A Pilot Human Study on Dual Photoplethysmography for Continuously Monitoring Pulse Wave Velocity and Blood Pressure

Abstract

Hypertension is a leading contributor to cardiovascular disease, stroke, and chronic kidney disease worldwide. However, current blood pressure (BP) monitoring techniques, whether auscultatory, oscillometric, or invasive, are constrained by episodic measurement, user variability, and limited feasibility for real-world, continuous monitoring. These limitations hinder early detection of abnormal BP patterns and long-term cardiovascular risk assessment. This paper presents a noninvasive device based on photoplethysmography (PPG) signals, with calculated Pulse Wave Velocity (PWV), that combines off-the-shelf sensors with an infrared-based dual PPG sensor. The device combines PPG signals and physiological measurements to predict blood pressure comprehensively using machine learning, with initial trials on 25 participants demonstrating promising accuracy with mean differences of 1.1 mmHg for Systolic Blood Pressure (SBP), −0.9 mmHg for Diastolic Blood Pressure (DBP), and −0.2 mmHg for Mean Arterial Pressure (MAP) compared to standard measurements with the Vicorder device. While primarily tested on a young population, the device shows potential for continuous, real-world BP and PWV monitoring, offering greater usability and accessibility than traditional tonometry- and oscillometry-based devices.

I. INTRODUCTION

Cardiovascular disease (CVD) remains the leading cause of death globally [1] with hypertension serving as a significant risk factor [2,3]. While early detection of hypertension is crucial for implementing timely interventions and preventive strategies, current monitoring methods face substantial technical and practical limitations. Intermittent clinical measurements limit blood pressure (BP) assessments or require invasive or inaccurate procedures for continuous monitoring [4]. The constraints of existing technologies, from white coat hypertension in clinical settings (where patients exhibit transient elevated BP measures in clinical settings) to the complexity of continuous monitoring systems highlight the pressing need for advanced solutions that can provide reliable and continuous measurements in real-world environments [5,6]. Continuous monitoring of BP may facilitate the early identification of hypertension and enable the implementation of preventative medical interventions to slow or reduce the progression of associated health complications, such as stroke, chronic kidney disease, and cerebrovascular disease [7,8,9,10,11].

Intermittent blood pressure monitoring can be performed through several established methods, including auscultatory and automated oscillometric devices, conducted in clinical settings and at home. Intermittent blood pressure monitoring techniques face several limitations. Traditional auscultatory devices require calibration, vary in accuracy depending on operator skill, and give measurements that only show a snapshot of a patient’s cardiovascular health [12,13]. Handwritten logbooks, still used by many patients, especially older adults, can suffer from missing entries or misreported values. Even with digital tracking, challenges persist as data may be recorded inconsistently, stored over varying time frames, or include measurements from multiple users sharing the same device, complicating interpretation [2]. Recognizing these limitations, recent international guidelines have emphasized the role of out-of-office BP monitoring in diagnosis and treatment planning [14].

Continuous non-invasive blood pressure monitoring would likely benefit both patients and clinicians. However, the current gold standard for continuous BP monitoring is arterial catheterization. Arterial catheterization is highly accurate but is invasive and restricted to hospital settings with specialist equipment and training [15]. In recent years, a concerted effort has been made to explore the continuous monitoring of blood pressure by exploiting its relationship with arterial stiffness [16].

The walls of large elastic arteries, such as the aorta, comprise vascular smooth muscle cells and connective tissues, including highly elastic elastin and stiffer collagen fibers [17]. Because of this composition, the pressure-volume relationship is non-linear. At low distending pressure (pulse pressure), the tension is borne by elastin fibers, whereas at higher distending pressure, stiffer collagen fibers are recruited, making the artery stiffer [18]. As such, blood pressure and arterial stiffness are interrelated. Arterial stiffness is hence a predictive of cardiovascular morbidity and mortality, particularly in elderly populations, and has been endorsed as a clinical marker by multiple studies [19]. Pulse wave velocity (PWV) is a non-invasive measure of arterial stiffness. PWV represents the velocity at which the forward pressure wave generated by the contraction of the left ventricle is transmitted through the arterial tree. The determination of PWV involves dividing the path length (the distance between proximal and distal recording sites on the body) by the time taken for a pulse wave to travel that distance (pulse transit time [PTT]). PTT can be reliably assessed non-invasively using PPG [20]. PPG can detect changes in blood flow volume to detect arterial pulse waves. In contrast to traditional measures of PWV, PPG-derived estimates have typically utilized electrocardiogram (ECG) signals as the proximal waveform, whilst using PPG to detect a distal pulse wave. This approach uses the peak of the R wave on the ECG to represent the initial transmission of the pulse wave through the arterial tree and the foot of the distal PPG pulse wave. By using the peak of the R wave, this approach does not account for the pre-ejection period (PEP) – the delay between depolarization of the cardiac tissue and the mechanical contraction of the ventricle – and thus represents pulse arrival time (PAT) rather than PTT. Previous research using this approach has shown moderate agreement with reference BP devices [21]. However, by not accounting for the PEP in its derivations, the utility of this approach in the measurement of continuous blood pressure may be limited. By contrast, monitoring continuous blood pressure more accurately may be possible by utilizing PPG sensors at both proximal and distal sites and using the recommended foot-to-foot approach to determine PTT.

