Feasibility and performance evaluation of PPG on a Galaxy Watch in continuous central blood pressure monitoring

Abstract

Aims

Wearable, cuffless blood pressure (BP) monitoring systems using photoplethysmography (PPG) offer promising alternatives to traditional methods for out-of-office BP but lack rigorous validation against invasive continuous central BP measurement. This study aimed to develop and evaluate a smartwatch-based algorithm for detecting changes in mean arterial pressure (MAP) using PPG signals from the Samsung Galaxy Watch, with invasive central BP as the reference standard.

Methods and results

A novel algorithm was developed and calibrated using 6117 measurements from 440 participants. External evaluation was performed prospectively in 114 participants undergoing left and/or right heart catheterization, where central aortic BP was recorded at 1-min intervals using pigtail catheters. Simultaneous non-invasive PPG signals were collected from Galaxy Watch 6. The primary endpoint was the detection of a ≥15% increase in MAP from baseline, with receiver operating characteristic analysis performed to evaluate diagnostic accuracy. After calibration, the smartwatch-estimated MAP changes demonstrated strong correlation with invasive measurements (r = 0.92, P < 0.001), with a mean bias of 0.51 mmHg (Bland–Altman analysis). The algorithm showed excellent diagnostic performance for detecting ≥15% increases in MAP (AUC = 0.85; 95% CI = 0.80–0.91) and remained robust in an uncalibrated dataset. The time delay between invasive and non-invasive BP changes was minimal (median = 1.0 min), indicating near real-time tracking capability.

Conclusion

This study demonstrates the preliminary feasibility and accuracy of a PPG-based smartwatch algorithm for detecting meaningful BP changes in real time without the need for frequent calibration. Since our protocol did not meet the relevant validation requirements, further studies are needed to substantiate its potential clinical applicability.

Graphical Abstract

Graphical Abstract.

Introduction

Accurate and timely assessment of blood pressure (BP) outside of the clinical setting is essential for early identification of hypertension and for evaluating treatment effectiveness for hypertension.¹ This need is emphasized in the current clinical guidelines to optimize the management of hypertensive individuals.^2,3 However, conventional cuff-based techniques to detect out-of-office BP, such as ambulatory BP monitoring or home BP monitoring, present notable limitations, including patient discomfort, interference with daily activities, and the possibility of failure to capture transient BP fluctuations due to the sporadic nature of measurements.⁴ Recent advances in digital health technologies have facilitated the development of wearable, non-invasive BP monitoring systems using photoplethysmography (PPG), which overcome these challenges. These systems provide frequent BP assessments with minimal disruption to the patient, offering a promising alternative to traditional methods.^5,6

Several studies previously identified favourable correlations between PPG-based algorithms and intermittent cuff-based measurements.^7–11 Nevertheless, data are limited regarding the validation of wearable cuffless BP monitoring under various conditions using the gold standard of invasive continuous central BP measurement, which is crucial for detecting rapid clinically meaningful BP changes. Furthermore, most previous studies have primarily focused on the prediction ability for systolic BP (SBP) and diastolic BP (DBP) rather than mean arterial pressure (MAP), which is a better prediction marker of cardiovascular events.¹² While detecting relative changes rather than absolute values is important for patient notification, most previous studies used algorithms that attempted to estimate the exact value through calibration. To date, no PPG-based BP monitoring device has met the rigorous standard criteria required for clinical implementation.

Therefore, this study aimed to develop a smartwatch-based change of MAP estimation algorithm using the Galaxy Watch. We rigorously evaluated its accuracy against the changes of invasively measured central BP every minute during cardiac catheterization.

Methods

Model development

The pulse wave analysis-based algorithm for BP estimation was developed using PPG signals obtained from the Samsung Galaxy Watch 6 (Samsung Electronics Co., Suwon, Republic of Korea) (see Supplementary material online, Figure S1). The development phase included 6117 measurements from 440 participants with BP profiles ranging from normotension to hypertension (see Supplementary material online, Table S1). The data were sourced from six independent databases that adhered to the selection criteria outlined in International Organization for Standardization (ISO) 81060-2:2018, ensuring standardization in BP estimation and compliance with medical-grade accuracy standards. The algorithm follows a structured process consisting of three key stages: pre-processing, calibration, and BP estimation (see Supplementary material online, Figure S2).

