Calibration-Free Cuffless Blood Pressure Estimation Using Multimodal ECG-PPG Fusion on a Google Pixel Watch

Jathushan Kaetheeswaran; Boyi Ma; Ali Abedi; Shehroz S. Khan; Milad Lankarany

Source-linked AI summary

Calibration-Free Cuffless Blood Pressure Estimation Using Multimodal ECG-PPG Fusion on a Google Pixel Watch

Jathushan Kaetheeswaran, Boyi Ma, Ali Abedi, Shehroz S. Khan, Milad Lankarany

arXiv:2608.26325v1Aug 26, 2026cs.HC

TL;DR

Inadequate out-of-clinic BP monitoring motivates alternatives to inconvenient cuff-based devices. This study evaluates calibration-free BP estimation from synchronized Google Pixel Watch ECG and PPG signals in 40 participants, comparing algorithm approaches and demographic fusion. The proposed deep learning model performed best overall, while subgroup variability and dataset limitations constrain clinical reliability.

Problem
Accessible long-term BP monitoring remains limited because cuff-based devices are cumbersome, provide snapshot readings, and can be affected by body position, motion, and cuff dimensions.
Method
The study uses synchronized Google Pixel Watch ECG and PPG recordings with cuff BP measurements from 40 participants across semi-naturalistic conditions to evaluate calibration-free estimation models.
Results
The proposed deep learning model achieved the strongest overall performance, while demographic fusion improved generalizability to unseen individuals.
Takeaways & Limitations
Consumer-grade smartwatches show potential as accessible platforms for calibration-free cuffless BP estimation and longitudinal monitoring algorithms.
Takeaways & Limitations
Clinical reliability remains constrained by higher errors among obese participants and the need for larger, more diverse populations and additional sensing modalities.

Abstract

from arXiv · show

Inadequate blood pressure (BP) monitoring and management outside of clinical settings can worsen major cardiovascular risk factors such as hypertension. While cuff-based devices are commonly used for at-home monitoring, these devices can be inconvenient for daily use due to their sensitivity to body positions, upper-arm constrictions, and limited portability. A promising alternative is emerging in the form of consumer-grade smartwatches, where physiological signals related to cardiac activity can be used to estimate BP non-invasively and continuously across daily living conditions. In this work, we use data collected from a Google Pixel Watch in 40 participants to develop and compare several algorithm approaches for BP estimation. We found that our proposed deep learning model achieved the strongest overall performance, and that fusing smartwatch signals with demographic information improved model generalizability to unseen individuals. However, we also identified that model accuracy was not consistent across participant subgroups, with obese individuals yielding higher estimation errors than others. This study highlights the feasibility of consumer-grade smartwatches as accessible platforms for deploying robust BP estimation algorithms, though clinical reliability will require larger, more diverse populations and additional sensing modalities.

AUTHOR SUMMARY

The study evaluates Google Pixel Watch signals for non-invasive, continuous BP estimation and finds that deep learning with demographic fusion performed best overall. Accuracy varied across subgroups, with higher errors among obese participants.

Cuff-based devices are inconvenient for daily monitoring because of body-position sensitivity, upper-arm constrictions, and limited portability.
40 participants provided Google Pixel Watch data for comparing several BP-estimation algorithm approaches.
The proposed deep learning model achieved the strongest overall performance.
Fusing smartwatch signals with demographic information improved generalizability to unseen individuals.
Obese participants yielded higher estimation errors than other participant subgroups.

1 INTRODUCTION

The introduction motivates accessible longitudinal BP monitoring by describing limitations of cuffs and challenges in wrist-based estimation. It positions the Google Pixel Watch study as an evaluation of multimodal cuffless BP estimation under consumer-device constraints.

Hypertension is a leading modifiable cardiovascular-disease predictor, making high-quality longitudinal BP-monitoring tools important.
Cuff-based sphygmomanometers provide snapshot readings and can be inaccurate with body-position, motion, or cuff-dimension changes.
Wrist-worn devices combine ECG and PPG sensing with portability and flexibility for continuous, non-invasive home monitoring.
Commercial cuffless BP watches often show poor agreement with cuff measurements, partly because global validation standards are limited.
Wrist-worn BP estimation remains challenging because physiological confounders, poor contact, and motion artifacts affect signal-to-BP mappings.
The HEART-Watch dataset provides synchronized ECG, PPG, and cuff BP measurements from a commercial Google Pixel Watch under semi-naturalistic conditions.

2 METHODS

The methods use synchronized Google Pixel Watch ECG and PPG recordings with cuff BP measurements from 40 participants across semi-naturalistic physical states. Analyses assess protocol-related BP variation and smartwatch ECG agreement with chest ECG.

