Update on Estimation of Blood Pressure and Pulse Wave Velocity Using Signals Obtained from Wearable Devices

In recent years, advances in very large-scale integration technology have spurred the flourishing expansion of wearable devices. This progress comes at a critical juncture as the global community grapples with the escalating challenges posed by an ageing population, which places strain on medical and social care systems. Wearable devices have emerged as a pivotal solution, facilitating pervasive healthcare and enabling continuous physiological assessment.¹ The projected compound annual growth rate of the global wearable technology market continues to rise. This trend has been particularly pronounced since the onset of the COVID-19 pandemic, as consumers seek wearable devices for purposes ranging from remote healthcare to health rehabilitation

Cardiovascular diseases (CVDs) remain a leading cause of mortality globally, contributing to a significant proportion of deaths each year.² Despite advances in medical knowledge and technology, combating CVDs still presents as a multifaceted task to public health systems worldwide. Holistic approaches, including prevention strategies, early diagnosis, effective management and ongoing research into new treatments and interventions, are involved. Collaborative efforts across disciplines and sectors are essential to reduce the prevalence and impact of CVDs. As a result, long-term health monitoring with wearable devices to support digital therapeutics for CVDs provides a promising solution and attracts much attention.

Several indexes and biomarkers are essential for CVD risk assessment. Among them, blood pressure (BP), pulse wave velocity (PWV) and heart rate variability (HRV) are essential indexes that can be continuously monitored through wearable devices. Undoubtedly, BP is regarded as one of the important indicators associated with CVDs while HRV reflects the underlying autonomic nervous system activity and cardiovascular health.³ Additionally, PWV, a measure of arterial stiffness, offers insights into vascular health. By leveraging wearable technology for monitoring these key indexes, proactive interventions and personalised management strategies can be adopted.

In this paper, we aim to investigate the recent progress in BP and PWV estimations using signals gathered from wearable devices. The diverse range of wearable devices designed for acquiring physiological signals is first illustrated. Pertinent concerns and key considerations are also addressed. Potential algorithms to enhance the estimation accuracy are discussed. Also, the expectations for future applications and the anticipated challenges for its developments will be described.

Physiological Signal Acquisition and Processing in Wearable Devices

Various wearable devices have been developed to capture physiological signals for health monitoring (Figure 1).^4–12 The integration of sensors into an armband, vest and belt provides physiological insight during physical activity while a smartwatch and ring can offer 24-hour continuous observation. ECG and photoplethysmography (PPG) are the most commonly gathered signals because both can capture physiological features. Besides, signals from accelerometer and oscillometer can also be used. Thus, the heart rate, HRV, BP, respiratory rate, peripheral oxygen saturation, sleep condition and activity index are calculated and estimated.

From an ECG, P waves, QRS complexes, and T waves are often observed, which reflect the contraction of the atria and ventricles during one heartbeat, as well as the subsequent depolarisation and repolarisation processes. PPG, obtained from a low-cost optical module, is also increasingly prevalent in wearable technology.¹³ The blood volume changes in the microvascular bed of tissue can be detected and thus the pulsatile nature of blood flow can be observed. In addition, pulse arrival time between ECG R peak and PPG fiducial point or pulse transit time (PTT) between multiple PPG signals has been shown to be highly correlated with arterial stiffness. Accordingly, it offers valuable information about the haemodynamic state, which supports applications in CVD risk assessment consisting of heart rate monitoring, BP estimation, arrhythmia detection, sleep apnoea monitoring and vascular age estimation.¹⁴⁻¹⁶

Since PPG morphological features embed the haemodynamic conditions, feature extraction is an important step in the signal processing flow for detection or estimation of indexes related with CVDs. However, it is worth noting that PPG morphology varies not only with the arterial compliance but also with the measurement sites due to tissue characteristics, blood flow dynamics, and vascular anatomy. Thus, several PPG types have been defined according to their morphology.^17,18 PPG signals from the fingertip and the earlobe usually exhibit sharp systolic peak following dicrotic notch and diastolic peak, especially for young people.¹⁹ PPG signals from the wrist tend to be smoother than those from the fingertip, with a less prominent dicrotic notch. For a finger PPG, when only a single peak is present, the diastolic peak is likely to vanish, with over a 90% probability, due to the steeply rising slope from the onset.²⁰ However, for a wrist PPG, the observed single peak could represent either the systolic peak or diastolic peak. Hence, the wrist PPG morphological features are usually less recognisable.¹⁹ Researchers need to consider these differences when interpreting PPG data and developing algorithms for the corresponding signals from specific wearable devices.

