Data-hugging shields proprietary AI models from research that could disprove them

“Water, water everywhere, nor any drop to drink” from The Rime of the Ancient Mariner, by Samuel Taylor Coleridge, nicely describes almost every academic AI health researcher’s challenges with access to data.

Medical AI has always promised to revolutionize healthcare, from early disease detection to personalized treatment, but its success hinges on robust evidence and public trust. Yet today, many cutting-edge health AI models are developed behind closed doors on proprietary datasets, precluding independent verification. Data and code sharing is remarkably scarce in medical AI literature: a recent systematic review of 1342 studies of AI in critical care found that 85% of studies did not make their datasets available, and 87% of studies did not provide relevant code¹. Such widespread absence of transparency and reproducibility impedes external validation, hampers cumulative scientific progress, and, as we show here, leads us to believe we are safer than we are.

Several high-profile failures of prominent AI algorithms have already surfaced: Epic’s widely used sepsis prediction algorithm was shown to have poor performance on external retrospective validation, after it had already been deployed in hospitals across the United States². Likewise, Philips recently issued a Class I software correction for its outpatient telemetry monitoring service when the algorithm failed to transmit ECG alerts for atrial fibrillation and ventricular tachycardia, contributing to two patient deaths and 109 injuries³. Perhaps these were anomalies; however, could this happen more often for applications where we do not see immediate effects, and we just do not know about it? How broad a scale do these errors reach? One prominent case arises from the recent revolution in heart monitoring enabled by smartwatches. Wearable-based detection of heart issues presents a major opportunity, as many people remain unaware of cardiovascular problems until serious events occur. Atrial fibrillation (AFib), in particular, is a common yet often undiagnosed condition that significantly increases the risk of stroke⁴. Apple claims to detect AFib from smartwatch-derived photoplethysmographic (PPG) signals with exceptionally high accuracy⁵, offering the potential to improve outcomes for millions through early intervention. However, the lack of access to Apple’s dataset, or any comparable large public dataset, prevents independent researchers from verifying these claims. Should such claims from entities like Apple be accepted without independent scrutiny? They performed experiments on a proprietary large-scale dataset and reported extensive results to the FDA⁶. However, a previous independent evaluation on the Apple Watch showed that the detection algorithm could not classify 27.9% of ECG signals that were provided to it⁷. More disconcertingly, other claims presented by Apple strain credulity. Specifically, Apple recently presented results indicating that they could predict a person’s biological age along with other demographic features purely from PPG signals collected via Apple Watch, with an average error of roughly three years for age prediction⁸, while the rest of the field was achieving error around seven to ten years^{9,10,11,12,13,14,15,16,17,18,19,20}. Subsequent Apple papers reported even lower errors (2.89 years using PPG alone and 2.46 years with behavioral features²¹; ~2.4 years in a self-reported healthy cohort and ~3.2 years in the general cohort²²). However, these results still lack routine independent replication because the Apple Heart and Movement Study dataset is not publicly available; releasing code and synthetic data cannot substitute for external validation on real, independent datasets²². These escalating, state-of-the-art claims, produced on the same inaccessible Apple Heart and Movement Study dataset, further highlight the widening chasm between proprietary results and what is verifiable by the broader scientific community.

Any of these results, if validated, would be remarkable. They would imply that age information, accurate to within approximately three years, could be collected with light beamed into surface vessels on one’s wrist over a short period of time, independent of genetics or lifestyle factors such as exercise or smoking history—factors that typically create discrepancies between vascular age measures and chronological age.

While not entirely discounting biological plausibility, several factors render these claims challenging to accept at face value. First, measurements of vascular aging generally are not able to pinpoint age that precisely¹⁷; even epigenetic measures of aging are only able to estimate age to about 3.6 years²³. Second, neither of Apple’s papers used a single external or public dataset. The authors’ models are not public, so there is no way to test model performance on external datasets, and their data remains proprietary, preventing any validation of their findings. Third, their claims lack interpretability, and as such, the paper does not provide any additional scientific insight or benefit to the scientific community that could be used to verify the claims or inform future research.

The alternative text for this image may have been generated using AI.

Full size image

The main panel shows two representative groups of PPG signals (Group 1: blue; Group 2: red), each containing six signals from different 5-year age buckets spanning 40–70 years. These results demonstrate high similarity within each group and clear dissimilarity between groups, illustrating consistent morphology despite age variations. Inset panels display individual signals with participant ages (top: Group 1; bottom: Group 2).

The alternative text for this image may have been generated using AI.

Full size image

This figure illustrates substantial variation within each age group.

If we could determine whether Apple’s claim on age prediction is replicable, it would help determine whether its accompanying results on atrial fibrillation (which is a more clinically significant condition to detect) are valid. In other words, by testing claims that are plausible to test using public data, we will either verify an amazing discovery or cast doubt on related (important) results that we have insufficient data to verify.