The first stage was pre-processing. The raw PPG signal was acquired from the smartwatch sensor and underwent a series of pre-processing steps to ensure signal quality. A noise cancellation filter was applied to remove motion artefacts and unwanted signal components. The signal quality estimator then evaluated the PPG waveform based on multiple consecutive pulse signals. To minimize the influence of motion artefacts and ambient noise, a 4th-order IIR bandpass filter (0.4–8 Hz) was applied to the raw PPG signal. Signal quality was assessed using a signal quality index (SQI), calculated based on pulse-to-pulse correlation. Only pulse waveforms with an SQI above a predefined threshold were selected for further feature extraction and blood pressure estimation. If the signal quality surpassed a predefined threshold, a representative pulse waveform was generated, capturing the key morphological features of the PPG signal corresponding to a single cardiac cycle. Simultaneously, reference BP values were obtained through non-invasive measurements based on auscultation, allowing real-time mapping between the PPG-derived waveform and ground-truth BP values.

The second stage was calibration. The algorithm established a reference framework for BP estimation. This process was repeated until a sufficient number of PPG waveforms and corresponding reference BP values were collected, forming the calibration dataset.

The final stage, BP estimation, involved a multi-step approach to dynamically compute BP changes. The extracted features from newly acquired PPG waveforms were normalized against the baseline features obtained during calibration. A feature scaling process was then applied to account for non-linear physiological variations and attenuate measurement fluctuations. The algorithm subsequently combined multiple PPG-derived waveform features to enhance robustness and accuracy. Finally, the estimated MAP change was calculated by applying a scale factor to the combined features. The final BP value was determined by adding this estimated BP change to the reference MAP obtained during calibration. The algorithm was developed for SBP, DBP, and MAP estimation and the calibration for SBP, DBP, and MAP estimation is performed separately.

Model assessment

Study design and participants

External assessment of performance was conducted as a prospective, single-arm, temporal study to evaluate the diagnostic accuracy of the smartwatch-based BP monitoring algorithm (Figure 1). The study enrolled adults (19 years or older) who needed to undergo left and right heart catheterizations as a clinical requirement at Samsung Medical Center, Seoul, Republic of Korea, between November 2023 and December 2024. Participants were excluded if they had significant mitral or tricuspid regurgitation, a history of constrictive pericarditis, or an inability to wear a smartwatch on their wrist. The study protocol was approved by the Institutional Review Board of Samsung Medical Center (SMC 2023-07-193), and all participants provided written informed consent.

Figure 1.

Scheme of the current assessment of performance study. BP = blood pressure; PPG = photoplethysmography.

Invasive BP measurement protocol

Invasive BP measurements were obtained using a catheter-based system (Mac-Lab, GE HealthCare, Chicago, Illinois, USA). Invasive MAP values were directly calculated by the system through time-averaged integration of the area under the aortic pressure waveform, representing the gold standard for MAP measurement. This method does not rely on formula-based approximations. Participants were placed in the supine position. Central BP was recorded at 1 min intervals with a pigtail catheter placed in the ascending aorta through a radial or femoral artery sheath. Time synchronization was achieved using timestamped recordings, ensuring that invasive and non-invasive BP measurements were aligned. Calibration was performed using the first three 1-min interval measurements, during which PPG signals and invasive BP were simultaneously recorded under stable haemodynamic conditions. This yielded three calibration data points per participant, forming the baseline reference for estimating BP changes over time

Endpoint

The primary endpoint was a 15% increase in MAP from baseline. Hypertension was defined as SBP ≥140 mmHg or DBP ≥90 mmHg, while normal BP was defined as SBP <120 mmHg and DBP <80 mmHg. The transition from normal BP to hypertension stage 2 involved an approximate 16.7% increase in SBP (from 120 to 140 mmHg), a 12.5% increase in DBP (from 80 to 90 mmHg), and a 14.3% increase in MAP (from 93.33 to 106.67 mmHg). Given that the threshold for defining hypertension in clinical practice is based on an absolute BP increase, a 15% increase from baseline MAP was considered a physiologically relevant threshold that approximates the transition from normotension to hypertension. As a secondary endpoint, we assessed a 12 mmHg absolute increase in MAP from baseline. Additionally, we noted 15% increases in SBP and DBP and absolute increases of 15 mmHg in SBP and 10 mmHg in DBP.