HEART-Watch Dataset: The HEART-Watch dataset contains synchronized wrist ECG and PPG recordings from 40 participants, with five snapshot BP readings per session and 200 independent observations.
HEART-Watch Dataset: Measurements span sitting, standing, and walking states under realistic wearing conditions and a diverse participant cohort.
HEART-Watch Dataset: The dataset includes self-reported healthy participants and lacks the ESH-IP2 minimum cohort size for medium- or high-BP ranges.
Protocol variation: SBP and DBP decreased after seated rest, increased when sitting changed to standing, and increased after walking.
Smartwatch ECG: Smartwatch ECG was evaluated against chest ECG using Bland-Altman analysis of QRS intervals across five measurement periods.
Smartwatch ECG: QRS biases ranged from 0.56 to 1.55 ms, and smartwatch ECG showed sufficient agreement for feature extraction such as R-peak detection.

2.3 Signal Preprocessing

ECG and PPG recordings were filtered and normalized to reduce baseline drift and preserve signal morphology. Saturated smartwatch ECG recordings were detected and replaced with chest ECG when invalid.

ECG preprocessing: ECG preprocessing removes baseline wander with a median filter and applies a zero-phase FIR bandpass filter from 0.67 to 45 Hz.
ECG preprocessing: ECG recordings are segmented into 2-second windows and checked for saturation artifacts.
ECG quality control: Recordings with more than 10% saturation artifacts are considered invalid and substituted with chest ECG.
PPG preprocessing: PPG signals are inverted, median-filtered, and bandpass-filtered from 0.5 to 8 Hz before subsequent processing.
Signal normalization: Min-max normalization is avoided to reduce the influence of sharp amplitude spikes from artifacts such as lead fall-offs.

2.4 Feature Extraction

The study segmented synchronized ECG and PPG recordings into nonoverlapping windows, extracted demographic and waveform features, and calculated PAT measures from cardiac and peripheral fiducial points. Signal delineation used physiological constraints to reduce artifacts and implausible detections.

Windowing: Recordings were segmented into 8, 10, 15, 20, 25, and 30 second nonoverlapping windows, each assigned the recording’s BP label.The windowing addressed the non-stationary nature of PPG signals.
Demographic Features: Participant features included age, sex, race, height, weight, numerical BMI, and categorical BMI classes from Underweight through Obese Class II.BMI categories were generated using the Canadian BMI Nomogram.
PPG Peak and Onset Selection: PPG peaks and onsets were detected with MSPTD, aligned greedily, and filtered using crest-time bounds of 70–500 ms.When multiple candidates existed, maximum-amplitude peaks and lowest-amplitude onsets were selected under the stated pairing logic.
PPG Morphology: Maximum-slope points were obtained from the largest first derivative between each PPG onset and peak, with values stored as rates of change per second.
PAT Features: PAT was computed from ECG R-peaks to PPG onsets, maximum slopes, or peaks, with physiological bounds applied to remove outliers.The extracted PAT variants correspond to PPG cycle peak, maximum slope, and valley fiducial choices.
ECG Features: ECG delineation identified QRS and RR intervals, retaining RR durations corresponding to heart rates between 35 and 130 BPM to reject likely artifact-driven outliers.The Peak Prominence ECG Delineator algorithm was used for ECG delineation.

2.5 Calibration-based vs Calibration-free Modelling

The paper distinguishes calibration-based models, which use data from the same subjects during training and validation, from calibration-free models evaluated on unseen subjects. It focuses on calibration-free generalization using leave-one-subject-out cross-validation.

Calibration-based Modelling: Calibration-based modelling includes the same subjects in training and validation, sometimes calibrating parameters with one or two subject-specific data points.
Calibration-free Modelling: Calibration-free modelling validates using features from subjects excluded from training, targeting performance on unseen individuals.The paper also calls this approach generalizable modelling.
Evaluation Protocol: Leave-one-subject-out cross-validation treats each subject as validation once and trains on the remaining subjects.This design avoids subject overlap between training and validation and is intended to evaluate realistic unseen-subject performance.

2.6 Mechanism-based Models

The mechanism-based approaches relate BP to PAT through vascular-elasticity and elastic-tube models. Their evaluation used leave-one-subject-out validation and error-based performance metrics.