PPG sensing technology has been well established in various smart watches and equivalent products to measure heart rate and blood oxygen saturation. While the combination of ECG and PPG has been tested as a means of estimating BP, the capability of using multiple PPG sensing units to yield measurement of BP remains to be explored [22]. Here we report a wearable optoelectronic system that consists of two separate PPG sensors with one placed at the toe and the other at the finger, in order to achieve high accuracy of PPT tracking.

Our system will leverage a machine learning algorithm to improve the estimation precision of continuous BP measurements based on a combined consideration of measured PPT and other physiological parameters, including height, weight, average heart rate, average PWV, and average period of pulse. This pilot study aims to investigate the feasibility of a novel wearable device that uses photoplethysmography (PPG) and finger-toe pulse wave velocity (ft-PWV) [23] oscillometry-based BP and PWV monitoring devices. The device seeks to comprehensively predict blood pressure by leveraging PPG signals and physiological measurements, demonstrating promising accuracy compared to standard measurements. The pilot study will compare the data collected from the wearable device to a reference device (VICORDER) during rest and select everyday activities to assess the feasibility and reliability of the novel approach for continuous, real-world BP monitoring.

II. Methods

A. Design

This pilot study aims to 1) investigate the accuracy and precision of our continuous PPG Blood Pressure tracking device, 2) evaluate data from our novel wearable pulse wave sensor and data acquired with FDA-approved devices (Vicorder), and 3) evaluate the difference in accuracy between our fabricated wearable pulse oximeter and the Vicorder. To accomplish these aims, 25 healthy adult participants were recruited in this prospective observational study to evaluate the wearability and efficacy of our BP wearable monitor in comparison to the FDA-approved Vicorder device, which uses oscillometric cuffs at the carotid and femoral sites to measure pulse waveforms. Our study design collects reference data using a blood pressure cuff, i.e., systolic and diastolic BP measurements. This was followed by data acquisition with the Vicorder sensor placed at the carotid and femoral arteries. The distance between the two locations was calculated to compute PWV using PTT. Once this data was acquired, two PPG sensors were utilized for data acquisition at the left index finger and left big toe. Both sensors were attached to the skin using hypoallergenic cloth, PDMS (polydimethylsiloxane) materials, and silk to prevent skin irritation. This setup was used for data acquisition with each participant for three trials. The primary and secondary efficacy endpoints for this study are 1) our fabricated BP wearable monitor is safe and comfortable to wear, and 2) the performance of the BP wearable monitor is comparable to the commercial Vicorder device. The outcomes of this study will inform the future next steps for full-scale clinical studies utilizing our BP wearable monitor.

B. Participants and Consent

The human subjects study was approved by the office of human research ethics at the University of North Carolina at Chapel Hill (Protocol no. 22–0163). All human subjects gave written and informed consent before participating in the studies. The target audience for our pilot study was healthy adults with no prior history of cardiovascular-related medical conditions, since the study seeks to demonstrate the functionality of our BP wearable monitor. Participants were required to have no clinically significant skin disorders, no history of allergic reactions to silicone or adhesives, and no broken or irritated skin at sensor application sites (e.g., finger, wrist, temple, carotid, arm, neck, thigh). Individuals physically or cognitively unable to perform activities of daily living were excluded, as were pregnant or lactating women. All participants provided IRB/IEC-approved written informed consent before study initiation. Information on the purpose of the study, methodology, and extent of participant involvement was communicated in verbal and written formats while providing the participant with an ample opportunity to inquire about the study details. Participant confidentiality was maintained by assignment of a coded number to each subject. Participant data was de-identified from personal contact and identifying information and stored in a password-protected data management system used by the research team. Research data corresponding to each participant was determined with a unique identification number. All study databases were properly archived after the study.