Covariates

Comprehensive clinical data were collected through electronic medical record reviews, structured medical interviews, and physical examinations. Collected demographic information included age, sex, height, weight, body mass index, and body surface area. Baseline BP measurements, including SBP, DBP, and MAP, and heart rate were recorded.

Cardiovascular diagnoses were documented, including heart failure with preserved ejection fraction, heart failure with reduced ejection fraction, pre-capillary pulmonary hypertension, and non-cardiac dyspnoea, on the basis of echocardiographic and catheterization findings. Additional clinical factors such as hypertension, diabetes mellitus, dyslipidaemia, atrial fibrillation, smoking status, prior heart failure hospitalizations, and history of coronary artery disease were also collected.

Functional status was assessed using the New York Heart Association classification, with participants categorized into Class I–IV based on symptom severity. Laboratory data included haemoglobin, serum creatinine, and estimated glomerular filtration rate to assess renal function. N-terminal pro b-type natriuretic peptide level was measured as a biomarker of cardiac stress.

Statistical analysis

The sample size was estimated using an area under the curve of receiver operating characteristic (AUROC)-based hypothesis testing framework to evaluate the ability of the smartwatch-based BP estimation algorithm to detect a ≥15% increase in invasive MAP. Assuming a null hypothesis AUROC of 0.50, an expected AUROC of 0.80, an event prevalence of 1.5%, a two-sided α of 0.05, and 90% power, the required number of observations was calculated to be 1667. Accounting for a 10% dropout or data loss due to poor signal quality, the total required sample size was 1853 observations. As each participant undergoing invasive catheterization was expected to yield approximately 15 valid paired MAP measurements, 120 participants were considered sufficient to meet the analytical requirements.

Continuous variables were reported as mean [standard deviation (SD)] or median [interquartile range (IQR)] and categorical variables as n (%). Differences between invasive and non-invasive BP values were assessed using paired t-tests, and Pearson’s correlation coefficient was used to quantify associations. Bland–Altman analysis was used to evaluate systematic bias and limits of agreement.

A logistic regression model to detect a ≥15% increase in invasive MAP, SBP, and DBP from baseline was developed to assess the predictive performance of changes in MAP, SBP, and DBP from baseline, measured by non-invasive estimation, as predictors. ROC curve analysis was performed, and AUC values were reported. We performed additional analyses using absolute thresholds (12 mmHg MAP, 15 mmHg SBP, 10 mmHg DBP) as a secondary analysis. As a sensitivity analysis, a time-delay analysis was conducted to examine the lag between peak invasive and non-invasive BP changes. A window-based notification performance assessment was performed using a sliding window method (Nwin = 1 and 2 min), and AUC was calculated for each window.

We used case wise-deletion methods for MAP, SBP, and DBP; thus, there were no missing variables in any of the analyses. All analyses were conducted using R 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria). A two-tailed P-value < 0.05 was considered statistically significant.

Results

Participants

A total of 120 participants were initially enrolled in the study, and six (5.0%) were excluded because of poor PPG signal quality. From the final sample of 114 participants, 2299 BP data points were recorded. Among the population, 94 participants underwent cardiac catheterization in resting status only, and 20 participants underwent cardiac catheterization in resting status and during bicycle exercise. Each participant had a median of 14 paired invasive and non-invasive BP measurements. The mean age of participants was 58.0 (SD: 15.7) years, and 50% of them were male. The most common diagnosis was pre-capillary pulmonary hypertension (42.1%), followed by non-cardiac dyspnoea (34.2%). Hypertension and diabetes mellitus were present in 44.7% and 27.2% of participants, respectively (Table 1). Baseline echocardiographic and cardiac catheterization data are presented in Supplementary material online, Table S2. The average measurement duration was 25.6 min (SD: 17.9), with a range of 9–104 min.

Table 1.