The first-order derivative and the second-order derivative PPG (FDPPG and SDPPG) waveforms are often computed to assist the evaluation of vessel stiffness and vascular age.^21,22 Pulse decomposition analysis has been proposed to reveal the latent forward component waves and reflected component waves.²³ Various numbers and shapes of component waves have been assumed to synthesise the PPG pulse.^24,25 It has been shown that although the morphologies of synchronised finger PPG pulse and wrist PPG pulse are different, the decomposed component waves have significantly high correlations.²⁰ The pulse decomposition analysis was adopted for PPG acquired from bracelet for BP estimation, which demonstrates the importance of latent property of PPG pulse to reveal the haemodynamic state.⁹

Health Monitoring Related to Cardiovascular Diseases

There is a growing interest in using wearable devices to offer real-time physiological status during medical interventions. Recent achievements in BP estimation and PWV/vascular age estimation are discussed.

Blood Pressure

The BP estimation from wearable devices using PPG has garnered significant attention. Cuffless BP estimation using PPG signals possesses advantages of continuous monitoring, portability, real-time feedback and longitudinal data collection.

The rapid growth and advance of machine learning and AI algorithms usher BP estimation into a new era.²⁶ The accuracy requirement under certain distribution of reference BP for clinical applications is specified in the validation protocol of the Association for Advancement of Medical Instrumentation.²⁷ The dataset comprises at least 255 BP measurements collected from 85 subjects in a relaxed and stationary state. The corresponding ratios of categories of systolic BP (SBP) ≤100 mmHg (diastolic BP [DBP] ≤60 mmHg) or SBP ≥160 mmHg (DBP ≥100 mmHg) must be >5% while the ratio of the category of SBP ≥140 mmHg (DBP ≥85 mmHg) must be >20%. Specifically, the precision of both the mean and SD of the error must fall below 5 and 8 mmHg, respectively.

On the other hand, due to the ‘white-coat’ effect, recent hypertension guidelines suggest multiple out-of-office BP measurements be used for hypertension management and diagnosis. Such BP measurements, obtained in environments outside the traditional clinical setting, capture variations that may not be apparent during clinic visits and can offer a fuller BP profile of a patient. Usually, BP exhibits a characteristic circadian pattern, with BP typically dipping during sleep and rising upon awakening. BP variability has been associated with cardiovascular conditions. Using wearable devices effectively supports the observation of BP variability and can assist the study of BP variability in clinical practice.²⁸

Various cuffless BP estimation approaches from wearable devices have been reported in the literature. The study protocols of cuffless BP estimations have been grouped into three categories: measurement in stationary condition; measurement in controlled-lab environment; and measurement in free-living setting.²⁹ Moreover, with the widespread adoption of machine learning or deep learning algorithms for BP estimation, there is the problem of data leakage. This occurs when the regression model obtains the information from the validation or test set in the training process, resulting in optimistic performance metrics. Therefore, it is imperative to adopt subject-split criterion to fairly demonstrate the generalisation ability of the regression model. Besides, PPG signals can be conveniently derived from non-contact optical devices, offering a more accessible and user-friendly option, whereas ECG requires electrodes placed on the body, enabling to capture certain cardiac parameters. Hence, studies using PPG-only approaches for BP estimation have the advantages of convenience and are suitable for continuous monitoring. Conversely, incorporating both PPG and ECG signals may allow for a more comprehensive assessment of cardiovascular function. Below, the trends of algorithmic development are investigated, and the performance of previous studies is examined in the context of these considerations.

