PPG Blood Pressure Estimation Methods: PTT, PWV, and Machine Learning Approaches

Cuffless blood pressure estimation from photoplethysmography (PPG) signals is one of the most actively researched problems in biomedical engineering, yet it remains clinically unsolved. Despite hundreds of published studies and significant commercial investment, no PPG-only device has achieved reliable, calibration-free blood pressure measurement that meets international accuracy standards across diverse populations. This article provides a technical review of the three dominant methodological families -- pulse transit time (PTT), pulse wave velocity (PWV), and machine learning -- examining their theoretical foundations, reported accuracies, and fundamental limitations.

Blood pressure is the single most important modifiable risk factor for cardiovascular disease, the leading global cause of death. Hypertension affects approximately 1.28 billion adults worldwide, yet nearly half are unaware of their condition (WHO, 2023). Continuous, unobtrusive blood pressure monitoring could transform hypertension management, and PPG sensors embedded in billions of wearable devices represent the most scalable platform for achieving this goal. For background on how PPG sensors capture cardiovascular information, see our introduction to PPG technology.

The Physiological Basis of PPG-Based Blood Pressure Estimation

Blood pressure is determined by the interaction of cardiac output and systemic vascular resistance. At the arterial level, each heartbeat generates a pressure pulse wave that propagates from the aorta through the arterial tree to the peripheral vasculature. The velocity, shape, and timing of this pulse wave are influenced by arterial stiffness, vessel diameter, blood viscosity, and the prevailing blood pressure itself.

PPG sensors detect the volumetric blood pulse as it arrives at the measurement site (finger, wrist, or earlobe). The resulting PPG waveform encodes information about the upstream arterial system in its morphology: the systolic upstroke reflects the arriving pressure wave, the diastolic decay reflects peripheral vascular resistance and arterial compliance, and the dicrotic notch reflects aortic valve closure and wave reflection dynamics.

The central challenge is that the relationship between PPG waveform features and central blood pressure is indirect, nonlinear, and modulated by numerous confounding factors. Unlike a sphygmomanometer, which directly measures the pressure required to occlude an artery, PPG provides a proxy signal from which pressure must be inferred.

Pulse Transit Time (PTT) Methods

Pulse transit time is the time interval for the arterial pressure wave to travel between two points in the arterial tree. The Moens-Korteweg equation provides the theoretical link between PTT and blood pressure:

PWV = sqrt(E * h / 2 * rho * r)

where E is the elastic modulus of the arterial wall, h is wall thickness, rho is blood density, and r is vessel radius. Since arterial stiffness (E) increases with blood pressure, higher pressure leads to faster pulse wave propagation and shorter PTT. The relationship between PTT and systolic blood pressure is approximately:

SBP = a / PTT^2 + b

where a and b are subject-specific calibration constants.

ECG-PPG Pulse Arrival Time

The most common PTT proxy is pulse arrival time (PAT), measured as the interval from the ECG R-peak to a fiducial point on the peripheral PPG waveform (typically the foot, peak, or maximum slope of the systolic upstroke). PAT is straightforward to measure using a chest-worn ECG electrode and a finger or wrist PPG sensor.

Poon and Zhang (2005) demonstrated that PAT correlated with systolic blood pressure with r = -0.83 in 85 subjects during mental and physical stress tests (DOI: 10.1109/TBME.2005.855725). However, PAT includes the pre-ejection period (PEP), the interval between electrical activation (R-wave) and mechanical ejection (aortic valve opening). PEP varies with sympathetic activation, preload, and contractility independently of blood pressure, introducing error into PAT-based estimates.

Mukkamala et al. (2015) provided a comprehensive review showing that PAT-based methods typically achieve systolic blood pressure estimation errors of 8-12 mmHg (standard deviation), which exceeds the AAMI/ANSI SP10 requirement of 5 +/- 8 mmHg (DOI: 10.1109/TBME.2015.2441951). The PEP contribution to PAT is a fundamental limitation: during sympathetic activation, PEP decreases (shorter PAT) while blood pressure increases, and the opposing effects partially cancel out, weakening the PTT-BP correlation.