Thus, in this work, we used public PPG data to try to evaluate Apple’s reported findings on age. If Apple’s results are replicable, we would expect to see at least some predictability in age. But we did not find much improvement beyond just predicting the mean population age. Our analysis, carried out using a deep neural network comparable to the one used in Apple’s studies^8,21, both trained on the UK Biobank²⁴, yields results more consistent with physiological understanding than Apple’s reported outcomes. Specifically, we observed substantial variability in PPG signals among individuals of identical chronological ages (see Fig. 1) and high similarity in some PPG signals across different chronological age groups (also see Fig. 2, Supplementary Information Fig. S1, and Supplementary Information Fig. S2), which indicates low predictability of age from PPG morphological features alone. Our results indicate that there is a fairly low ceiling for predictive accuracy in biological age estimation. They suggest that Apple’s claims may not be biologically plausible or generalizable without substantial additional covariates or interpretive context.

The alternative text for this image may have been generated using AI.

Full size image

Bars show the Q1–Q3 range, and dots show the median. For this visualization, data points falling below Q1 − 3*IQR or above Q3 + 3*IQR were considered outliers and removed. The within-age-group variation is always large.

The alternative text for this image may have been generated using AI.

Full size image

Left: PaCMAP of SMoLK model embeddings on validation fold. Right: PaCMAP of ResNet model embeddings on validation fold. Age does not dominate the learned representations.

Our results cast doubt on some of the largest-scale and most important types of results in AI health, which is problematic because atrial fibrillation detection is important. We place the blame for this circumstance purely on “data hugging,” i.e., not sharing data. There is simply no high-quality, large, public PPG dataset, nor a public testing platform for PPG, from any source on-par with the Apple study. If there were, we could test Apple’s results on atrial fibrillation directly.

While gathering and maintaining high-quality data is challenging²⁵, we claim that constructing and maintaining these datasets and testing platforms is completely within the realm of possibility and has precedent. The National Institute of Standards and Technology (NIST) has been running facial recognition vendor tests, as well as speaker recognition vendor tests, since 1996^26,27,28,29. This approach could readily extend to health data. If anyone were to devote the same effort to specific health data systems as NIST does to facial recognition and speaker recognition systems, it could lead to a measurable and almost immediate improvement in human health, with many lives saved.

Below, we first discuss Apple’s study, then examine the problems with public PPG data to contextualize the disparity between what companies possess and what the public can access. We then present our effort to predict age using public PPG data, discuss the implications of our results on age estimation, and propose remedies that were briefly outlined above. In short, our position is that data hugging is becoming increasingly problematic due to the increase in widespread proprietary models, and that there is a practical remedy, which is the development of a few, highly curated public benchmark datasets and evaluation platforms.

A research paper by Apple recently presented a striking advance: a foundation model trained on a large dataset of PPG and electrocardiogram (ECG) data from the Apple Heart and Movement Study⁸. The study enrolled approximately 141,000 participants aged 18 years and older, with no upper age limit, who contributed continuous wrist-based PPG and ECG recordings collected via Apple Watch over a period of up to 3 years. Apple’s self-supervised foundation model, trained on this proprietary dataset, reportedly encodes participant characteristics effectively enough that researchers can predict demographics with very high accuracy using linear models on its features. For example, using PPG data, the model was able to distinguish older vs. younger individuals (above vs. below 50) with an AUC of 0.976. Moreover, in the regression task, the model could predict a patient’s exact age with a mean absolute error of only 3.2 years. These results were followed by an even more striking claim from a subsequent Apple study, which reported a mean absolute error of 2.46 years by combining sensor and behavioral data²¹.

A closely related follow-up study, Miller et al. (2025), introduces a wrist-PPG “aging clock” (PpgAge) and reports associations between age-gap statistics and cardiometabolic disease, behaviors (smoking, exercise, sleep), and longitudinal physiological events²². However, neither the models nor the underlying data are publicly available, leaving the reported performance unvalidated on independent datasets.

Age estimation and atrial fibrillation detection are totally different problems. We know we cannot externally validate Apple’s claims with respect to the detection of atrial fibrillation since we do not know what data were omitted. Atrial fibrillation is easy to detect if the data are clean^30,31 however, the data are essentially never clean. One way to find almost perfect accuracy in atrial fibrillation detection is to report accuracy after omitting all of the noisy signals, which could be a sizable fraction of the data. But omitting much of the data negates the point of using the device in the first place. In contrast, age does not rapidly change over time like atrial fibrillation signals, and the reported age-prediction error of roughly 2–3 years is substantially tighter than previously demonstrated on public datasets. While potential confounders such as medication use or disease status may contribute to age prediction, these factors alone are unlikely to explain near-perfect accuracy, and possibilities like inadvertent data leakage or methodological errors cannot be independently assessed without access to the data or model. These considerations make age prediction a particularly informative case for assessing replicability.