Measurement in Stationary Condition

Previously, numerous studies employed various machine learning models for BP estimation, typically by first extracting information from physiological features and then employing these features in the regression techniques. The open data sets, including the MIMIC II Database and the University of Queensland Vital Signs Dataset were commonly used, while others used data gathered through their own experimental protocols.^30,31 The MIMIC II data set contains physiological signals and vital signs time series captured from patients in the intensive care unit. The University of Queensland Vital Signs Dataset is composed of signals recordings from 32 surgical patients who underwent anaesthesia. Various estimation algorithms have been adopted, including support vector regression, regression tree, adaptive boosting (AdaBoost), and random forest (RF).^32–34 Liu et al. developed BP estimation by using support vector regression and the new SDPPG features.³² Support vector regression finds a hyperplane in the feature space that is as close to the training data points as possible while minimising errors within a specific margin. Khalid et al. adopted the regression tree, a decision tree-based algorithm that partitions the feature space into subsets hierarchically and fits a smaller model to each partition.³³ Several machine learning algorithms have also been compared by Mousavi et al.,³⁴ such as decision tree regression, support vector regression, AdaBoost regression, and RF regression. The AdaBoost algorithm and the RF algorithm are members of the ensemble learning family. The former constructs a stronger learner by combining multiple weak learners through weighted aggregation while the latter produces its result by averaging the outputs of multiple decision trees generated by bootstrap aggregation and random feature selection. The AdaBoost algorithm outperformed the other three algorithms. Note that the waveform was used directly, and its properties could be exploited by the non-linear algorithms, resulting in mean error and SD error of −0.05 ± 8.90 mmHg for SBP.³⁴

Given the demonstrated efficacy of neural networks across various domains, artificial neural network and deep neural network were also used for BP estimation.^35,36 Priyanka et al. constructed two hidden layers for the artificial neural network and Song et al. inserted three hidden layers for their two-stage deep neural networks.^35,36 Besides the hand-crafted features, personal information like age, weight, sex, and height was also considered as the input.³⁶ Subsequently, an increasing number of studies have embraced deep learning techniques in this application of BP estimation. Furthermore, leveraging the remarkable capability of rapidly advancing deep learning algorithms to distinguish subtle differences, PPG waveform segments have often been used as direct inputs to the model.

Baek et al. adopted time and frequency waveforms as the inputs to a one-dimensional convolutional neural network with extraction and concentration block for identifying local features from periodic signals.³⁷ Long short-term memory (LSTM) with residual connection was designed by Li et al.,³⁸ which is known to excel in capturing long-range dependencies over sequences and thus can be applied to PPG sequences. Likewise, a streamlined variant of LSTM known as the gated recurrent unit was evaluated by Aguirre et al. for its ability to reveal long-term property of sequential data with fewer parameters and the demographic information was also included as input to the model.³⁹ The U-Net, distinguished by its unique U-shape architecture featuring skip connections between the encoder and the decoder, enabled effective feature extraction and has been verified by Yoshizawa et al. for its capability in BP estimation, where the University of California Irvine Dataset, extracted and processed from MIMIC II, was used.^40,41 The SBP and DBP estimation errors achieved −1.7 ± 10.9 mmHg and 0.95 ± 6.21 mmHg with only ECG input.⁴⁰

Since the attention-based technique in transformer has demonstrated its superiority in learning capability, its variants for BP estimation can also be seen.⁴²⁻⁴⁴ Kim et al. evaluated the residual U-Net with attention-based skip connections and outputs of U-Net were fed into the self-attention module with position encoding for the time sequence.⁴³ The transformer-based architecture with knowledge distillation, named KD-Informer, was also verified by Ma et al.⁴⁴ The learned knowledge is transferred from teacher Informer to lightweight student informer. Hand-crafted features were used as the inputs of prior knowledge. The mean and SD of the estimation error derived from the output arterial BP waveform were 0.03 ± 6.38 mmHg and 0.02 ± 4.49 mmHg for SBP and DBP, respectively, given the subject-split criterion.⁴⁴ Other techniques, such as gated recurrent unit with attention mechanism and knowledge distillation from multimodal teacher network to unimodal student network or convolutional neural network-LSTM cascaded network, have been realised.^45,46 Their performances are summarised in Table 1. Given advancements in deep learning, BP estimation accuracy can also be enhanced, especially for subject-wise estimation.

