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
TL;DR
Accessible long-term BP monitoring solutions remain limited. This study tested cuffless BP estimation from synchronized smartwatch ECG and PPG, finding the proposed deep-learning model strongest overall and demographic inputs important to performance on unseen subjects.
Problem
Accessible long-term blood pressure monitoring solutions remain limited.
Source: §1 · p. 2Method
The study evaluated cuffless BP estimation from synchronized wrist ECG and PPG recorded by a Google Pixel Watch during semi-naturalistic conditions.
Source: §1 · p. 3Results
The proposed deep-learning architecture achieved the strongest overall LOSO performance, and removing demographic inputs produced the largest increase in LOSO MAE.
Source: §4 · p. 27Takeaways & Limitations
Consumer-grade smartwatches show potential for longitudinal home BP monitoring, but further improvements are needed before reliable real-world deployment.
Source: §4 · p. 28The 40-participant dataset contained only 200 independent recordings, whose autocorrelated window segments may have constrained learning generalizable physiological patterns.
Source: §4 · p. 28
Abstract
from arXiv · showhide
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 · 1 INTRODUCTION
This introduction motivates accessible long-term blood-pressure monitoring and presents a study of synchronized ECG and PPG from a consumer smartwatch, evaluated with cuff measurements under semi-naturalistic conditions.
1 INTRODUCTION
- Long-term BP trends predict cardiovascular outcomes, and timely cardiovascular management depends on effective longitudinal monitoring.Hypertension is a leading modifiable predictor of cardiovascular disease, and a 1996–2021 study found hypertensive children were more than twice as likely to experience major cardiac events in adulthood.Source: §1 · p. 2
- Infrequent, fixed-setting clinical measurements can miss masked or white-coat hypertension, while accessible long-term home monitoring remains limited.Continuous self-monitoring paired with telemedicine has been associated with improved health outcomes.Source: §1 · p. 2
- Cuff-based monitors are cumbersome for daily use, provide snapshot readings, and can be inaccurate with body position, motion, or improper cuff dimensions [61, 72, 51].These conditions are common in uncontrolled home environments.Source: §1 · p. 2
- Wrist-worn ECG and PPG devices offer portable, non-invasive continuous monitoring, but available cuffless watches can agree poorly with cuff measurements.An Apple Watch received FDA clearance for an AI-based hypertension notification feature using PPG.Source: §1 · pp. 2–3
- Wrist-based BP estimation is challenged by physiological confounders and daily-use signal artifacts, while prior studies often rely on custom devices that may not reflect consumer-watch performance.Many works use PPG alone or additional instrumentation; custom and dedicated devices may have higher signal quality, whereas commercial-watch design and processing constraints can affect biosignal quality and estimation performance.Source: §1 · p. 3
- The study evaluates consumer-grade Google Pixel Watch signals for BP estimation and introduces HEART-Watch, described as the first dataset combining smartwatch ECG and PPG with cuff BP measurements.It assesses cuffless estimation under consumer-device sensing constraints during semi-naturalistic conditions.Source: §1 · p. 3
2 METHODS · 2.1 HEART-Watch Dataset · 2.2 Smartwatch ECG
The methods use 200 BP observations from 40 participants in a semi-naturalistic smartwatch dataset and assess smartwatch ECG against chest ECG. BP changed significantly across protocol transitions, while ECG QRS agreement was sufficient for feature extraction; the healthy cohort lacked adequate medium- and high-BP representation.
2.1 HEART-Watch Dataset
- The HEART-Watch dataset contains synchronized wrist ECG and PPG from a 2022 Google Pixel Watch, with five snapshot BP readings per session for 40 participants, totaling 200 observations.Measurements were collected across sitting, standing, and walking under semi-naturalistic wearing conditions [25].Source: §2.1 · p. 5
- The cohort met ESH-IP2 participant and sex criteria, but its self-reported healthy participants left medium- and high-BP ranges below the protocol’s minimum cohort size.ESH-IP2 specifies at least 10 males and 10 females in a cohort of at least 33.Source: §2.1 · pp. 5–6
- After Bonferroni correction, SBP and DBP decreased after seated rest, increased on standing, and increased after four minutes of walking.Wilcoxon signed-rank tests assessed sequential changes across protocol measures.Source: §2.1 · p. 7, Figure 2 · p. 8
2.2 Smartwatch ECG
- Participants recorded Lead I smartwatch ECG by placing the right index finger on the watch crown during BP measurements; concurrent chest ECG was collected, but the study focused on smartwatch ECG.The focus was intended to reflect wearable monitoring situations where only the smartwatch is available; agreement was assessed using QRS interval Bland–Altman analysis.Source: §2.2 · p. 8
- 0.56–1.55 ms mean QRS differences across five periods supported sufficient smartwatch–chest ECG agreement for feature extraction, including R-peak detection.The first-period bias was statistically significant (p < 0.05), but small relative to typical QRS intervals of 80–100 ms.Source: §2.2 · pp. 8–9, Figure 3 · p. 9
2.3 Signal Preprocessing
ECG and PPG recordings were filtered and normalized to preserve waveform features, while saturated watch ECG recordings were screened and, when invalid, replaced with chest ECG recordings.