Baseline characteristics and clinical presentation of the patients (N = 114)

Characteristic
Age, years, mean ± SD	58.0 ± 15.7
Sex
Male	57 (50.0%)
Female	57 (50.0%)
Body mass index, kg/m2	24.9 ± 5.8
Body surface area, m2	1.7 ± 0.2
Diagnosis
Heart failure with preserved ejection fraction	21 (18.4%)
Heart failure with reduced ejection fraction	6 (5.3%)
Pre-capillary pulmonary hypertension	48 (42.1%)
Non-cardiac dyspnoea	39 (34.2%)
Hypertension	51 (44.7%)
Diabetes mellitus	31 (27.2%)
Dyslipidaemia	54 (47.4%)
Current smoking	12 (10.5%)
Coronary artery disease	17 (14.9%)
Atrial fibrillation	11 (9.6%)
Previous heart failure admission	13 (11.4%)
NYHA classification
I	15 (13.2%)
II	38 (33.3%)
III	52 (45.6%)
IV	9 (7.9%)
Laboratory findings
Haemoglobin, g/dL	13.4 ± 2.1
Creatinine, mg/dL	1.05 ± 0.97
Estimated glomerular filtration rate, mL/min/1.73 m2	83.3 ± 26.5
NT-proBNP, pg/mL, mean (IQR)	322.5 (110.0–1201.0)
Medications
ACEI, ARB, or ARNI	25 (21.9%)
Beta-blocker	32 (28.1%)
Calcium-channel blocker	18 (15.8%)
Loop diuretics	47 (41.2%)
Spironolactone	38 (33.3%)
SGLT-2 inhibitor	15 (13.2%)

Data are presented as mean (SD), median (IQR), or n (%).

ACE, angiotensin-converting enzyme; ARB, angiotensin receptor blocker; ARNI, angiotensin receptor neprilysin inhibitor; IQR, interquartile range; NT-proBNP, n-terminal pro-B-type natriuretic peptide; NYHA, New York Heart Association; SD, standard deviation; SGLT-2, sodium-glucose cotransporter 2.

The mean baseline invasive MAP was 97.1 mmHg (SD: 15.1), with values ranging from 57.7 to 132.3 mmHg. The mean change in invasive MAP from baseline was 0.6 mmHg (SD: 5.8), with a maximum increase of 26.3 mmHg and a maximum decrease of −29.0 mmHg. The mean non-invasive MAP during initialization was 97.1 mmHg (SD: 15.0), and the mean change in non-invasive MAP from baseline was 0.5 mmHg (SD: 3.9). There were no significant differences between invasive and non-invasive MAP values (Table 2).

Table 2.

Invasive and non-invasive mean arterial pressure, systolic blood pressure, and diastolic blood pressure

Characteristic	Invasive (n = 114)	Non-invasive (n = 114)	P value
Mean arterial pressure
Baseline measurement, mmHg	97.3 ± 16.1	97.1 ± 15.0	0.58
Average change from prior values, mmHg	0.1 ± 4.5	0.0 ± 3.8	0.61
Average change from baseline, mmHg	0.6 ± 5.8	0.5 ± 3.9	0.45
Systolic blood pressure
Baseline measurement, mmHg	133.1 ± 25.6	133.1 ± 25.4	0.53
Average change from prior values, mmHg	0.0 ± 6.1	0.0 ± 5.0	0.90
Average change from baseline, mmHg	0.2 ± 10.0	0.2 ± 5.4	0.74
Diastolic blood pressure
Baseline measurement, mmHg	71.7 ± 11.5	71.7 ± 11.4	0.84
Average change from prior values, mmHg	0.1 ± 4.6	0.0 ± 3.4	0.79
Average change from baseline, mmHg	0.9 ± 4.9	0.3 ± 4.1	<0.01

Data are presented as mean ± SD.

The mean baseline SBP was 133.1 mmHg for both invasive and non-invasive measurements. The mean change in invasive SBP from baseline was 0.2 mmHg, with no significant difference between invasive and non-invasive measurements (Table 2). The mean baseline DBP was 71.7 mmHg for both invasive and non-invasive measurements, with a mean change of 0.9 mmHg in invasive measurements and 0.3 mmHg in non-invasive measurements (Table 2).

Correlation and agreement between invasive and non-invasive MAP, SBP, and DBP

After calibration, a strong positive correlation (r = 0.92, P < 0.001) was observed between invasive and non-invasive MAP values (Figure 2A). Bland–Altman analysis showed a mean bias of 0.51 mmHg, indicating close agreement between the two measurement methods. Similarly, a strong correlation (r = 0.93, P < 0.001) was observed between invasive and non-invasive SBP, with a mean bias of −0.04 mmHg in Bland–Altman analysis (Figure 2B). For DBP, the correlation remained high (r = 0.87, P < 0.001), with a mean bias of 0.7 mmHg (Figure 2C).