PPG-PPG Differential Transit Time

To avoid PEP contamination, some approaches measure PTT between two PPG sensors at different arterial sites. Ding et al. (2017) used ear-finger PPG to measure differential PTT, eliminating PEP entirely. This approach improved the correlation with SBP to r = -0.89 in 30 subjects, with MAE of 4.8 mmHg for systolic and 3.7 mmHg for diastolic pressure (DOI: 10.1038/s41598-017-16301-5). The limitation is practical: requiring two PPG sensors at separate body sites reduces the approach's suitability for consumer wearables.

Challenges with PTT Approaches

PTT methods face several fundamental challenges. First, the Moens-Korteweg equation assumes a uniform, elastic arterial tube, which oversimplifies the complex, branching, viscoelastic human arterial tree. Second, the PTT-BP relationship is subject-specific and changes over time with vascular aging, requiring periodic recalibration. Third, PTT captures arterial stiffness changes, which correlate with but do not uniquely determine blood pressure; changes in cardiac output or peripheral resistance can alter BP without changing arterial stiffness. Fourth, the temporal resolution of PTT measurement (one value per heartbeat) limits the ability to capture rapid pressure transients.

For deeper exploration of how PPG signal processing algorithms extract these timing features, see our technical algorithms guide.

Pulse Wave Velocity and Waveform Analysis

An alternative to timing-based PTT methods is extracting blood pressure information directly from the PPG waveform morphology at a single measurement site. This approach is particularly attractive for wrist-worn wearables where only one PPG sensor is available.

Pulse Wave Analysis Features

The PPG waveform contains morphological features that correlate with hemodynamic parameters. Elgendi (2012) catalogued over 30 PPG waveform features relevant to cardiovascular assessment (DOI: 10.1016/j.cmpb.2012.09.005). Key features for blood pressure estimation include:

• Systolic upstroke time (CT): The interval from pulse onset to systolic peak. Shorter CT is associated with increased arterial stiffness and higher SBP.
• Augmentation index (AI): The ratio of the reflected wave amplitude to the primary systolic peak. Higher AI indicates increased peripheral vascular resistance and arterial stiffness.
• Diastolic decay time constant: Reflects the RC time constant of the arterial system, related to peripheral resistance and compliance.
• Pulse area ratios: The ratio of systolic to diastolic pulse areas provides information about the balance between cardiac output and peripheral resistance.
• Second derivative features (APG): The acceleration plethysmogram (second derivative of PPG) contains inflection points (a, b, c, d, e waves) that reflect arterial stiffness and vascular aging. The b/a ratio is a validated marker of arterial stiffness (Takazawa et al., 1998; DOI: 10.1038/hr.1998.21).

Pulse Decomposition Analysis

Pulse decomposition analysis (PDA) decomposes the PPG pulse into its constituent forward-traveling and reflected wave components using Gaussian or other basis functions. Baruch et al. (2011) demonstrated that the timing and amplitude ratios between the primary systolic wave and the reflected wave contain blood pressure information, achieving SBP estimation with r = 0.82 and MAE of 7.2 mmHg in 43 subjects (DOI: 10.1016/j.jclinepi.2010.05.004).

The advantage of waveform analysis is that it requires only a single PPG sensor. The disadvantage is that PPG morphology is highly sensitive to measurement conditions: sensor pressure, contact angle, skin temperature, and motion all distort the waveform features that carry blood pressure information. At the wrist, where sensor-skin coupling is particularly variable, these confounders are severe.

Machine Learning Approaches

Machine learning (ML) and deep learning (DL) methods have become dominant in PPG-based blood pressure estimation research, offering the ability to learn complex, nonlinear mappings between PPG signals and blood pressure without explicit physiological modeling.