2.3 Signal Preprocessing
- ECG signals received median baseline-wander filtering, then zero-phase 375th-order FIR bandpass filtering from 0.67 to 45 Hz with a Hamming window.Baseline wander was described as low-frequency drift between 0.8 and 1.25 Hz, associated with electrode shifts, breathing, and other movements.Source: §2.3 · p. 9
- Watch ECG recordings were segmented into 2-second windows for saturation checks; recordings with saturation in more than 10% were invalidated and substituted with chest ECG.Only p03 measure 5 and p20 measure 1 were identified as completely invalid and substituted.Source: §2.3 · p. 9
- PPG signals were inverted to restore pressure-like waveform morphology, median-filtered, and zero-phase Butterworth-filtered from 0.5 to 8 Hz for peak detection.Source: §2.3 · p. 10
- Both signals were trimmed by 2 seconds at the start and Z-score-normalized in 5-second sliding windows with 50% overlap; min-max normalization was avoided to limit sharp artifact spikes.The trimming removed initial recording artifacts, and the normalization was selected to preserve waveform shapes and avoid potential lead-fall-off spikes.Source: §2.3 · p. 10
2.4 Feature Extraction
The feature-extraction pipeline segmented synchronized ECG and PPG into nonoverlapping windows and derived demographic, pulse-timing, and ECG features. It used physiological constraints to select waveform fiducials and filter implausible timing measurements.
2.4 Feature Extraction
- 8-, 10-, 15-, 20-, 25-, and 30-second nonoverlapping windows were assigned the recording’s BP label for feature extraction.Windowing addressed the non-stationary nature of PPG signals.Source: §2.4 · p. 10
2.4.1 Demographic Features
- Participant age, sex, race, height, weight, and BMI were used as numerical and categorical features.BMI categories from the Canadian BMI Nomogram were Underweight, Normal, Overweight, Obese Class I, and Obese Class II.Source: §2.4.1 · pp. 10–11
2.4.2 Peak and Onset Detection
- ECG R-peaks were delineated with the Peak Prominence ECG Delineator, while MSPTD identified PPG systolic peaks and onsets.PPG detection was challenging because motion artifacts, ambient lighting, and sensor contact affect the waveform; diastolic features were omitted because wrist PPG details vary with posture and contact pressure.Source: §2.4.2 · p. 11
2.4.3 PPG Peak and Onset Selection
- A greedy alignment paired PPG onsets and peaks with crest times of 70–500 ms.When multiple candidates occurred, the maximum-amplitude peak and, where needed, the lowest-amplitude onset were selected; maximum slope was calculated from the largest first derivative between onset and peak.Source: §2.4.3 · p. 12
2.4.4 Pulse Wave Velocity
- PAT-based estimation was used with synchronized wrist ECG and PPG; PAT spans R-peaks to PPG onsets, maximum slopes, or peaks and includes PEP.The paper relates decreasing PAT to higher BP and cites evidence that PAT-based estimation may better tolerate physical stress; synchronized wrist ECG has been shown to improve PPG-based techniques. PTT instead measures delay between points on the same arterial branch.Source: §2.4.4 · p. 12
2.4.5 Pulse Arrival Times
- Valid PAT3 and PAT2 values were bounded at 70–400 ms, while PAT1 values were bounded at 100–650 ms.PAT was measured from each ECG R-peak to the closest PPG onset, peak, or maximum-slope point; the wider PAT1 bounds accounted for demographic differences in arterial stiffness.Source: §2.4.5 · p. 13
2.4.6 ECG Time-domain Features
- ECG features included mean QRS duration and mean RR interval, with RR intervals restricted to durations corresponding to 35–130 BPM.The RR filter guarded against artifact-induced amplitude spikes; the cuff’s reported heart-rate range was 50–115 BPM.Source: §2.4.6 · p. 13
2.5 Calibration-based vs Calibration-free Modelling
The study focuses on calibration-free BP modelling, which tests generalization to subjects unseen during training, rather than calibration-based modelling that uses data from the same subjects in training and validation. It evaluates models with leave-one-subject-out cross-validation (LOSO CV).