Measurement in Controlled-lab Condition

Controlled laboratory protocols for BP estimation often involve specific procedures and conditions designed to elicit standardised responses and minimise external influences. Some common protocol settings are:

• Exercise: participants may undergo controlled exercise protocols, such as treadmill walking or cycling, before BP measurements to realise the BP response to physical activity.
• Cold pressor test (CPT): participants may immerse a hand or forearm in ice-cold water for several minutes to stimulate a sympathetic nervous system response and transiently increase BP for assessing vascular reactivity and autonomic function.
• Posture changes: participants may be asked to change positions, such as transition from lying down to standing before BP measurements to evaluate postural influence on BP.

Owing to the lack of an open data set for this specific measurement, the number of studies on BP estimation under these conditions is not as extensive as those conducted under stationary conditions. Nevertheless, more researchers are gradually focusing on developing the tracking ability of estimated BP.

The participants were asked to perform exercise consisting of sequential cycling of three levels of intensity and BP estimation during alternative posture changes of sitting, supine and standing, investigated by Sun et al.⁴⁷ PPG signals from finger and forehead as well as ECG were collected. Features were extracted as the input to the multiple linear regressor. The estimation performance during posture changes was better than that during exercise. Liu et al. gathered multiple-wavelength PPG signals from the finger both before and after a 2-hour rehabilitation exercise.⁴⁸ The properties of responses generated by different wavelengths were employed to distinguish between arterial pulsation, capillary pulsation and motion artefact. Principal component analysis was used to separate signal sources, followed by estimation conducted via linear regression.

Paliakaitė et al. implemented the CPT, and participants immersed forearm in water.⁴⁹ One-minute recordings before and after 7°C water immersion were used for training, while another recording after the second 10°C water immersion was used for testing. Both finger PPG and wrist PPG were captured, but estimation accuracy derived from multiple linear regression using finger PPG surpassed those using wrist PPG. The moderate and heavy cycling exercises in predefined profiles were conducted by Landry et al. in the study protocol, while PPG and ECG were collected from the vest.⁵⁰ A nonlinear autoregressive model with exogenous inputs was established to estimate mean arterial pressure while periodic calibration was incorporated to improve the performance during the experiments. Sel et al. included hand-gripper exercise and CPT in the protocol and used bioimpedance signals for BP elevation manoeuvres.⁵¹ Physics-informed neural network was constructed and Taylor’s approximation for time-varying cardiovascular phenomena was performed to reduce the required training labels for personal model. The mean and SD of SBP and DBP estimation errors were 1.30 ± 7.60 mmHg and 0.6 ± 6.40 mmHg, respectively.

Table 2 lists the settings for these studies in controlled-lab environments. Although the number of subjects in the respective datasets is limited, recent studies have also evaluated advanced algorithms in personal models to track the BP variation caused by external stimulations.

Measurement in Free-living Condition

The use of out-of-office BP measurements is recommended by hypertension guidelines. The circadian cycle is evident in BP readings, typically following a pattern characterised by higher readings during wakefulness and lower readings during sleep. The BP dip phenomenon refers to the normal nighttime BP decrease that occurs during sleep. Typically, a reduction of 10–20% in both SBP and DBP compared to daytime levels is observed. The 24-hour out-of-office BP measurement offers valuable insights into an individual’s BP profile.²⁸

The importance of 24-hour BP estimation in a free-living context lies in its ability to provide a comprehensive assessment of BP variation throughout day and night. This continuous monitoring allows for detection of abnormal BP pattern, such as nocturnal hypertension or non-dipping status, which increases cardiovascular risk. However, in the study protocol of 24-hour BP estimation in free-living settings, one challenge is to ensure that each participant wears the BP monitoring device for the entire duration of the experimental period, which may span 24 hours or longer. Non-adherence can result in insufficient or inaccurate data, while potential measurement errors caused by motion artefacts and cuff position error may affect the reliability of the experiments. Another challenge is the synchronisation between the reference BP values and the long-recording signals, as using an external wireline trigger source can be inconvenient for participants, and wireless approaches may exceed power constraints.