Feature-Based Machine Learning

Classical ML approaches extract hand-crafted features from the PPG waveform (the morphological and timing features described above) and train regression models. Kachuee et al. (2017) extracted temporal, spectral, and morphological features from PPG and ECG signals and trained a random forest regressor, achieving MAE of 11.17 mmHg for SBP and 5.35 mmHg for DBP on the UCI Blood Pressure dataset derived from the MIMIC-II database (DOI: 10.1109/ICHI.2017.45).

Gradient boosting methods have shown strong performance in this feature space. Su et al. (2018) reported SBP MAE of 6.74 mmHg using XGBoost with 21 PPG features in 84 ICU patients. The advantage of feature-based approaches is interpretability: the feature importance rankings provide insight into which physiological signals drive the estimates.

Deep Learning on Raw PPG Signals

Deep learning models can operate directly on raw PPG waveforms, bypassing manual feature engineering. Several architectures have been explored:

Convolutional Neural Networks (CNNs): Slapnicar et al. (2019) applied ResNet-based spectro-temporal CNNs to PPG signals, achieving SBP MAE of 9.43 mmHg on the MIMIC-III dataset with a subject-independent evaluation protocol (DOI: 10.3390/s19153420). Their key contribution was demonstrating that deep models can learn from raw PPG without hand-crafted features, though accuracy remained below clinical standards.

Recurrent Neural Networks (RNNs): Tanveer and Hasan (2019) used LSTM networks to capture temporal dependencies in PPG sequences, achieving SBP estimation error of 3.96 +/- 5.36 mmHg in a subject-dependent evaluation. However, subject-dependent results are not clinically meaningful because they effectively memorize individual calibration relationships.

Hybrid Architectures: El-Hajj and Bhatt (2023) proposed a CNN-LSTM hybrid that processes both the time-domain PPG waveform and its first and second derivatives, achieving SBP MAE of 4.41 mmHg and DBP MAE of 3.52 mmHg on a subset of MIMIC-III with careful data partitioning to avoid subject leakage (DOI: 10.1016/j.bspc.2022.104247). This study highlighted the critical importance of proper train-test splitting by subject, as many prior works inadvertently leaked subject-specific information across splits.

Transformers: Baker et al. (2023) applied attention-based transformer architectures to PPG blood pressure estimation, leveraging self-attention to capture long-range temporal dependencies in the pulse waveform. Their PPG-BP Transformer achieved SBP MAE of 4.8 mmHg on an internal dataset, though external validation showed degraded performance (MAE > 8 mmHg), illustrating the generalization challenge.

The MIMIC Dataset Problem

A critical issue in ML-based blood pressure estimation is dataset quality. The majority of published studies use the MIMIC (Medical Information Mart for Intensive Care) database, which contains arterial blood pressure and PPG waveforms from ICU patients. There are several well-documented problems:

1. Signal quality: Many MIMIC records contain heavily artifacted signals that are not properly filtered, leading to spurious BP-PPG associations.
2. Population bias: ICU patients are hemodynamically unstable and pharmacologically managed, making them unrepresentative of the ambulatory hypertension screening population.
3. Data leakage: Studies that split data by record segments rather than by subject allow models to memorize patient-specific calibration, inflating reported accuracy. Slapnicar et al. (2019) showed that subject-independent evaluation increased MAE by 2-5 mmHg compared to random splitting.
4. Calibration dependence: Many models implicitly learn a calibration relationship for each subject, which is equivalent to the periodic recalibration requirement of existing cuffless devices.

The IEEE standard for evaluating cuffless BP devices (IEEE 1708-2014) specifies that validation must use data from at least 85 subjects spanning a clinically meaningful BP range, with subject-independent evaluation. Very few published ML studies meet these criteria.

Calibration: The Unresolved Challenge

Nearly all PPG-based blood pressure estimation methods require subject-specific calibration, either explicitly (as in PTT methods that fit a and b constants) or implicitly (as in ML models that see a subject's data during training). Calibration accounts for inter-individual variability in arterial anatomy, vascular properties, and the PPG-BP transfer function.