C. Continuous Finger-Toe Sensing Device

A dual-channel PPG sensor was designed to provide continuous PWV and BP measurements. Two MAX30102 devices emitted near-infrared light (wavelength of 870–900 nm) into the finger and the toe. The amount of light absorbed varies inversely with blood volume at the recording location. The photo detector picked up the backscattered light modulated by tissue scattering and absorption, generating a proportional photocurrent [23]. An integrated ADC converted the analog signal from the photodetector into a digital signal. This digital signal represents the intensity of the reflected light over time. Each sensor was set to a sampling rate of 400 Hz, with a sample average of 4 adjacent data points for a smoother output. The signal was sent as I2C; hence, a switch controller (TCA9548A) was used to switch between the sensors to read both finger and toe data. A microcontroller (Arduino UNO) was used to send data to the Universal Asynchronous Receiver/Transmitter (UART), which was read and compiled by Python to create data arrays. A strap was fastened onto the device to facilitate easy placement and optimal signal acquisition, allowing for consistent and reliable signal acquisition throughout the study. The combined data from the sensors was read by a microcontroller (Arduino UNO) and transmitted to a computer for further processing. These functions were implemented using Arduino Integrated Development Environment (IDE) scripts, enabling seamless data acquisition and transfer for subsequent analysis.

D. Signal Processing

The photoplethysmography (PPG) data is organized into a file with a time index and sensor readings from both the finger and toe. The columns include the time index (representing the timestamp of each measurement), sensor readings from the PPG sensor placed on the toe, and sensor readings from the PPG sensor placed on the index finger. In addition to the PPG signals, the dataset also includes metadata for each participant, specifically age, height, weight, average pulse transit time, average heart rate, average PWV, and average period.

Raw PPG signals transmitted from the sensors underwent a series of processing steps to extract meaningful physiological information. The raw PPG signals were passed through two subsequent filters. The first filter applied was a moving average filter with a window size of 5 data points, designed to smooth the signal by averaging out short-term fluctuations caused by noise and high-frequency noise. The choice of a window size of 5 data points strikes a balance between eliminating spurious noise and preserving the inherent temporal dynamics of the physiological signal. This filtering step was essential in improving the signal-to-noise ratio, reducing fluctuations in the signal, and allowing for easier trend detection.

Following the initial smoothing, the signal was refined using a third-order Butterworth filter with a passband between 0.5 and 6.5 Hz. The Butterworth filter, known for its maximally flat frequency response within the passband, was selected to eliminate low-frequency noise, such as baseline wander, and higher-frequency interference that may arise from sensor or environmental noise. The choice of this frequency range ensured the retention of physiologically relevant components of the PPG signal, corresponding to the heart rate and respiratory variations, while discarding unwanted noise. This dual-stage filtering approach optimized the signal for downstream analysis, ensuring PPG waveform fidelity was preserved for the subsequent peak detection process. The PPG signal was finally inverted to resemble a traditional pulse waveform.

After this filtering process, the find_peaks function from the SciPy library was applied, identifying local maxima (peaks or troughs on uninverted plots) with tailored height and distance thresholds to ensure the primary peaks are detected. These detected peaks determine the average peak-to-peak PTT. PTT is a surrogate for the time it takes for the arterial pulse wave to travel between two sites within the cardiovascular system and is inversely related to arterial stiffness.

The finger-toe PTT (ft-PTT) is calculated using the maximum of the second derivative algorithm for foot wave detection. For PWV calculation, while a simple heuristic equation, PWV = height/PTT, can be used, a more refined approach employs a height-dependent correction factor k, where PWV = k×height/PTT.

This correction factor k is derived from anatomical studies showing that the aortic valve is a stable anatomical landmark [9]. The correction helps account for the actual arterial path length, which varies non-linearly with height. Peripheral arterial stiffness shows minimal age-related changes compared to central arteries [18]. This method was validated using height measurements from 187 individuals in occupational medicine settings [9]. Higher PWV values indicate increased arterial stiffness, which is associated with elevated cardiovascular risk.

This processing pipeline, from initial noise reduction to the final computation of PTT and PWV, was rigorously designed to extract high-quality physiological data from raw PPG signals, providing a reliable assessment of arterial function. Following the extraction of PTT, PWV, heart rate, and participant biometrics, these features were combined with the raw PPG waveforms for blood pressure estimation.