Zheng et al. adopted resampling and filtering to remove abrupt transients of the time-varying PTT, acquired from PPG and ECG signals recorded by armband, and to generate the slowly varying curve.⁶ The estimation was mainly calculated by linear regression from PTT. Various algorithms, including linear regression, RF, multilayer perceptron network, and LSTM, have been verified by Radha et al. for 24-hour BP estimation using hand-crafted features including features from pulse decomposition analysis.²⁹ The same resampling and low-pass filtering techniques were also applied to the extracted features. The RF algorithm outperformed the remaining algorithms in this case. Yilmaz et al. evaluated nocturnal BP estimation using PPG segments in the window preceding ambulatory BP monitor.⁵² A total of 75 hand-crafted features, consisting of two component waves, systolic wave and residual wave which were decomposed from each PPG pulse, were fed into the RF algorithm for exploiting the non-linear relationship between the features and the reference BP. From their experimental results, 7-second window length was shown to be adequate. To reduce the calibration burden and the training overhead, Yang et al. adopted convolutional neural network with the model-agnostic meta-learning technique to initialise the model weights for few-shot learning tasks.⁵³ The synchronisation issue was addressed by detecting the cuff inflation time from recorded waveforms, allowing for the alignment of the BP measurements with the physiological signals. The performances of these studies are summarised in Table 3. Thus, using wearable devices to track 24-hour BP variability throughout the day offers a potential solution for understanding the circadian rhythms of BP.

Pulse Wave Velocity

PWV, typically measured by the propagation speed of pressure wave between a proximal and a distal point, is regarded as an important index which reflects the stiffness of the arterial walls. Since pressure waves travel more slowly in the elastic vessels according to haemodynamics, higher PWV values indicate stiffer arteries. Elevated PWV has been associated with either various cardiovascular conditions, such as hypertension or atherosclerosis, or risk of future cardiovascular events. Both carotid-femoral PWV (cfPWV) and brachial-ankle PWV (baPWV) offer valuable insights into vascular health conditions and help to guide treatment decisions and interventions for clinical use.

While BP measurements can be obtained using instruments either in hospital or at home, PWV measurements require specialised equipment and are typically performed by trained professionals in hospital settings. Furthermore, the interval between successive PWV measurements could extend to several months. Therefore, estimation of PWV by wearable devices is also recognised as an essential task for assessing the risk of CVDs. Like validation of BP measurement devices, validation of non-invasive PWV measurement devices is recommended recently.⁵⁴ A sample size of 85 subjects with relatively uniform distribution in four age groups (≤29, 30–49, 50–69 and ≥70 years) is suggested. The reference cfPWV values should be >5% for readings <6 m/s and >10 m/s. At least 20% of readings should be >8 m/s. The Artery Society also suggests an acceptable accuracy threshold of 1.5 m/s.⁵⁵

The features in SDPPG have been shown to be highly correlated with vascular age.⁵⁶ Jang et al. estimated baPWV by linear regression using features extracted from finger PPG and SDPPG.⁵⁷ Compensation using Fridericia equation with respect to heart rate has been demonstrated to be effective.⁵⁷ The performance was evaluated by error rate, which is defined as the estimation error divided by the reference value. The estimation was calculated based on the arrival time of the reflected wave with compensation achieved performance with average error rates of 8.36% for men and 9.52% for women, respectively. Finger PPG signals were collected from 123 subjects and the normalised feature of time interval between pulse peak to the next pulse onset with respect to pulse length was adopted for the estimation of baPWV by multiple linear regression together with age, which was also proposed by Jang et al.⁵⁸ Correlation coefficients of 0.899 were observed for men and 0.775 for women between the estimates and the references.