Figure 2.

Correlation between invasive and non-invasive measures and bland-altman plots. DBP = diastolic blood pressure; MAP = mean arterial pressure; SBP = systolic blood pressure.

The correlation of percent change values between the calibrated and uncalibrated estimates was also very high for MAP (r = 0.94), SBP (r = 0.91), and DBP (r = 0.95) (see Supplementary material online, Figure S3).

ROC curve analysis

In the primary analysis, there were 49 (2%), 74 (3%), and 79 (3%) recorded cases of an absolute increase of 15% in invasive MAP, SBP, and DBP, respectively. ROC curve analysis demonstrated that non-invasive MAP measurements exhibited strong predictive ability for a ≥ 15% increase in invasive MAP, with an AUC of 0.85 (95% CI: 0.80–0.91, Figure 3A). In a subgroup analysis, the algorithm demonstrated an AUC of 0.81 (95% CI: 0.72–0.90) for detecting MAP increases ≥15% among participants with atrial fibrillation (n = 11), and an AUC of 0.84 (95% CI: 0.78–0.92) in those without AF (n = 103). For detecting a ≥ 15% increase in invasive SBP, the AUC was 0.79 (95% CI: 0.74–0.84, Figure 3B). In contrast, the AUC for detecting a ≥15% increase in invasive DBP was lower, at 0.64 (95% CI: 0.57–0.71, Figure 3C). The AUC values obtained from the uncalibrated dataset were comparable to those from the calibrated dataset (Figure 3).

Figure 3.

ROC curve to detect a 15% increase in invasive measurement. AUC = area under the curve; DBP = diastolic blood pressure; MAP = mean arterial pressure; SBP = systolic blood pressure; ROC = Receiver operating characteristic.

In the secondary analysis, there were 63 (3%), 105 (5%), and 70 (3%) recorded cases, with an absolute increase of 12 mmHg in invasive MAP, 15 mmHg increase in SBP, and 10 mmHg increase in DBP, respectively. ROC curve analysis demonstrated that non-invasively measured MAP showed a strong predictive ability for detecting an absolute increase of 12 mmHg in invasive MAP, with an AUC of 0.80 (95% CI: 0.74–0.86; Supplementary material online, Figure S4A). For detecting a 15 mmHg increase in invasive SBP, the AUC was 0.76 (95% CI: 0.72–0.81; Supplementary material online, Figure S4B). In contrast, the AUC for detecting a 10 mmHg increase in invasive DBP was 0.62 (95% CI: 0.56–0.70; Supplementary material online, Figure S4C). Notably, the AUC values from the uncalibrated dataset were also comparable to those from the calibrated dataset for detecting absolute increases of BP (see Supplementary material online, Figure S4).

Time-delay analysis

The median time delay between the peak change in invasive BP and the corresponding non-invasive BP measurement was 1.0 min (IQR: 0.0–2.0 min) for MAP and SBP and 1.0 min (IQR: 0.0–3.0 min) for DBP. Similar results were observed for the uncalibrated dataset. The window-based notification performance analysis showed AUCs for MAP of 0.83 and 0.84, respectively, with a window size of 1 and 2 min, with no significant improvement in detection performance (P > 0.05, Supplementary material online, Figure S5).

Discussion

In this development and feasibility study, we developed a smartwatch-based change of BP estimation algorithm and evaluated its accuracy against changes of invasive central BP. The principal findings from this study are as follows. First, we found that smartwatch-based changes of the BP estimation algorithm provide a strong correlation and agreement with invasive MAP after calibration and a high level of diagnostic accuracy for the detection of both absolute and relative BP changes from the continuous measurements of central BP as a reference. Second, the algorithm also showed excellent diagnostic performance for changes in BP, even in the absence of calibration. Third, the time delay between BP changes and detection by the smartwatch was short.