2.5 Calibration-based vs Calibration-free Modelling
- Calibration-based modelling uses data from the same subjects for training and validation, whereas calibration-free modelling validates on subjects excluded from training.Source: §2.5 · p. 13
- Because the dataset was small with few samples per subject, the study focused on calibration-free modelling to assess smartwatch BP estimation on unseen subjects.Source: §2.5 · p. 13
- LOSO CV validated each subject once using models trained on the remaining subjects, providing a realistic unseen-subject evaluation compared with holdout splits that can alter accuracy.Source: §2.5 · p. 13
2.6 Mechanism-based Models
Mechanism-based models estimate BP from PAT using vascular elasticity or elastic-tube relationships. They were evaluated with subject-separated cross-validation using signed-error metrics.
2.6 Mechanism-based Models
- Vascular-elasticity models use Moens-Korteweg, Bramwell-Hill, and Hughes equations to relate PAT to vascular and blood properties.These properties include elastic modulus, radius, area, thickness, and density.Source: §2.6 · p. 14
- Elastic-tube models use conservation of mass and momentum to nonlinearly relate BP to PAT through calibration-specific parameters.The listed mathematical models use subject-specific parameters determined through least-squares fitting.Source: §2.6 · p. 14, Table 3 · p. 14
- Mechanism-based models were evaluated using LOSO cross-validation split by subject ID, with mean error, mean absolute error, and signed-error standard deviation computed across window sizes and model types.For models incorporating heart rate, ECG-derived HR was used rather than BP cuff readings.Source: §2.6 · p. 14
2.7 Machine Learning Models
The section motivates flexible, feature-based machine learning because mechanism-driven BP models have constrained assumptions and may not generalize across cohort sizes. It evaluates supervised models using leave-one-subject-out cross-validation.
2.7 Machine Learning Models
- Mechanism-driven models face constrained assumptions about changing peripheral vasculature and may lack the flexibility to generalize across cohort sizes.The authors therefore considered traditional machine learning methods that incorporate handcrafted features.Source: §2.7 · p. 15
- The 15 handcrafted features combined beat-level PPG morphology, including crest time and maximum-slope amplitude, with height, weight, and BMI for BP estimation.The features were intended to capture waveform morphology and demographic characteristics associated with cardiovascular health, including blood viscosity and peripheral vascular resistance.Source: §2.7 · p. 15
- The study evaluated 17 supervised models for systolic and diastolic BP across window sizes using LOSO cross-validation and ME, MAE, and STD on unseen data.Source: §2.7 · p. 15
2.7.1 Kernel-based Models
- Kernel-based evaluation included GPR, SVM/SVR, and kernel approximation models, using mappings to high-dimensional spaces to represent nonlinear feature relationships.Kernel approximation generates nonlinear features and fits a linear model rather than using an exact kernel function.Source: §2.7.1 · pp. 15–16
2.7.2 Tree-based models
- Tree evaluation covered coarse, medium, and fine single trees with leaf-size bias–variance tradeoffs, LSBoost/AdaBoost residual fitting, and bagged bootstrap-trained trees averaged into one output.Source: §2.7.2 · p. 16
2.8 Deep Learning Models
The authors introduce a deep learning model that predicts SBP and DBP from wrist ECG and PPG time series plus demographic information, aiming to avoid brittle feature engineering for variable smartwatch signals.