Despite the unprecedented potential for heart monitoring to save lives through early disease detection, publicly accessible datasets for PPG signals, particularly those annotated with critical labels like age, remain limited in size, scope, and detail. Currently, the largest publicly accessible dataset is the UK Biobank, consisting of finger-based PPG recordings for ∼170,000 patients, roughly between the ages of 40 and 70^24,32 at the first assessment visit. Each waveform consists of a single pulse of 100 points collected after the waveform stabilizes over 10–15 s. Other public PPG datasets are typically orders of magnitude smaller. For example, the MIMIC database consists of on the order of 10,000 critically ill ICU patients rather than a general population cohort³³. Other public datasets are often drawn from specialized contexts, such as sleep studies or physical activity trials, and include many fewer patients. Some notable datasets include the WESAD dataset, consisting of PPG signals from 15 participants undergoing a lab study of stress detection³⁴, PPG3 DaLiA with continuous physiological PPG recordings from 15 participants during daily activities³⁵, Stanford Wearables Dataset consisting of about 50 patients monitored longitudinally over several months^31,36,37, VitalDB consisting of about 6000 PPGs recorded during surgical cases³⁸, MESA with about 2,000 patients who underwent a sleep study and had PPG signals recorded throughout³⁹, and the nuMoM2b study, recording overnight PPG signals during a sleep study for over 3000 moms in pregnancy⁴⁰. Only recently have initiatives like PaPaGei begun to aggregate existing public datasets, achieving about 13,500 patients from publicly available datasets for the purpose of training a foundation model⁴¹. However, these efforts remain nascent, and widespread adoption is still limited.

All publicly available datasets have critical limitations, including short recording lengths (typically 30-s clips), preventing long-term analyses, inadequate sample sizes (sometimes as few as 5 patients), and incomplete covariate information, such as smoking status or body mass index. Even the UK Biobank, currently one of the most comprehensive public health datasets, presents substantial challenges for training deep learning models, including cost, long wait times for access, inability to download individual data directly, and infrastructure issues on the DNANexus Platform. Operationally, utilizing the UK Biobank via DNANexus constrains researchers to work within its specific cloud infrastructure, necessitating careful resource management to avoid job termination during computationally intensive tasks like deep learning model training and to manage usage-based costs. These costs are incurred for compute utilization, storage of data and models, and data egress. For example, running an on-demand instance with 144 GiB of memory for 48 hours (mem1_ssd1_v2_x72) would cost £94 on the platform, in addition to a £3000 3-year Tier 1 access fee. The requirement to perform computations exclusively within this environment, along with the high costs, influences the structure and feasibility of large-scale, iterative deep learning experiments.

Lack of public data means: (1) potential biases go unchecked until deployment [see the dermatology study of Ibrahim et al. ⁴²], including those due to the goal of misleading readers [see the critique of Kan-Tor et al. ⁴³, about an IVF study] (2) underrepresented groups may be excluded [e.g., “health data poverty,” ⁴⁴], (3) errors in coding or reporting go unchecked⁴⁵, (4) generalizable knowledge or insight about the medical condition itself cannot be gained—the revolution in health data science is stalled precisely because of these issues.

We aimed to replicate Apple’s findings using the UK Biobank dataset. The dataset provides PPG measurements alongside a wealth of other clinical data. We focused on data from each patient’s first assessment visit conducted between 2006 and 2010, which includes finger PPG recordings from 170,624 patients.

Analysis 1: arterial stiffness index, a known age estimator

We first studied how the arterial stiffness index varied as a function of age. The arterial stiffness index (ASI) is an established biomarker for vascular aging and is strongly linked to cardiovascular risk factors^46,47. It quantifies the ability of arterial walls to expand and contract in response to blood flow pulses. Diminished arterial compliance, often associated with aging, can be quantified directly from PPG signals using the formula ASI \(=\frac{\,{\rm{height}}}{{\rm{peak-to-peak\; distance}}}\), where peak-to-peak distance is defined as the time interval between the initial systolic peak and the subsequent (or reflected) peak in the pulse waveform, effectively serving as a surrogate for the pulse wave transit time through a patient’s arterial tree. In our analysis, we see a trend where ASI increases as patients age. Figure 3 and Supplementary Information Fig. S3 present summary statistics for the distribution of the arterial stiffness index for the UK Biobank. For instance, the median arterial stiffness index rises from approximately 7.3 m/s at age 40 to nearly 9.9 m/s at age 70. However, the substantial inter-quartile range observed within each age group indicates that age alone does not fully explain the variance seen in arterial characteristics. The notable spread in these values points to the existence of additional factors—such as lifestyle choices and comorbidities—that may significantly modulate arterial stiffness. For example, the third quartile ASI for age 40 at 8.8 m/s is much larger than the median ASI for age 50 at 8.4 m/s. The implication is that ASI, one of the most well-defined and reliably calculated features from a patient’s PPG⁴⁸, is not sufficient to determine a patient’s age to within 3.2 years. To quantify this ceiling directly, we evaluated an ASI-only baseline for predicting chronological age using fivefold cross-validation, fitting (i) a univariate linear model and (ii) a spline-based generalized additive model (GAM) within each training fold and evaluating on the held-out fold. The linear baseline achieved an average MAE of 6.85 years, while the GAM provided a small improvement with an MAE of 6.78 years. Both are only slightly better than predicting the mean age (Table 1). These results provide a reference point for how much age signal is captured by a single, explicitly defined vascular feature derived from PPG.