Vascular-elasticity Models: Vascular-elasticity models use the Moens-Korteweg, Bramwell-Hill, and Hughes equations to relate PAT to vascular and blood mechanical properties.These properties include elastic modulus, radius, area, thickness, and density.
Elastic-tube Models: Elastic-tube models use conservation of mass and momentum to relate BP nonlinearly to PAT through calibration-specific parameters.
Evaluation: Mechanism-based models were evaluated with leave-one-subject-out cross-validation using mean error, mean absolute error, and standard deviation of signed errors.Windows were split by subject ID to avoid leakage between training and validation sets.
Model Parameters: The listed mathematical models contain subject-specific parameters a, b, c, d, and k determined through least-squares fitting.

2.7 Machine Learning Models

The machine-learning analysis evaluated supervised models using handcrafted ECG, PPG, and demographic features under leave-one-subject-out validation. It emphasized nonlinear regressors and compared kernel, tree, and ensemble approaches.

Motivation: Mechanism-driven models face constrained assumptions about peripheral vascular mechanics and limited calibration parameters, which can hinder generalization across dynamic cardiac conditions and diverse cohorts.
Features: The study derived 15 handcrafted features capturing waveform morphology and demographic characteristics associated with BP estimation.PPG crest time and maximum-slope amplitude describe beat-level morphology, while height, weight, and BMI describe body size and obesity.
Evaluation: Seventeen supervised machine-learning models were trained and validated across window sizes using leave-one-subject-out cross-validation and evaluated by ME, MAE, and STD.
Kernel-based Models: Kernel-based regressors were selected to capture nonlinear BP relationships that linear regression may not represent.The kernel trick maps low-dimensional inputs into a higher-dimensional space where a linear model corresponds to a nonlinear original-space solution.
Kernel-based Models: The evaluated kernel families included Gaussian process regression, support vector machines, and kernel approximation models.
Tree-based Models: Tree-based models included coarse, medium, and fine single regression trees, plus boosted and bagged tree ensembles.Leaf-size settings trade lower bias and higher variance against higher bias and lower variance; bagging averages outputs from trees trained on bootstrapped subsets.

2.8 Deep Learning Models

The proposed deep learning model uses wrist-based ECG and PPG time-series together with demographic information to predict systolic and diastolic blood pressure. Its architecture encodes signals and demographics separately, then uses demographic-wise modulation before prediction.

2.8 Deep Learning Models: The model accepts wrist-based ECG and PPG time-series plus demographic information to predict SBP and DBP.It was developed to improve on mechanism-driven and traditional machine-learning approaches requiring complex feature extraction.
2.8 Deep Learning Models: The signal encoder uses modality-specific dilated CNN processing, concatenation, pointwise fusion, normalization, and strided downsampling.The design expands temporal receptive fields while retaining separate ECG and PPG feature extraction before fusion.
2.8 Deep Learning Models: A lightweight MLP encodes demographic variables into a 64-feature embedding.Categorical variables are one-hot encoded, while continuous variables are normalized using training-split statistics.
2.8 Deep Learning Models: A FiLM-inspired layer linearly scales and shifts encoded signal features as a function of demographic embeddings.This demographic-wise modulation incorporates participant characteristics into intermediate signal representations.
2.8 Deep Learning Models: The modulated representations are pooled, flattened, and passed through a prediction MLP producing two outputs for SBP and DBP.The architecture contains a signal encoder, demographics encoder, demographic-wise modulation layer, and prediction layer.
2.8 Deep Learning Models: Signals were segmented into non-overlapping 8-, 10-, 15-, 20-, 25-, and 30-second windows, with LOSO cross-validation reserving each held-out subject for testing.The remaining 39 participants were used for training in each fold.

2.9 Model Evaluation

Model evaluation used subject-level leave-one-subject-out cross-validation, clinical and technical benchmarking standards, modality ablations, and BMI-based subgroup analysis. These procedures assessed generalization, input contributions, and performance differences between obese and non-obese participants.

2.9 Model Evaluation: The best model was compared with a training-population-mean baseline and a fixed 120/80 mmHg naïve estimator.Both estimators were evaluated using MAE, ME, and STD to test whether the model captured individualized BP variation.
2.9 Model Evaluation: Performance was assessed against AAMI requirements of ME within ±5 mmHg and STD within ±8 mmHg, alongside IEEE 1708 and BHS grading criteria.A device was recommended when both SBP and DBP passed AAMI and received Grade A or B under BHS.
2.9 Model Evaluation: Input ablations systematically removed ECG, PPG, or demographic features and retrained the model under identical LOSO conditions.Ablated variants were evaluated using ME, MAE, STD, AAMI, IEEE 1708, and BHS standards.
2.9 Model Evaluation: Participants were divided into obese and non-obese groups using a BMI threshold of 30 kg/m2 for subgroup evaluation.The analysis compared MAE, ME, and STD and assessed subgroup differences with a two-sided non-parametric test and Cliff’s δ effect size.