2.8 Deep Learning Models
- The model uses wrist ECG and PPG time series with demographic information to predict SBP and DBP, avoiding reliance on hand-engineered features for variable smartwatch signals.The motivation is that wrist-signal morphology can vary with sensor contact, skin elasticity, and skin pigment, making idealized feature-extraction assumptions unreliable.Source: §2.8 · p. 17
2.8.1 Model Architecture
- Separate dilated CNN branches extract ECG- and PPG-specific features before fusion and cross-modality learning, while a lightweight MLP encodes demographic variables.The signal encoder uses four output channels per modality and downsamples fused features by 50%.Source: §2.8.1 · p. 17
- A demographic-conditioned FiLM layer scales and shifts signal features before pooled representations enter an MLP that predicts SBP and DBP.Demographic embeddings generate the FiLM scale and shift parameters; the prediction layer maps representations to two outputs.Source: §2.8.1 · pp. 17–18
2.8.2 Data Preparation
- Categorical sex, race, and BMI class were one-hot encoded, continuous demographics were normalized using training-split statistics, and signals were segmented into 8- to 30-second windows.Numerical BMI values were excluded from the demographic feature set based on preliminary experimentation.Source: §2.8.2 · p. 18
2.8.3 Training Settings
- Training used leave-one-subject-out cross-validation, with 39 participants for training and the held-out participant for testing; L1 loss directly targeted MAE.Each fold trained for 20 epochs with AdamW, and the checkpoint with the lowest combined SBP and DBP training MAE was evaluated on the held-out subject.Source: §2.8.3 · p. 19
2.9 Model Evaluation
Evaluation compared the best-performing model with population-mean and fixed-normal-BP estimators, assessed it against international standards, and tested modality contributions and BMI subgroup differences.
2.9.1 Baseline and Naïve Comparison
- The best-performing model was compared with LOSO-fold training-population mean BP estimates and fixed naïve estimates of 120/80 mmHg to assess whether it learned demographic- and biosignal-specific relationships.Comparisons used MAE, ME, and STD.Source: §2.9.1 · p. 19
2.9.2 International Standards
- The model was evaluated against AAMI ME and STD limits of ±5 and ±8 mmHg, IEEE 1708 MAE grades, and BHS cumulative MAE thresholds [69, 74, 42].A device was recommended if SBP and DBP both passed AAMI and earned BHS grade A or B; Table 5 summarizes the standards.Source: §2.9.2 · p. 19, Table 5 · p. 19
2.9.3 Ablation Study
- Input modalities were systematically removed from the best configuration, and each variant was retrained under identical LOSO cross-validation conditions.Variants were evaluated using ME, MAE, and STD, as well as AAMI, IEEE 1708, and BHS standards.Source: §2.9.3 · p. 20
2.9.4 Subgroup Study
- Subgroup analysis compared obese participants (BMI ≥30 kg/m2) with non-obese participants using the best configuration and the same LOSO framework.ME, MAE, and STD were assessed; MAE differences were tested two-sided with Mann–Whitney U, and effect size was measured with Cliff’s δ.Source: §2.9.4 · p. 20
3 RESULTS
The proposed deep learning model achieved the strongest overall estimation performance, while subgroup results showed higher errors among obese participants, particularly for DBP. Results also reveal differences across timing features and model types.
3.1 Correlation between Blood Pressure and Pulse Arrival Time
- PAT1 and PAT3, but not PAT2, showed significant BP correlations, supporting PPG peaks and onsets over maximum-slope timing.PAT1 correlated positively with SBP (r = 0.30, p = 1.5 x 10-5), but its DBP relationship was weaker and nonsignificant (r = 0.13, p = 0.066). PAT3 correlated negatively with SBP (r = -0.30, p = 1.5 x 10-5) and DBP (r = -0.22, p = 0.0014); PAT2 correlations were nonsignificant for SBP (r = 0.04, p = 0.6) and DBP (r = 0.00, p = 0.95).Source: §3.1 · p. 20, Figure 10 · p. 21
3.2 Mechanism-based Models
- Demographic-feature models performed best among mechanism-based approaches, and the 30-second window gave the best combined MAE and STD.The 30-second model had SBP 11.88 ± 6.30 and DBP 8.75 ± 3.98; PAT2 gave the lowest SBP MAE across window sizes despite weak linear correlation, while PAT2 and PAT3 were comparable for DBP.Source: §3.2 · pp. 21–22, Table 6 · p. 22
3.3 Machine Learning Models
- ML models had slightly lower SBP MAE than mechanism-based models but higher variability, with STD exceeding the AAMI limit for both BP estimates.ML SBP MAE ranged from 11.08 to 11.23 mmHg and DBP MAE from 8.78 to 9.13 mmHg; STD was approximately 14 mmHg for SBP and 11 mmHg for DBP, making predictions inconsistent across held-out subjects.Source: Table 7 · p. 22, §3.3 · p. 22