Analysis 2: deep learning for age prediction

It is possible that Apple’s foundation models^8,21 learned new features or a mix of features to more precisely predict a patient’s age. To explore this possibility, we conducted experiments using two distinct deep learning architectures: SMoLK⁴⁹, a lightweight yet powerful model suitable for wearable devices, and a ResNet-based model⁵⁰, which is commonly used in PPG-related tasks, including classification^31,51 and as the backbone for PPG foundation models⁵². Our guiding hypothesis was that a sophisticated architecture might be able to learn latent features beyond ASI to predict a patient’s age from PPG data. Our results, however, were modest at best compared to Apple’s results, suggesting that even sophisticated deep learning architectures face significant challenges in accurately predicting age from PPG data in isolation.

We show the results of our models in Table 1. These performance metrics indicate that while the models capture some signal related to age, likely related to vascular aging, the magnitude of confounding factors inherent in the dataset severely limits the models’ ability to predict age with greater precision than within half a decade. The details of model training and evaluation are described in the Supplementary Information.

To further understand the nature of the features learned by our models, we employed PaCMAP (Pairwise Controlled Manifold Approximation Projection), a dimensionality reduction technique designed to preserve both local and global structures in high-dimensional datasets⁵³. PaCMAP was used to project the embeddings derived from the trained ResNet and SMoLK models into a two-dimensional space, thereby facilitating visual inspection of how well the age-related signal was separated by these models (Fig. 4). Our PaCMAP visualizations reveal that, although one of the primary axes in the embedding space exhibits a discernible gradient corresponding to age, the overall distribution remains highly overlapped. Thus, while age does have an observable influence on the PPG signals, it does not dominate the learned representations.

Instead, the embeddings may capture the influence of other factors—likely including lifestyle aspects, genetic background, and additional health markers—which collectively introduce substantial confounders into the age prediction latent space. For instance, we know that hypertension and elevated triglyceride levels tend to increase with age and contribute to arterial stiffness and a change in vascular profile⁵⁴, but these do not increase at the same rate for everyone.

Thus, as with all previous age indices, we are able to extract some features that correlate with age, but none are able to predict it to within an average of 3 years, or even 5 years.

There are several differences between our study and Apple’s, some stemming from limitations of public data. First, our empirical analysis relies on the UK Biobank dataset, which, while large, has specific characteristics: the PPG signals are recorded from the finger at rest for a short duration, unlike the continuous wrist-based PPG in the Apple study. This difference in sensor placement (finger versus wrist) introduces variability in signal characteristics as finger PPG often provides a higher fidelity signal and allows for more stable measurements^55,56. Furthermore, the UK Biobank cohort primarily includes individuals between the ages of 40 and 70, while the Apple Heart and Movement Study consisted of patients as young as 18 years old with no upper limit. While we explored several state-of-the-art deep learning architectures (e.g., SMoLK and ResNet), it remains possible that other model architectures or more extensive tuning could yield different results. However, we do not believe these differences alone would allow for the substantially superior age estimation accuracy reported by Apple.

As discussed, Apple did not provide any interpretation of their results as to what features in the PPG signal the model may be leveraging. Our UK Biobank findings raise important questions about whether Apple’s model found a true biological signal of aging or if it capitalized on idiosyncrasies of the Apple Heart and Movement Study. In the absence of Apple releasing its model, data, or any insight into what its model has detected, we cannot dissect its features. But the broader point is that closed models invite skepticism because our results disagree, and alternate explanations cannot be ruled out.

In summary, our analyses using openly available data serve as a reality check on Apple’s claims. This reinforces the need for external model validation. Without that, even well-intentioned studies risk painting an overly optimistic picture. For the field of medical AI to progress and for the public to benefit, we must bridge the gap between isolated, privatized results and generalizable science.

The challenges previously identified are not insurmountable, even if we cannot expect a sea change in transparency anytime soon. We propose several concrete, realistic remedies to foster a more open, trustworthy, and efficient ecosystem for medical AI, as well as some more general remedies.

