and the signals. Then, the signals are cut to a duration of 20 minutes, according to this study scenario.

3.3. Key measurements for BP estimation: PTT and HR

A critical step involves detecting R-peaks and S-peaks from ECG and PPG signals. Experimentally selecting optimal thresholds for peak detection in both signals is necessary. To enhance peak detection precision, local maxima detection is performed by eliminating smaller peaks within a specific time interval. This interval is set experimentally as half of the time difference between adjacent peaks time stamps. R-peaks and S-peaks timestamps are then used for HR and PTT calculations. First, each S-peaks timestamp is matched with the corresponding R-peaks timestamp, sorted chronologically. The selected R-peaks timestamp is one that precedes the current S-peaks timestamp in processing order. If the R-peak timestamp is found, PTT is calculated by subtracting the R-peak timestamp from the current S-peak timestamp. The identified R-peaks timestamps are then used to calculate HR, as shown in Figure 6.

This approach offers a dual advantage. It eliminates the requirement for ECG and PPG signals to be sampled at the same frequency. This flexibility proves beneficial in conserving bluetooth bit transfer rate, especially for wearable devices. Furthermore, PPG signal quality poses challenges as peaks can be significantly attenuated. By selecting key points from S-peaks timestamps, more reliable measurements can be obtained. This is because S-peaks detection involves selecting a threshold based on visual inspection, ensuring accurate identification of peaks.

3.5. Selection of training and test points

It is necessary to generate an interpolating time array from ECG and PPG timestamps to facilitate proper interpolation of the reference device timestamps and blood pressure values (SBP and DBP). The same interpolating time array is utilized for interpolating HR and PTT values. Subsequently, a function is applied to diminish the dimensionality of the data, yielding arrays representative of essential measurements over 10-second intervals.

3.6. Linear Regression

The final step involves preparing the training and test datasets, comprising 70% and 30% of the entire dataset, respectively. A cleaning process is performed on the training set. The outliers are identified using the mean and standard deviation. The mean and standard deviation of the PTT and HR data are calculated, and any data points that fall outside the range of mean ± standard deviation are considered outliers. PTT and HR serve as input data for the selected linear regression model in this study. The model is trained and tested using the SBP and DBP values of the reference device. Subsequently, the model provides predicted values for SBP and DBP, along with the corresponding mean absolute errors (MAE) and standard deviations (SD). In Figure 7 predicted and actual values of SBP and DBP using the BP algorithm on SINTEC signals.

4. Results

According to ANSI/AAMI/ISO 810602:2018 [20], for a database of at least 85 measurements, the difference between the investigated method and the reference device should be:

(4)

The BP algorithm, employing the timestamp approach, has demonstrated promising accuracy. Table 2 presents the average MAE and SD values obtained across the entire database. The results, when compared with the ANSI/AAMI/ISO ISO 81060-2:2018 guideline [22], indicate significant potential for the practical application of the BP algorithm in real scenarios involving wearable devices.

In Table 3, a comparison of BP accuracy results for SINTEC and Shimmer signals in a clinical environment is presented. A clinical setting, characterized by power line interference and other noises stemming from the concentrated presence of instruments, was considered [23]. Despite SINTEC devices facing challenges such as Bluetooth interruptions and high noise sensitivity, the results remained promising and demonstrated a slight improvement over the outcomes obtained in the same acquisition session with Shimmer 176 devices.

These results are encouraging for the feasibility of wearable solutions in BP monitoring algorithms using this study algorithm. It appears that achieving an acceptable level of accuracy in BP estimation does not necessarily require highly-frequency sampled signals, as demonstrated by the comparable performance of SINTEC and Shimmer signals. Focus on peaks, rather than the morphological goodness of signals, contributes to the robustness of BP estimation even at lower fs. This not only benefits Bluetooth communication and battery life but also highlights the resilience of the algorithm to interruptions and noise in clinical environments.

5. Discussion

BP algorithm aims to long-term monitoring of BP in both clinical settings and daily activities. However, prolonged monitoring poses challenges as both ECG [24,25] and PPG [26] signals become more susceptible to corruption by MA, potentially impacting prediction accuracy. To overcome this challenge, two strategies involving accelerometer signals have been envisioned. The following strategy has been assessed on 40 Shimmer signals deliberately corrupted by motion artifacts, and the results were discussed in [27], representing a starting point for future strategies.

5.1. Adaptive filter

Ad discussed in [27] to mitigate potential motion artifacts induced by subtle movements, such as hand gestures or walking, an adaptive filter can be employed. Specifically, the filter utilizes the norm of accelerometer signals, recorded simultaneously with physio-logical signals, as the target signal for the Least Mean Squares (LMS) adaptive filter, [28]. The LMS filter takes the noisy PPG and ECG signals as input signals, producing denoised ECG and PPG signals by computing the difference between the noisy signals and the error output of the LMS filter, as illustrated in Figure 8.

5.2. Peaks cleaning

An additional improvement to the BP algorithm could involve the introduction of a signal peak cleaning mechanism, specifically designed to address instances where peaks may be attributed more to MA than physiological activity. This improvement utilizes the acceleration norm as a reference signal [29], enabling the algorithm to identify potential time windows susceptible to severe MA, as proposed in [27] and illustrated in Figure 9

Thresholds employed to determine whether R-peaks and S-peaks fall within a critical time window should be meticulously chosen. Furthermore, by selecting the time window duration, the algorithm discards points with a heightened likelihood of being corrupted by motion artifacts.

6. Conclusion

This study investigated the timestamp approach algorithm, demonstrating its suitability in terms of accuracy for predicting BP using wearable and comfortable devices, even with low-frequency sampled signals. The proposed method for cuffless blood pressure monitoring has great potential for prevention and monitoring of diseases such as hypertension. While the results show promise, further investigation is needed, especially in exploring machine learning coefficients for long-term monitoring applications. This additional exploration should focus on potential changes over time and identify the optimal calibration approach for sustained accuracy [27,30]. Technological improvements are deemed essential to enhance market availability. The implementation of edge computing algorithms on wearable sensors holds the potential to further improve BP accuracy by deploying effective denoising techniques tailored to specific cases and integrating alarm systems. Future wearable systems should not only address technological alarms, such as electrode detachments, but also incorporate alerts for anomalous BP values [31].