Each subject’s finger and toe PPG measurements were segmented into 45 500-sample (5 seconds) windows. Each individual window was fed into the model along with their corresponding subjects’ seven physiological features (age, height, weight, PTT, heart rate, PWV, and measurement period). Each subject had one set of blood pressure measurements and one set of physiological features, but multiple corresponding PPG segments corresponding to it.

A supervised machine learning architecture was implemented, consisting of dense layers and long short-term memory (LSTM) units to capture both morphological and temporal characteristics of the signals. The network output was trained to predict systolic, diastolic, or mean arterial pressure. This overall mapping can be expressed as

$$\begin{gathered} f(PP{G}_{toe},PP{G}_{finger},biometricfeatures) \\ \to \{SBP,DBP,MAP\} \end{gathered}$$

where f denotes the trained LSTM-based regression model. Model training used participant-level randomization (80% training, 20% testing) with a fixed random seed for reproducibility. Full architectural details and training procedures are provided in the Supplementary Materials (Section 1).

E. Materials and Device Fabrication

The materials used for PPG measurement included MAX 30102 devices, TCA9548A I2C Multiplexer, Arduino UNO module, and iHealth Track Smart Upper Arm Blood Pressure Monitor. The MAX30102 device was purchased from Amazon and manufactured by Devsolution. Headers were soldered in the lab to the VIN, SDA, SCL, and GND pins and attached by wires to the respective pins of the Arduino UNO module. These wires were bound together by tape at their header attachment points and at regular intervals along their length to prevent entanglement. An elastic band was attached to the MAX30102 module using super glue. The elastic band for the finger device was attached as a simple loop, but the band for the toe was cut at angles. Hence, it was narrower at one end to account for the changing diameter of the toe.

F. Method of Investigation

All measurements were conducted in a temperature-controlled room at 21°C to minimize external influences. Participants were instructed to lie at 30° and rest for 10 minutes to achieve a stable physiological baseline. Following this rest period, brachial blood pressure (BP) was measured using the Vicorder device, which employs a cuff-based oscillometric method. Three BP readings were obtained at one-minute intervals and averaged for each participant.

After brachial BP measurement, participants underwent PWV testing using the same Vicorder device. PPG signals were simultaneously collected using our custom-built infrared dual PPG sensor. Measurements were recorded during three separate sessions, spaced one minute apart. Although our pilot study did not assess central blood pressure directly, our methodology is informed by Obeid et al., who demonstrated strong agreement between peripheral PPG-based pulse transit times (PTTs) and carotid–femoral PWV—the gold standard for aortic stiffness assessment (r² = 0.81, p < 0.001). This supports the use of peripheral PPG-derived measurements as a noninvasive proxy for arterial stiffness and justifies comparison of our device’s BP outputs to standard cuff-based peripheral BP and Vicorder-derived central PWV.

G. Study Protocol

A total of 25 healthy adult participants (mean age 27.3 years; range 18–56 years; 20 males, five females) were enrolled in the study. All participants reported no history of cardiovascular-related medical conditions and were instructed to abstain from food, caffeine, and strenuous exercise for at least two hours before testing. Baseline demographic and anthropometric data, including age, sex, ethnicity, height, weight, arm length, and leg length, were collected. Participants then lay elevated at a 30° angle on a stretcher and rested quietly for 10 minutes to stabilize their baseline. Blood pressure (BP) was measured using an automated sphygmomanometer, with a cuff applied to the arm. Three BP readings were taken at one-minute intervals, and the systolic blood pressure (SBP), diastolic blood pressure (DBP), and their averages were recorded.

Following the BP measurements, the blood pressure cuff was removed, and preparation for the PPG measurements commenced. The participant’s carotid artery was palpated, and its location was marked for accurate sensor placement. A femoral cuff was applied near the participant’s groin area, and the distance between the carotid artery and the top of the femoral cuff was measured. This distance was multiplied by 0.8 to account for the arterial path. The participant’s bed was elevated to prevent jugular interference, and the carotid cuff was applied.

The PPG device was powered on, and data were recorded three times at one-minute intervals, with the participant remaining still to ensure accurate readings. The Vicorder system was set to capture ten peaks during each recording to ensure consistent PWV calculation. Upon completion of the measurements, all equipment was carefully removed, the area was cleaned, and the collected data were recorded in the study log. Finally, the participant’s comfort was ensured before concluding the session.

A separate, internally conducted experiment, independent of the main study, was performed to investigate the performance of device measurements while sleeping, typing, writing, and biking.

H. Analysis Plan

The main evaluation metrics for predictions were r² value, mean residual difference, and standard deviation of residuals. A higher r² value suggests the model’s predictions fit the actual data. A lower mean residual difference suggests the model predictions are more accurate on average. The standard deviation of residuals provides insight into variation in errors. A lower standard deviation implies that the model predictions are consistent. Low residual difference, combined with low standard deviation, provides complementary insight into how accurate projections are.

A Bland-Altman plot was employed to visually identify the bias between predicted and actual blood pressure. It helps check for extreme predictions by visualizing whether the difference falls within the expected limits of agreement (mean difference ± 1.96 times the standard deviation of difference).

Model prediction consistency under noisy input conditions was assessed by performing intraclass correlation coefficient analyses using repeated Gaussian-perturbed test inputs. Full details and results are provided in Supplemental Section 3a.

Repeatability per volunteer was assessed by calculating within-subject variability (standard deviation) and bias per subject (mean difference).

III. Results

Our model, which predicts blood pressure with a mean residual difference of +1.1 mmHg (standard deviation (SD) = 4.8 mmHg, r² = 0.815, p < 0.001) for systolic blood pressure (SBP), −0.9 mmHg (SD = 2.8 mmHg, r² = 0.696, p < 0.001) for diastolic blood pressure (DBP), and −0.2 mmHg (SD = 3.8 mmHg, r² = 0.326, p < 0.001) for mean arterial pressure (MAP). Bland–Altman analysis yielded 95% limits of agreement of −8.31 to +10.51 mmHg for SBP, −6.39 to +4.59 mmHg for DBP, and −7.65 to +7.25 mmHg for MAP. These plots and results show minimal systematic bias and acceptable limits of agreement for a pilot study, supporting the agreement between cuff-based and model-predicted measurements. According to AAMI/ISO criteria (mean error ≤ ±5 mmHg and SD ≤ 8 mmHg for SBP/DBP), our pilot data meets the preliminary accuracy thresholds for SBP and DBP. We emphasize this result is preliminary because full ISO validation requires a larger, protocolized cohort.

For SBP, the within-subject standard deviation averaged 2.78 mmHg and the mean bias per subject was −1.37 mmHg. For DBP, the SD and mean bias were 0.88 mmHg and 0.29 mmHg. For MAP, the SD and mean bias were 2.12 mmHg and 0.25 mmHg. The low mean variability and bias across each subject’s 45 PPG inputs indicates that measurements were consistent within subjects.

The within-subject standard deviation and bias represent measurement repeatability while the model residuals indicate accuracy of model predictions

To evaluate reproducibility, we assessed the stability of predictions under noisy conditions. As detailed in the Supplementary Materials (Section 3), intraclass correlation coefficients (ICC(2,1)) were computed across 30 perturbed test sets with Gaussian noise ranging from 5–30%. ICC values remained above the commonly accepted threshold for good reliability (0.75) for most noise levels, only dipping slightly under the most extreme conditions (≥25% noise). All ICC values were significantly greater than zero (p < 0.001), indicating that the model predictions are reproducible and robust to input variability.

Saliency maps were used to visualize which regions of the PPG signal and specific biometrics that most influenced the model’s prediction. Weight and heart rate appear to be the most impactful. All seven features’ saliency value were found statistically different from zero based on a one-sample t-test (p < 0.001), indicating they all exhibit statistically significant relevance to the model’s predictions. Additionally, one-sample t-tests on the cross-correlation between saliency maps and their corresponding PPG signals revealed statistically significant alignment with both finger and toe PPG measurement (p < 0.001; see Supplemental Section 3b). It suggests the model attends to physiologically important timepoints, such as the systolic peak, when predicting blood pressure.

During sleep, the autonomic nervous system undergoes a crucial shift as parasympathetic activity becomes dominant, leading to reduced heart rate, blood pressure, and PWV [24]. Our device captures these cardiovascular changes, revealing how PWV progressively decreases as sleep deepens, reflecting heightened parasympathetic tone and diminished sympathetic activity. PWV trends remain steady during REM sleep, based on typical stage progression patterns. As a non-invasive monitoring tool, it offers valuable insights into sleep architecture and cardiovascular health, positioning it as a promising advancement in wearable health technology. These results are consistent with prior findings that link increasing sleep depth to a gradual decline in arterial stiffness, as measured by PWV [25]. Analysis done while walking, biking, and typing can be found in Supplemental Section 2.

This performance aligns with the findings of several critical studies focused on the predictive value of PWV for cardiovascular outcomes and blood pressure estimation. For instance, the study by Garcia-Carretero et al. highlights the predictive significance of PWV in a high-risk population (obesity, hypertension, and hyperglycemia) without prevalent cardiovascular disease. Their LASSO Cox regression model identified PWV as a critical prognostic factor for cardiovascular events, with an adjusted hazard ratio of 1.199 (p < 0.001). The inclusion of PWV as a significant variable in their model emphasizes its importance in assessing arterial stiffness and cardiovascular risk, much like our model’s ability to achieve accurate blood pressure predictions. However, while their study focuses on cardiovascular outcomes, our model directly targets blood pressure prediction, achieving an impressive mean absolute error (MAE) of 4.160 mmHg for SBP and 2.311 mmHg for DBP, reflecting accuracy in clinical settings.

IV. Conclusion

The results presented in the Bland-Altman and scatter plots demonstrate that the device’s predictions for systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP) align closely with reference measurements from the Vicorder system. While a minor bias was observed, particularly a slight overestimation of SBP, most data remained within acceptable agreement limits. This suggests that the wearable device can track blood pressure fluctuations with sufficient precision for continuous monitoring, while offering improved usability and accessibility compared to traditional oscillometric and tonometric instruments.

The wearable BP monitor uses dual-site PPG sensors to derive PWV from PTT, rather than relying on ECG-based pulse arrival time, which does not account for the pre-ejection period and can introduce estimation errors. Furthermore, the system enables real-time tracking of cardiovascular dynamics across different states, including rest and sleep, offering potential value in early hypertension detection and long-term cardiovascular monitoring.

Despite these strengths, the study has several limitations. The participant population was limited to healthy, predominantly young adults, which may reduce generalizability to older individuals or those with cardiovascular conditions. The machine learning model used for blood pressure prediction was also trained on this limited population and may require adaptation for broader clinical use. While preprocessing methods were employed to manage noise and signal dropout, the system remains susceptible to PPG signal degradation during motion or due to inconsistent sensor placement.

Future work will focus on expanding the study population to include a broader demographic range, eliminating wired components through full Bluetooth integration, and enhancing sensor housing to reduce pressure-induced variability. Improvements to the signal processing pipeline will also aim to address intermittent signal loss and motion artifacts. These refinements will help establish a robust platform capable of delivering continuous, user-friendly blood pressure and PWV monitoring across diverse settings. As hypertension and cardiovascular disease continue to represent leading global health burdens, noninvasive wearable technologies such as the one presented in this study may provide meaningful contributions to the early identification and long-term management of cardiovascular risk.

Appendix

Fig. 1.

(a) Light transmission pathway of the pulse oximeter, showing LED emission, tissue penetration, and detection by the photodetector. AC and DC signal components are labeled. (b–c) Temporal delay between finger and toe PPG signals used to calculate PTT from systolic peaks. (d) Comparison between the wearable BP monitor and the reference Vicorder device, illustrating signal acquisition, PWV calculation, and BP prediction via a machine learning model.

Fig. 2.

(a–d) Preprocessing steps for finger and toe PPG signals. (a) Raw signals with spike artifacts. (b) Signals after applying a moving average filter (window size = 5). (c) Signals after a third-order Butterworth filter, with spikes removed. (d) Flipped PPG signal showing primary and secondary peaks and dicrotic notch; plotted points mark primary peaks used for PTT and PWV calculation.

Fig. 3.

Scatter (top) and Bland-Altman (bottom) plots comparing actual and predicted systolic (SBP), diastolic (DBP), and mean arterial pressure (MAP) in 25 participants. Scatter plots show predicted vs. actual values with a line of identity; Bland-Altman plots show bias (red dashed line) and 95% limits of agreement (gray dashed lines), indicating strong prediction accuracy with minimal bias.

Fig. 4.

(a) Correlation between Vicorder cf-PWV and device-measured ft-PWV (R² = 0.69). (b) Residual plot from ridge regression model predicting cf-PWV, showing random distribution around zero. (c) Predicted vs. actual cf-PWV with strong linearity (R² = 0.78). (d) PWV fluctuations during sleep stages across three nights; PWV decreases during deep sleep (N3) and increases during REM and wakefulness, showing consistent cardiovascular trends.

Fig. 5.

(a) Prototype of the wearable dual-PPG sensor device for continuous blood pressure and pulse wave velocity monitoring. (b) Finger placement on the sensor during signal acquisition. (c) Device housed in a white 3D-printed enclosure for improved usability and protection.