A few comprehensive open-source benchmark datasets and independent evaluation platforms can have a huge impact

A crucial remedy lies in the strategic development and maintenance of a few high-quality, large-scale, open-source benchmark datasets along with independent evaluation platforms, focusing on specific high-impact health domains. The selection criteria for these domains should prioritize conditions that (1) represent common deadly diseases; (2) allow for meaningful prediction or early detection using data collected before severe disease manifestation; and (3) involve data types that are widely collected and are not difficult to collect, making large-scale dataset creation feasible.

Let us discuss several prime examples meeting these criteria, starting with cardiovascular monitoring for heart arrhythmias such as atrial fibrillation. As discussed, while wearables collect vast data with potential for early arrhythmia detection, proprietary control of these data blocks the independent validation needed to trust life-altering claims – a dangerous paradox given the stakes. A dataset could be easily collected with a concerted effort from a large group of individuals at risk for heart arrhythmia.

A second prime example is medical imaging for cancer screening. Ubiquitous mammography data offers a clear path to AI contributions, yet frustrating access limitations stall independent validation and broader progress. For instance, a recent breakthrough in AI-mammography shows that breast cancer can now be predicted up to 5 years in advance from subtle asymmetries between the patients’ left and right sides⁵⁷; however, little can be done to advance this technology to the clinic without access to large datasets, despite the fact that doing so could potentially prevent numerous cancers and save lives. Mammography data are collected in virtually every large medical center and numerous additional clinics, but they are not generally available to academic researchers.

There are several other examples. Digital pathology requires large datasets to train robust models for analyzing whole-slide images for cancer diagnosis and grading; yet, access is often institutionally siloed. Similarly, developing and validating generalizable, fair models for predicting chronic disease onset or progression using longitudinal electronic health records demands large, diverse EHR datasets. Dermatology could benefit from diverse benchmark datasets to improve AI-based skin cancer detection. Furthermore, standardized public validation datasets are scarce for many at-home monitoring applications, like cuffless blood pressure estimation, sleep staging from wearables, and diabetic foot ulcer imaging, despite the feasibility of frequent and inexpensive data collection.

These benchmark datasets must be accompanied by robust, independent evaluation platforms. There is a clear precedent for empowering neutral agencies like the National Institute of Standards and Technology (NIST) to evaluate high-stakes AI systems. A similar initiative in medical AI, potentially led by NIST, the FDA, or a trusted independent consortium such as the proposed Health AI Assurance Lab⁵⁸, could establish transparent and standardized evaluations with both overall and demographic-specific performance metrics.

Importantly, these platforms and their associated datasets would require investments for regular updates and versioning to reflect evolving clinical practice and prevent overfitting to static benchmarks⁵⁹. Aligning these evaluation standards with regulatory needs for post-market surveillance and managing algorithm changes would further enhance their value. While establishing such infrastructure requires investment, the potential return in terms of accelerated innovation and improved human health could be immense.

Personalized algorithm evaluations can help users choose between AI-health products

While the evaluations discussed above would lead to major health innovations, they do not address patient-specific decisions, for instance, whether you should buy an Apple Watch or a Fitbit. The answer would depend on which brand’s algorithm works better for you. Currently, there is no way for consumers to evaluate that information. It could be incredibly helpful to consumers of AI-health algorithms to be able to upload personal data into an evaluation platform to determine which AI-health product is best for them. In doing this, it is important to protect individual medical data uploaded into such an evaluation platform – no customer’s data should be stored or used for any other purpose than to evaluate algorithms so users can choose between them.

The paradigm of data ownership and sharing needs to change

The current view of data as investigator-owned should be reconsidered. Patients and, by extension, the general public have a stake in the data they contribute. Datasets should generally be treated as shared assets serving broader societal interests, not just proprietary assets for investigators with privileged patient access.

This paradigm shift may require examining the role of Institutional Review Boards (IRBs) in research ethics. One of the IRB’s mandates is to safeguard participants while upholding ethical standards and regulations⁶⁰. Yet a critical question emerges: Is it possible that study participants are treated unethically if they are never informed that their data will be involved in a non-reproducible study? Patients should be able to choose to share their de-identified data to enable scientific reproducibility and further studies. It is also possible that study participants and the public are substantially worse off when either (1) the study is not reproducible, and the results turn out to be wrong; (2) the data is used by only one research group rather than many. Might it be possible, then, that IRBs can inadvertently perform ethics-washing for data-hugging?

In the United States, this dynamic is compounded by the unintended consequences of privacy regulation. HIPAA was originally intended to prevent private health insurance company malfeasance, but it has since become a barrier to data access for legitimate research. This policy regime allows EHR and PACS vendors to keep their versions of collated and extensive datasets proprietary. As we have noted, these very vendors are often the ones making consequential AI mistakes (e.g., Epic’s sepsis algorithm and Philips’ telemetry failures). The result is a paradox: regulations designed to protect patients now shield the organizations whose algorithms may be putting patients at risk from the external scrutiny that could identify and correct these errors.