Recent advancements in digital health technologies, particularly PPG-based pulse waveform analysis, have led to improvements in continuous, non-invasive, cuffless BP monitoring.^13,14 Since the emergence of this concept, there have been many efforts to measure true BP using a variety of non-invasive cuffless devices, including finger probes, wristbands, armbands, chest belts, and vests.¹⁵ Despite these advancements, no cuffless device has met the criteria necessary for inclusion in clinical practice guidelines. The European Society of Hypertension has emphasized the need for specific, clinically meaningful, and pragmatic validation procedures from the different types of cuffless devices intended to capture dynamic BP changes in real-world settings.¹⁶ This perspective suggests the limitations of traditional validation protocols, which may not adequately capture the performance of wearable, trend-based technologies.¹⁷

Many previous studies did not separate the datasets used for development from those used for validation, raising concerns about model overfitting and limited generalizability. Furthermore, validation was often performed using intermittent, non-invasive methods like 24-h ambulatory BP monitoring, which cannot adequately reflect the demands of continuous BP monitoring.^{7–11,18–22} While a few studies used invasive continuous BP as a reference standard, most were limited by small sample sizes.^23–25 Prior research primarily focused on the device’s ability to match absolute BP values, even though the ability to detect clinically meaningful increases from an individual baseline BP, such as hypertensive transitions, may be more relevant in a real-world setting. To address these gaps, we designed the current study with several key considerations. First, we developed and evaluated a smartwatch-based algorithm for estimating changes in MAP, validating minute-by-minute against invasively measured central BP, using a large and sufficiently powered dataset in accordance with clinical and methodological standards.¹⁶ Additionally, we applied a rigorous development framework that separated the training and assessment of performance cohorts, improving generalizability. Notably, rather than focusing solely on absolute BP estimation, we assessed the algorithm’s ability to detect a ≥15% increase from baseline MAP—a physiologically and clinically meaningful threshold that aligns with the transition from normotension to hypertension. This approach directly reflected the European Society of Hypertension’s emphasis on trend detection and patient-specific changes as key aspects of cuffless BP monitoring utility. In this study, we found that the smartwatch effectively tracked BP changes from baseline and detected clinically significant BP elevations. The current results add to the accumulating evidence in support of cuffless BP devices and suggest the possibility for the clinical use of non-invasive cuffless devices in real-world clinical settings. Nevertheless, the study protocol of this study did not exactly follow the standard ISO 81060-3:2022 guidelines. Although the number of subjects with measurements was satisfactory, the measurement interval (ISO 81060-3:2022 guideline recommendation: ≤30 s, but the current study: 1 min) and the conditions under which BP was measured (primarily resting state, with exercise in only a subset) did not satisfy the standard protocol. Accordingly, the results should be interpreted with caution. Additionally, the number of discrete BP change events per subject (≥50) required for ISO-defined change-tracking metrics could not be satisfied. While we performed an exploratory analysis using a relaxed threshold (≥5 events), the performance did not meet the ISO acceptance criteria (P50 ≤ 25%, P85 ≤ 50%). Future studies employing longer observation windows and stress-induced BP changes will be necessary for ISO-compliant evaluation.

An additional strength of the algorithm is that its accuracy was not dependent on calibration. Many existing cuffless BP monitors require frequent recalibration using a traditional BP cuff to maintain accuracy over time. The Aktiia bracelet showed high precision in controlled settings but required regular calibration to prevent measurement drift.²⁵ In contrast, the smartwatch algorithm used in this study maintained accuracy even without recalibration, as evidenced by similar AUC values for detecting a clinically significant increase in BP in both calibrated and uncalibrated datasets. In addition, percent change values estimated from the calibrated and uncalibrated algorithms showed excellent agreement (r = 0.94 for MAP, 0.91 for SBP, and 0.95 for DBP), further supporting the algorithm’s robustness to calibration drift. This suggests that the algorithm’s reliance on multiple PPG waveform features, rather than static calibration models, contributes to its long-term stability and reliability.

In the current study, the time delay between BP changes and detection by the smartwatch was short. The minimal time delay observed in BP tracking represents a key advantage of this smartwatch-based system. Previous studies have reported that wearable BP monitors exhibit lags of several seconds to minutes in detecting BP fluctuations, limiting their applicability in hypertension management.²⁶ The present study demonstrated that the smartwatch detected BP elevations within one second of the actual increase, indicating its potential for real-time haemodynamic assessment. The rapid BP tracking observed in this study is likely attributable to high-frequency signal sampling and real-time adjustments in pulse wave analysis, allowing for near-instantaneous BP estimation even in unstable environments.

Limitations

This study has several limitations. First, the measurement was evaluated by wrist-based BP monitoring. Unlike upper-arm cuff measurements, which are closer to the brachial artery, wrist-based PPG sensors are influenced by changes in wrist positioning, varying levels of arterial compression, and differences in peripheral vascular tone. Further studies should explore whether alternative devices, such as finger rings or ear-worn sensors, provide improved BP measurement stability, particularly for DBP estimation. Second, while the smartwatch was validated in both stationary and ambulatory settings, the measurement duration was limited to a relatively short period, typically lasting only several minutes. This may not fully capture the long-term stability and accuracy of the device, particularly in real-world conditions where individuals experience varying physiological states over extended periods. Although our algorithm showed strong agreement with invasive BP changes and demonstrated accurate detection of clinically meaningful BP elevations (e.g. ≥12 mmHg or ≥15% MAP increase), this study was not designed to assess the device’s performance in classifying individuals as hypertensive according to clinical diagnostic thresholds. In particular, our exploratory ISO-based analysis did not satisfy the strict criteria due to shorter monitoring intervals and limited event counts; therefore, our findings should be interpreted as preliminary. Further studies are warranted to validate its use in point-of-care or ambulatory settings where absolute BP classification is essential. Third, calibration and assessment of performance were performed in close temporal proximity, raising concerns that the algorithm’s accuracy might have been overestimated due to limited intra-individual BP variation. However, approximately 20% of participants underwent graded exercise during the feasibility testing phase, introducing substantial dynamic blood pressure changes and enabling meaningful assessment of real-time BP tracking. Future studies incorporating delayed or repeated calibration over extended periods would help further evaluate long-term calibration stability and the robustness of the algorithm in everyday conditions. Fourth, this was a single-centre study with a relatively small sample size and short monitoring duration, typically limited to several minutes. A substantial proportion of the assessment of performance cohort had pulmonary hypertension, which may alter vascular compliance and affect PPG waveform morphology. These factors may limit the generalizability of the findings to broader populations and real-world settings. Future studies should include longer-term monitoring across diverse clinical conditions and care environments.

Conclusions

This study demonstrates the feasibility of using a smartwatch-based, cuffless BP monitoring system to detect clinically meaningful changes in MAP with high temporal resolution and without the need for frequent recalibration. These results highlight the potential of this system for real-time, non-invasive haemodynamic monitoring. Given the small sample size, single-centre design, and brief monitoring window, these findings should be interpreted as preliminary work. Therefore, future studies should evaluate the long-term stability and usability of this technology in diverse populations and settings, including ambulatory and home-based care.

Supplementary material

Supplementary material is available at European Heart Journal – Digital Health.

Author contributions

Ki Hong Choi (MD PhD (Conceptualization [equal]; Formal analysis [equal]; Investigation [equal]; Writing—original draft [lead])), Chang Soon Park (PhD (Investigation [equal]; Methodology [equal]; Resources [lead]; Software [lead])), Danbee Kang (PhD (Formal analysis [lead]; Methodology [lead]; Visualization [lead]; Writing—original draft [equal])), Ji Hyun Cha (MD (Supervision [equal]; Writing—review & editing [equal])), Jihoon Kim (MD (Supervision [equal]; Writing—review & editing [equal])), Boram Lee (RN (Data curation [equal]; Investigation [equal])), Eunjin Kim (RN (Data curation [supporting]; Resources [supporting])), Seorim Kim (RN (Data curation [supporting]; Resources [supporting])), Dae-Geun Jang (PhD (Methodology [equal]; Software [equal]; Validation [equal])), Eunkyu Oh (MS (Software [equal]; Validation [equal])), Sung-Hwan Cho (PhD (Resources [equal]; Supervision [equal]; Validation [equal])), Jongmin Choi (PhD (Investigation [equal]; Resources [equal]; Software [equal]; Supervision [equal])), and Jeong Hoon Yang (Conceptualization [lead]; Investigation [lead]; Supervision [lead]; Writing—review & editing [lead])

Funding

This study was supported by Samsung Electronics.

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author.