3.4 Deep Learning Model
- The proposed DL model had the lowest MAE across approaches, with best performance at 8 seconds: SBP MAE 9.13 mmHg and DBP MAE 7.77 mmHg.Its STD values were also lowest (SBP ≈ 5.50 mmHg, DBP ≈ 3.80 mmHg) and stable across window lengths.Source: Table 8 · p. 23, §3.4 · p. 23
3.5 Baseline and Naïve Comparison Results
- The proposed DL model outperformed the baseline and naïve estimators for both SBP and DBP, primarily by reducing MAE.STD values remained comparable, suggesting the model leveraged demographic and biosignal inputs rather than collapsing toward constant predictions.Source: Table 9 · p. 24, §3.5 · p. 24
3.6 Ablation Study Results
- 9.13 mmHg SBP MAE was the complete model’s best result; removing demographic features caused the largest degradation for both BP estimates.For SBP, the complete model had ME 0.06 mmHg, STD 5.60 mmHg, and passed AAMI but received Grade D under IEEE 1708 and BHS; removing demographics raised MAE to 12.99 mmHg and STD to 7.78 mmHg. Removing ECG raised SBP MAE to 10.90 mmHg, while removing PPG yielded 9.30 mmHg. 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 resulted in Grade D under BHS.Source: Table 10 · p. 24, Table 11 · p. 25, §3.6 · p. 25
3.7 Subgroup Study Results
- Obese participants had higher DL-model MAE than non-obese participants for SBP (10.72 vs 8.90 mmHg) and DBP (13.96 vs 6.88 mmHg).The DBP difference was significant (U = 148, p = 0.011, δ = 0.691); the SBP difference was not (U = 120, p = 0.197, δ = 0.371).Source: §3.7 · p. 25, Table 12 · p. 26
4 DISCUSSION
The discussion argues that deep learning and multimodal smartwatch inputs support calibration-free BP estimation on unseen subjects, while subgroup disparities and study limitations constrain clinical reliability. Larger, more diverse studies and additional sensing modalities are needed before real-world deployment.
DISCUSSION
- Mechanism-based and ML models generalized worse than the proposed DL model; DL learns waveform representations modulated by demographics, helping maintain plausible predictions for unseen subjects.The discussion attributes greater ML error variability to sensitivity to inter-subject variation and says baseline analysis supports learning subject-specific patterns rather than collapsing predictions toward the training-population mean.Source: §4 · pp. 26–27
- Removing demographics caused the largest LOSO MAE increase; ECG-only models had lower SBP MAE than PPG-only models (13.57 vs. 14.55 mmHg), whereas PPG-only models had lower DBP MAE (9.37 vs. 10.63 mmHg).The modalities may provide complementary information, and fusion consistently outperformed single-modality models, with further improvements after adding demographics. PPG morphology may degrade with motion, poor wrist contact, and ambient lighting.Source: §4 · p. 27
- The proposed DL model achieved the strongest overall LOSO performance and met AAMI requirements for SBP and DBP, but received IEEE 1708/BHS Grade D for SBP and Grade C for DBP.The authors caution that prediction errors do not consistently remain within clinically acceptable thresholds to replace cuff-based standards, despite comparable errors to prior wrist-worn studies evaluated under strict LOSO-CV.Source: §4 · p. 27
- Obese participants (n = 5) had significantly higher DBP MAE and a nonsignificant SBP MAE increase, indicating demographic encodings did not eliminate subgroup bias.The authors suggest altered peripheral hemodynamics and ECG skin-electrode effects may contribute; multi-wavelength PPG has improved BP estimation over single-channel approaches, and additional physiological information or advanced fusion may improve equity.Source: §4 · pp. 27–28
- Interpretation is limited by 200 recordings from 40 participants split into autocorrelated windows, low subgroup-analysis power, self-reported demographics, and data that did not capture fully free-living conditions.Cuff-reference effects and behaviors such as pre-acquisition coffee may add variability, and smartwatch ECG requires active user interaction, limiting continuous multimodal monitoring.Source: §4 · p. 28
- The authors envision ECG-PPG fusion during routine or clinically relevant HBPM sessions, passive PPG monitoring between sessions, and categorical BP labels, while calling for larger, diverse, free-living studies before reliable deployment.Future work should add sensing modalities and advanced fusion strategies to improve generalizability to unseen subjects.Source: §4 · p. 28