IRBs could extend ethics oversight beyond study participants to include all stakeholders, such as funders, taxpayers, and the broader public who may be affected by the study’s outcomes. Specifically, IRBs would require actionable data custodianship plans that treat data as a public resource to help prevent data hugging.

IRBs could reduce the restrictive language from informed consent forms and procedural loopholes that sometimes relieve investigators of the responsibility for meaningful data sharing. Consent forms could also default to permit future secondary use, while distinguishing real privacy risks from administrative hurdles. Lastly, institutions should offer standardized data preparation pipelines and vetted templates that ensure compliance with HIPAA⁶¹, GDPR⁶², and FAIR principles⁶³.

The full analysis code (data processing, training, and figure generation) can be found at https://github.com/agk2000/PPG-Age-Reproducibility.

1. Gallifant, J. et al. Artificial intelligence for mechanical ventilation: systematic review of design, reporting standards and bias. Br. J. Anaesth. 128, 343–351 (2021).

2. Wong, A. et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern. Med. 181, 1065–1070 (2021).

   Article  Google Scholar 

3. U.S. Food and Drug Administration. Outpatient telemetry correction: Philips issues correction for monitoring service application used with mobile cardiac outpatient telemetry due to potential for missed information or notifications that may impact patient care. https://www.fda.gov/medical-devices/medical-device-recalls/outpatient-telemetry-correction-philips-issues-correction-monitoring-service-application-used-mobile (2025).

4. Healey, J. S. et al. Subclinical atrial fibrillation and the risk of stroke. N. Engl. J. Med. 366, 120–129 (2012).

   Article  Google Scholar 

5. Perez, M. V. et al. Large-scale assessment of a smartwatch to identify atrial fibrillation. N. Engl. J. Med. 381, 1909–1917 (2019).

   Article  Google Scholar 

6. U.S. Food and Drug Administration. Decision Summary for Apple ECG App (DEN180044). https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN180044.pdf (2018).

7. Pepplinkhuizen, S. et al. Accuracy and clinical relevance of the single-lead apple watch electrocardiogram to identify atrial fibrillation. Cardiovasc. Digital Health J. 3, S17–S22 (2022).

   Article  Google Scholar 

8. Abbaspourazad, S. et al. Large-scale training of foundation models for wearable biosignals. In Proceedings of the International Conference on Learning Representations (ICLR) (2024).

9. Khalid, S. et al. A low-cost PPG sensor-based empirical study on healthy aging based on changes in PPG morphology. Physiol. Meas. 46, 025005 (2025).

10. Asyrafi, H. et al. Photoplethysmography (PPG) signal-based chronological age estimation algorithm. In 2022 International Conference of Science and Information Technology in Smart Administration (ICSINTESA), pp 128–132 (IEEE, 2022).

11. Park, J. et al. Vascular aging estimation based on artificial neural network using photoplethysmogram waveform decomposition: retrospective cohort study. JMIR Med. Inform. 10, e33439 (2022).

    Article  Google Scholar 

12. Abrisham, K. P., Alipour, K., Tarvirdizadeh, B. & Ghamari, M. Vascular age prediction with PPG-based LSTM model. In 2023 9th International Conference on Control, Instrumentation and Automation (ICCIA), pp 1–6 (2023).

13. Chiarelli, A. M. et al. Data-driven assessment of cardiovascular ageing through multisite photoplethysmography and electrocardiography. Med. Eng. Phys. 73, 39–50 (2019).

    Article  Google Scholar 

14. Nie, G., Zhao, Q., Tang, G., Li, Y. & Hong, S. Artificial intelligence-derived photoplethysmography age as a digital biomarker for cardiovascular health. Commun. Med. 5, 481 (2025).

    Article  Google Scholar 

15. Ferdinando, H., Huotari, M. & Myllylä, T. Photoplethysmography signal analysis to assess obesity, age group and hypertension. In 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp 5572–5575 (IEEE, 2019).

16. Shin, H., Noh, G. & Choi, B.-M. Photoplethysmogram based vascular aging assessment using the deep convolutional neural network. Sci. Rep. 12, 11377 (2022).

    Article  Google Scholar 

17. Shin, H. Xgboost regression of the most significant photoplethysmogram features for assessing vascular aging. IEEE J. Biomed. Health Inform. 26, 3354–3361 (2022).

    Article  Google Scholar 

18. Nie, G., Zhao, Q., Tang, G., Li, J. & Hong, S. Deep imbalanced regression to estimate vascular age from PPG data: a novel digital biomarker for cardiovascular health. Preprint at arXiv https://doi.org/10.48550/arXiv.2406.14953 (2024).