The time interval of the first zero crossing of FDPPG and the second valley of SDPPG from finger PPG were manually marked and regarded as pulse transit time for estimating cfPWV by multiple linear regression, support vector machine, and exponential Gaussian process regression.⁵⁹ From the comparison results provided by Gentilin et al., the mean and SD errors of the Gaussian process regression algorithm was 0.01 ± 0.75 m/s. Li et al. extracted multiple features from wrist PPG for estimating cfPWV by eXtreme Gradient Boosting algorithm.⁶⁰ The correlation coefficient between the predicted cfPWV and reference cfPWV achieved 0.927. Other machine learning algorithms such as Lasso regression and RF have been applied for estimation of cfPWV from finger PPG by Hellqvist et al.⁶¹ The root mean square error of 0.70 m/s for the better estimates is obtained from Lasso regression.

Compensated or normalised pulse transit time or addition, pulse arrival time extracted from PPG signals was often considered for linear regression model. Nevertheless, non-linear regression algorithms have also been verified for estimation of PWV and usually they outperform the linear regression algorithms.

Discussion and Future Perspectives

As shown in recent studies, the development trend of BP and PWV estimation from wearable devices is increasingly shifting towards the usage of advanced algorithms, transitioning from linear to non-linear models, and from traditional machine learning to deep learning approaches. Traditional linear algorithms have the advantages of simplicity and ease of implementation. As the need for more accurate estimations grows, non-linear algorithms gain traction. The advent of deep learning also enables the development of highly sophisticated models capable of learning intricacies of physiological phenomena.

Data availability is critical in the development and application of machine learning and deep learning models. However, acquiring abundant data of diverse categories in the medical field can still be challenging. Innovative techniques such as transfer learning, meta-learning, and zero-shot learning offer promising alternatives for learners given constrained training data.^53,62−65 Transfer learning allows fine-tuning a pre-trained model to a domain-specific task by leveraging its original knowledge, whereas meta-learning learns to adapt a model efficiently from limited data. Zero-shot learning extends the prediction or generalisation capability of a model without explicit training examples. The advancement of deep learning techniques also introduces opportunities for improvements and enhancements of the algorithms in the medical field.

Nevertheless, incorporating AI into wearable devices for biomedical applications still faces several limitations. First, the computational power and battery life of wearable devices are often limited, which constrains the complexity of AI algorithms that can be deployed. Second, the accuracy of AI models is likely to be influenced by the data collected from sensors, which may be affected by noise, motion artefacts, or user behaviours. Data security is also critical, and privacy must be maintained when training models or developing large collaborative models across research groups.

Despite the use of advanced algorithms, it is imperative not to underestimate issues related to data leakage and distribution of the testing data set in biomedical applications. Moreover, although neural networks can perceive the detailed and latent properties of the waveforms, hand-crafted features still help to stabilise model convergence.^44,51 The morphological features and various waveforms can be regarded as multimodal input signals. From the ablation study, incorporating morphological features alongside the PPG waveform reduced the SD error in SBP and DBP estimations by 0.1–2.4 mmHg, compared to using the PPG waveform alone.⁴⁴ On the other hand, age has been shown to be a key factor in estimating PWV. Including demographic features such as age and weight improves the SD error in PWV estimation.⁵⁹ Thus, domain knowledge also plays an important role in realising contextual understanding, model interpretability and clinical validation in personalised medicine or smart medicine. Figure 2 summarises the current development of wearable devices for BP and PWV estimations. Collaboration among domain experts, healthcare professionals and data scientists can advance healthcare through wearable devices.

Conclusion

Wearable devices bring opportunities for continuous and long-term health monitoring. Thus, there is a considerable focus among researchers on their applications in risk assessment for CVDs. Estimation of BP and PWV, essential indexes associated with CVD risks from wearable devices, becomes critical for the development and management of healthcare. We address the concerns, algorithms, recent progress and opportunities for improvements in this review article. Specifically, BP estimation approaches and results are categorised into three kinds of study protocols to highlight the diverse considerations and constraints. As innovations in this field continue to evolve, collaboration and interdisciplinary efforts will remain crucial for enhancing personalised smart healthcare.