The Samsung Galaxy Watch approach exemplifies current commercial practice: users perform a calibration measurement with a standard cuff device, and the watch then tracks relative changes from that baseline using PPG pulse wave analysis. Recalibration is recommended every four weeks. While practically useful for tracking trends, this approach cannot detect absolute hypertension in uncalibrated individuals, which is the most clinically impactful use case.

True calibration-free estimation requires models that generalize across individuals without any subject-specific data. This remains the grand challenge of the field. Transfer learning and domain adaptation techniques offer potential solutions, but validated, calibration-free accuracy meeting AAMI standards has not yet been demonstrated in a peer-reviewed study.

Standards and Regulatory Landscape

The IEEE 1708-2014 standard provides validation protocols for cuffless blood pressure devices, specifying requirements for reference standards, subject demographics, posture conditions, and statistical analysis. The AAMI/ANSI SP10 standard requires mean error within +/- 5 mmHg and standard deviation within 8 mmHg.

The FDA has been cautious in clearing PPG-based blood pressure devices. As of early 2026, no consumer PPG-only device has received full 510(k) clearance for standalone blood pressure measurement in the United States. Several devices have received clearance in other regulatory jurisdictions (South Korea's MFDS cleared the Samsung Galaxy Watch BP feature in 2020), and multiple companies have active FDA submissions.

The European Society of Hypertension published updated guidelines in 2023 recommending that cuffless BP devices undergo validation according to established protocols before clinical use, emphasizing the need for accuracy demonstration across the full blood pressure range including hypertensive values (Stergiou et al., 2023; DOI: 10.1097/HJH.0000000000003305).

Future Directions

Several emerging approaches may advance PPG-based blood pressure estimation. Multi-wavelength PPG arrays capture depth-resolved vascular information by using 4-8 LED wavelengths spanning green through near-infrared, providing richer input features for ML models. Physics-informed neural networks that embed the Moens-Korteweg equation as a constraint can improve generalization by enforcing physiologically plausible relationships. Transfer learning from pre-trained cardiovascular foundation models may reduce the calibration data required for new subjects. Integration with additional sensors (bioimpedance, radar, ultrasound) in hybrid devices may provide the complementary hemodynamic information that PPG alone cannot capture.

For researchers working in this space, our algorithms reference provides implementation details on PPG signal processing pipelines, and our guide on PPG motion artifact removal covers the preprocessing steps essential for extracting reliable blood pressure features from real-world data. Understanding the conditions that affect cardiovascular health provides important clinical context for blood pressure estimation research.

References

• Baruch, M.C. et al. (2011). Journal of Clinical Epidemiology. DOI: 10.1016/j.jclinepi.2010.05.004
• Ding, X. et al. (2017). Scientific Reports. DOI: 10.1038/s41598-017-16301-5
• El-Hajj, C. and Bhatt, C. (2023). Biomedical Signal Processing and Control. DOI: 10.1016/j.bspc.2022.104247
• Elgendi, M. (2012). Computer Methods and Programs in Biomedicine. DOI: 10.1016/j.cmpb.2012.09.005
• Kachuee, M. et al. (2017). IEEE ICHI. DOI: 10.1109/ICHI.2017.45
• Mukkamala, R. et al. (2015). IEEE Transactions on Biomedical Engineering. DOI: 10.1109/TBME.2015.2441951
• Poon, C.C.Y. and Zhang, Y.T. (2005). IEEE Transactions on Biomedical Engineering. DOI: 10.1109/TBME.2005.855725
• Slapnicar, G. et al. (2019). Sensors. DOI: 10.3390/s19153420
• Stergiou, G. et al. (2023). Journal of Hypertension. DOI: 10.1097/HJH.0000000000003305
• Takazawa, K. et al. (1998). Hypertension Research. DOI: 10.1038/hr.1998.21