19. Lombardi, S., Tavernise, F., Francia, P. & Bocchi, L. Analysis of age-related variations in photoplethysmography: A machine learning approach. In European Medical and Biological Engineering Conference, pp 95–105 (Springer, 2024).

20. Barthels, M., De Cooman, T., Gruwez, H., Duncker, D. & Vandervoort, P. Predicting age from a real world smartphone acquired PPG signal using deep learning (2024).

21. Erturk, E. et al. Beyond sensor data: Foundation models of behavioral data from wearables improve health predictions. In Proc. International Conference on Machine Learning (ICML) (2025).

22. Miller, A. C. et al. A wearable-based aging clock associates with disease and behavior. Nat. Commun. 16, 9264 (2025).

    Article  Google Scholar 

23. Horvath, S. DNA methylation age of human tissues and cell types. Genome Biol. 14, 1–20 (2013).

    Article  Google Scholar 

24. Sudlow, C. et al. UK Biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 12, e1001779 (2015).

    Article  Google Scholar 

25. Sambasivan, N. et al. “Everyone wants to do the model work, not the data work”: Data cascades in high-stakes AI. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pp 1–15 (2021).

26. Grother, P. J. & Ngan, M. Face recognition vendor test (FRVT), report 8009. Technical Report (Information Access Division, National Institute of Standards and Technology, 2014).

27. Grother, P., Ngan, M. & Hanaoka, K. Face recognition vendor test (FRVT) part 3: Demographic effects, report 8280. Technical Report (Information Access Division, National Institute of Standards and Technology, 2019).

28. Martin, A. F. & Przybocki, M. A. The NIST speaker recognition evaluations: 1996–2001. In Odyssey, pp 39–43 (2001).

29. National Institute of Standards and Technology. NIST speaker recognition evaluation (SRE) program overview https://sre.nist.gov/ (2024).

30. Ding, C. et al. Learning from alarms: a robust learning approach for accurate photoplethysmography-based atrial fibrillation detection using eight million samples labeled with imprecise arrhythmia alarms. IEEE J. Biomed. Health Inform. 28, 2650–2661 (2024).

    Article  Google Scholar 

31. Torres-Soto, J. & Ashley, E. A. Multi-task deep learning for cardiac rhythm detection in wearable devices. npj Digital Med. 3, 116 (2020).

    Article  Google Scholar 

32. Bycroft, C. et al. The UK Biobank resource with deep phenotyping and genomic data. Nature 562, 203–209 (2018).

    Article  Google Scholar 

33. Johnson, A. E. et al. MIMIC-III, a freely accessible critical care database. Sci. Data 3, 1–9 (2016).

    Article  Google Scholar 

34. Schmidt, P., Reiss, A., Duerichen, R., Marberger, C. & Van Laerhoven, K. Introducing wesad, a multimodal dataset for wearable stress and affect detection. In Proceedings of the 20th ACM International Conference on Multimodal Interaction, pp 400–408 (2018).

35. Reiss, A., Indlekofer, I., Schmidt, P. & Van Laerhoven, K. Deep PPG: large-scale heart rate estimation with convolutional neural networks. Sensors 19, 3079 (2019).

    Article  Google Scholar 

36. Dunn, J. et al. Wearable sensors enable personalized predictions of clinical laboratory measurements. Nat. Med. 27, 1105–1112 (2021).

    Article  Google Scholar 

37. Bent, B. et al. The digital biomarker discovery pipeline: an open-source software platform for the development of digital biomarkers using mHealth and wearables data. J. Clin. Transl. Sci. 5, e19 (2021).

    Article  Google Scholar 

38. Lee, H.-C. et al. Vitaldb, a high-fidelity multi-parameter vital signs database in surgical patients. Sci. Data 9, 279 (2022).

    Article  Google Scholar 

39. Zhang, G.-Q. et al. The national sleep research resource: towards a sleep data commons. J. Am. Med. Inform. Assoc. 25, 1351–1358 (2018).

    Article  Google Scholar 

40. Haas, D. M. et al. A description of the methods of the nulliparous pregnancy outcomes study: monitoring mothers-to-be (nuMoM2b). Am. J. Obstet. Gynecol. 212, 539–e1 (2015).

    Article  Google Scholar 

41. Pillai, A., Spathis, D., Kawsar, F. & Malekzadeh, M. Papagei: open foundation models for optical physiological signals. In Proc. Thirteenth International Conference on Learning Representations (ICLR) (2025).

42. Ibrahim, H., Liu, X., Zariffa, N., Morris, A. D. & Denniston, A. K. Health data poverty: an assailable barrier to equitable digital health care. Lancet Digital Health 3, e260–e265 (2021).