3 RESULTS

Across mechanism-based, traditional machine-learning, and deep-learning approaches, the proposed deep-learning model achieved the strongest overall BP estimation performance. Demographic and multimodal inputs supported generalization, while obese participants—especially for DBP—had higher errors.

3 RESULTS: PAT1 correlated positively with SBP (r = 0.30, p = 1.5 x 10-5), whereas PAT3 correlated negatively with SBP (r = -0.30, p = 1.5 x 10-5) and DBP (r = -0.22, p = 0.0014).PAT1’s DBP association was weaker and not statistically significant (r = 0.13, p = 0.066).
3 RESULTS: PAT2-based mechanism models yielded the lowest SBP MAE across window sizes, while PAT2 and PAT3 performed comparably for DBP.The best mechanism-driven result at 30 seconds was SBP: 11.88 ± 6.30 and DBP: 8.75 ± 3.98.
3 RESULTS: ML models achieved SBP MAE values of 11.08 to 11.23 mmHg but had substantially higher variability, with SBP STD ≈ 14 mmHg and DBP STD ≈ 11 mmHg.These STD values exceeded the AAMI criterion of STD ≤ 8 mmHg.
3 RESULTS: The proposed DL model achieved its best 8-second performance at SBP MAE: 9.13 mmHg and DBP MAE: 7.77 mmHg, with SBP STD ≈ 5.50 mmHg and DBP STD ≈ 3.80 mmHg.It had the lowest MAE and STD among the three modeling approaches.
3 RESULTS: The proposed DL model outperformed both baseline and naïve estimators, primarily through lower MAE while maintaining comparable STD.The findings indicate that predictions used demographic and biosignal inputs rather than collapsing to constant values.
3 RESULTS: The complete ECG + PPG + Demographics model achieved SBP ME: 0.06 mmHg, MAE: 9.13 mmHg, and STD: 5.60 mmHg.Removing demographics increased SBP MAE from 9.13 to 12.99 mmHg, while removing ECG increased it to 10.90 mmHg.
3 RESULTS: For DBP, removing demographics increased MAE from 7.77 to 9.96 mmHg and STD from 3.84 to 5.27 mmHg.Removing PPG slightly reduced MAE to 7.54 mmHg but lowered the BHS grade to D because variability increased.
3 RESULTS: Obese participants had higher errors than non-obese participants for SBP (10.72 vs 8.90 mmHg) and DBP (13.96 vs 6.88 mmHg).The DBP difference was statistically significant with a large effect size (p = 0.011, δ = 0.691), whereas the SBP difference was not significant.

4 DISCUSSION

The proposed deep-learning approach outperformed mechanism-based and traditional ML models for calibration-free BP estimation, with multimodal and demographic inputs improving generalization to unseen subjects. However, performance varied across subgroups and did not consistently meet clinical thresholds, so larger, more diverse, and realistic evaluations remain necessary.

Model performance: The proposed deep-learning model achieved the strongest generalizable performance compared with mechanism-based and traditional ML approaches.Evaluation used a leave-one-subject-out cross-validation framework for unseen-subject performance.
Multimodal fusion: Removing demographic inputs produced the largest increase in LOSO MAE, indicating that demographic encodings helped shift predictions toward unseen subjects’ physiological ranges.The ablation study also found ECG contributed more strongly than PPG to reducing both SBP and DBP error when combined with demographics.
Multimodal fusion: Fusion of ECG, PPG, and demographic inputs consistently outperformed single-modality models, supporting complementary information across sensing modalities.PPG-only models performed better for DBP than ECG-only models, while ECG contributed more strongly overall to error reduction when demographics were included.
Clinical evaluation: The model met AAMI requirements for SBP and DBP but achieved Grade D for SBP and Grade C for DBP under IEEE 1708 and BHS grading.Prediction errors therefore did not consistently remain within clinically acceptable thresholds for replacing cuff-based gold standards.
Subgroup performance: Obese participants had statistically significant higher DBP MAE than non-obese participants, while the corresponding SBP MAE increase was not statistically significant.The obese subgroup was small (n = 5), and the authors relate the disparity to obesity-associated vascular, PPG, and ECG-signal differences.
Limitations and future work: The study’s 40-participant dataset, temporally autocorrelated windows, limited demographic descriptors, and semi-naturalistic conditions constrain generalizability and population-level conclusions.The authors call for larger, more diverse cohorts, free-living data, additional sensing modalities, and advanced fusion strategies before reliable real-world deployment.