    Article  Google Scholar 

43. Kan-Tor, Y., Ben-Meir, A. & Buxboim, A. Can deep learning automatically predict fetal heart pregnancy with almost perfect accuracy? Hum. Reprod. 35, 1473 (2020).

    Article  Google Scholar 

44. Daneshjou, R. et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Sci. Adv. 8, eabq6147 (2022).

    Article  Google Scholar 

45. Beam, A. L., Manrai, A. K. & Ghassemi, M. Challenges to the reproducibility of machine learning models in health care. J. Am. Med. Assoc. 323, 305–306 (2020).

    Article  Google Scholar 

46. Choi, S.-Y. et al. Age-associated increase in arterial stiffness measured according to the cardio-ankle vascular index without blood pressure changes in healthy adults. J. Atheroscler. Thromb. 20, 911–923 (2013).

    Article  Google Scholar 

47. Mitchell, G. F. Arterial stiffness and wave reflection: biomarkers of cardiovascular risk. Artery Res. 3, 56–64 (2009).

    Article  Google Scholar 

48. Mitchell, G. F. Arterial stiffness in aging: does it have a place in clinical practice? recent advances in hypertension. Hypertension 77, 768–780 (2021).

    Article  Google Scholar 

49. Chen, S. F., Guo, Z., Ding, C., Hu, X. & Rudin, C. Sparse learned kernels for interpretable and efficient medical time series processing. Nat. Mach. Intell. 6, 1132–1144 (2024).

    Article  Google Scholar 

50. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 770–778 (2016).

51. Guo, Z. et al. Improving atrial fibrillation detection using a shared latent space for ECG and PPG signals. Harvard Data Sci. Rev. 7, (2025).

52. Ding, C. et al. Siamquality: a convnet-based foundation model for photoplethysmography signals. Physiol. Meas. 45, 085004 (2024).

    Article  Google Scholar 

53. Wang, Y., Huang, H., Rudin, C. & Shaposhnik, Y. Understanding how dimension reduction tools work: an empirical approach to deciphering t-SNE, UMAP, TriMAP, and PaCMAP for data visualization. J. Mach. Learn. Res. 22, 1–73 (2021).

    Google Scholar 

54. Vallée, A. Arterial stiffness determinants for primary cardiovascular prevention among healthy participants. J. Clin. Med. 11, 2512 (2022).

    Article  Google Scholar 

55. Charlton, P. H. et al. The 2023 wearable photoplethysmography roadmap. Physiol. Meas. 44, 111001 (2023).

    Article  Google Scholar 

56. Allen, J. Photoplethysmography and its application in clinical physiological measurement. Physiol. Meas. 28, R1 (2007).

    Article  Google Scholar 

57. Donnelly, J., Moffett, L., Barnett, A. J., Trivedi, H., Schwartz, F., Lo, J. & Rudin, C. AsymMirai: Interpretable Mammography-based Deep Learning Model for 1–5-year Breast Cancer Risk Prediction. Radiology 310, e232780 (2024).

58. Shah, N. H. et al. A nationwide network of health AI assurance laboratories. J. Am. Med. Assoc. 331, 245–249 (2024).

    Article  Google Scholar 

59. Recht, B., Roelofs, R., Schmidt, L. & Shankar, V. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pp 5389–5400 (PMLR, 2019).

60. U.S. Department of Health and Human Services. Lesson 1: What are IRBs? https://www.hhs.gov/ohrp/education-and-outreach/online-education/human-research-protection-training/lesson-3-what-are-irbs/index.html.

61. U.S. Department of Health and Human Services. Guidance regarding methods for de-identification of protected health information in accordance with the HIPAA privacy rule. https://www.hhs.gov/hipaa/for-professionals/privacy/special-topics/de-identification/index.html (2012).

62. European Union. General data protection regulation (GDPR)—official legal text. https://gdpr-info.eu/ (2016).

63. Wilkinson, M. D. et al. The fair guiding principles for scientific data management and stewardship. Sci. Data 3, 1–9 (2016).

    Article  Google Scholar 

Download references

Authors and Affiliations

1. Medical Scientist Training Program, Duke University School of Medicine, Durham, NC, USA

   Anish Karpurapu

2. Department of Computer Science, Duke University, Durham, NC, USA

   Zhicheng Guo & Cynthia Rudin

3. Nell Hodgson Woodruff School of Nursing, Emory University, Atlanta, GA, USA

   Xiao Hu

Contributions

A.K. executed all analyses, performed statistical evaluation, generated figures and tables, and wrote the main paper text. Z.G. assisted with analyses, data quality checks, replication of analyses, and paper editing. C.R. and X.H. conceived the study, provided supervision, contributed to interpretation, and revised the paper. All authors read and approved the paper.

Corresponding authors

Correspondence to Anish Karpurapu or Cynthia Rudin.

Competing interests

The authors declare no competing interests.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions