PPG Foundation Models Generalize Within Cohorts but Not Yet Across Them

Question: How have self-supervised foundation models built on raw photoplethysmography waveforms been evaluated for generalization across external cohorts and downstream physiological or cardiometabolic prediction tasks?

Answer

Taken together, the retrieved papers show an evaluation literature that is broad in task coverage and narrow in cohort coverage, and this asymmetry is what inflates the apparent generalization of PPG foundation models. The dominant design is a large unlabelled pretraining corpus followed by many downstream tasks evaluated under subject-disjoint cross-validation within a single dataset or dataset family — 17 of 19 tasks from ~32,000 smartwatch users [1], 18 tasks across eight datasets [16], 14 of 15 tasks after ICU pretraining [2], six tasks across five public datasets under frozen linear probing [13]. Where a single open benchmark puts five public PPG foundation models against each other on nine datasets and 26 tasks with simple non-learned controls, the advantage largely evaporates: 35.3% of 68 task–regime combinations show no meaningful improvement over a random/constant baseline, and in 83.8% no PPG-specific foundation model outperforms both a deterministic pulse-rate-variability feature set and a generic time-series foundation model [6]. The clearest evidence of real cross-cohort transport comes from one model evaluated against an external out-of-distribution cohort and then applied unchanged to 3,721 ICU patients [17], and from a model pretrained on ICU waveforms that improved 14 of 15 tasks including field-like consumer-grade data [2]. Physiological endpoints (heart rate, blood pressure, SpO₂, labs) are the ones actually shown to move; cardiometabolic and hard cardiovascular outcomes are still carried by supervised, purpose-built models in single biobank cohorts rather than by frozen self-supervised embeddings.

Task breadth substitutes for cohort breadth in most published evaluations

The standard claim of generalizable representation is supported by task multiplicity rather than by cohort multiplicity. Masked multiscale wavelet reconstruction, pretrained on ~17 million unlabelled 10-second segments from ~32,000 smartwatch users, improves over or matches state-of-the-art open-source PPG foundation models, time-series foundation models, and other self-supervised baselines on 17 of 19 health-related tasks [1]. Duration-grounded Retain-and-Extend pretraining likewise reports the best observed score on 12 of 18 tasks from eight datasets against five existing PPG foundation models [16]. Retrieval-augmented ECG-linked representation augmentation is evaluated on five public datasets covering six downstream tasks under a unified frozen-feature linear-probing protocol [13]. In each case the evidence is a wide task sweep, and the number of genuinely independent cohorts drawn from different devices, care settings, and populations is far smaller than the number of tasks.

Pretraining-corpus choice has become the explicit lever for field generalization

A stated motivation in this literature is that ICU-derived pretraining data and its curation requirements complicate generalization to field-like data, and the proposed remedy is to use accompanying ECG and respiratory signals to select contrastive samples so that noisy PPG segments are retained rather than discarded, yielding improvements on 14 of 15 diverse downstream tasks including field-like daily activity and heart rate prediction with three times fewer pretraining subjects than prior state of the art [2]. The complementary strategy is to pretrain directly on consumer-scale field data and then transfer inward: a model trained on nearly seven million pulse-oximetry segments from 42,282 overnight sleep recordings in 10,704 adults was applied unchanged to ICU patients, where two-minute windows from the first day of admission improved survival prediction beyond age, sex, and the heart rate and oxygen saturation the same device reports [17]. Both results indicate that the pretraining distribution, not the architecture, is doing most of the transport work — which means "generalization" claims are not comparable across papers with different pretraining corpora.

Genuine external-cohort claims are usually anchored to a proxy target, most often chronological age

The most-cited cross-cohort demonstration uses chronological age as the global health benchmark, with the model outperforming existing open-source and proprietary feature extraction methods and showing robust generalization in an external out-of-distribution cohort before extending to 71 phenotypic targets spanning cardiometabolic and neuropsychiatric domains [17]. Age is a well-powered, universally available label that is strongly correlated with almost every physiological parameter, so it validates representation quality without testing transport of a clinical decision. The cardiometabolic follow-through in that work is within-individual tracking of next-day glycemic, dietary and activity state rather than incident disease prediction.

Cardiometabolic and cardiovascular outcomes are still delivered by supervised models in single cohorts

Where hard cardiometabolic endpoints have been predicted from raw PPG, the models are supervised and cohort-specific rather than self-supervised and transported. A multimodal framework integrating raw PPG waveforms with haemodynamic features in UK Biobank participants built an arterial aging clock whose residual, ArtAgeGap, correlated with pulse pressure (r=0.60), was elevated in diabetes (+0.93 years) and hyperlipidaemia (+0.49), and predicted incident hypertension, major adverse cardiovascular events, and cardiovascular and all-cause mortality in 96,615 participants followed for a median of 13 years, with a genome-wide association study in 114,098 individuals [9]. That is a strong cardiometabolic result, but it is a supervised clock trained and tested within one biobank, not a frozen self-supervised encoder evaluated out of cohort. Similarly, PPG foundation model embeddings have been used to estimate clinical laboratory trajectories across two ICU datasets with improvements over LSTM and carry-forward baselines [3] — a physiological rather than cardiometabolic target, and again an ICU-internal evaluation.

Discrimination is reported; calibration and subgroup performance are not

The axes on which waveform models most reliably fail are the ones least often measured. In an external validation of a pressure-derived haemodynamic predictor, discrimination transferred and was in fact larger than in development (AUROC 0.696 vs 0.638) while calibration did not transfer at all (external Brier skill −0.014), requiring local recalibration [11]. Subgroup degradation is documented where anyone looks: across four wrist-worn devices in 58 Hispanic adults with Fitzpatrick III–V skin, all devices deviated from criterion heart rate and higher BMI and darker skin tone were associated with increased error [15]; in subject-independent validation on 1.26 million segments from 2,154 perioperative patients, all deep learning models showed significant performance variation across demographic and cardiovascular conditions [14]. A related warning comes from PPG-to-camera-rPPG work showing that population-level plausibility does not establish recording-level preservation of physiological properties, with Fitzpatrick-associated heterogeneity in endpoint discrepancy that aggregate signal-to-noise did not explain [4]. None of the retrieved foundation-model papers reports calibration slope or demographic subgroup metrics for their clinical endpoints.

Conflicts and gaps

The central conflict is whether a PPG-specific foundation-model advantage exists at all. The unified benchmark finds no PPG-specific model beating both PRV and a generic time-series foundation model in 83.8% of comparisons [6], while model papers report wins on most of their own task suites against "state-of-the-art open-source PPG foundation models" [1] [16]. These claims have not been reconciled on shared tasks, shared splits, or shared baselines. A second conflict concerns whether blood pressure is recoverable from PPG in principle: a calibration-based Siamese ResNet benchmark found that normalized invasive arterial pressure meets AAMI standards (1.29±6.33 mmHg systolic) while normalized PPG does not (1.49±11.82 mmHg), concluding PPG contains information correlated with but possibly insufficient for BP [10]; a separate large MIMIC-III study reports direct PPG-to-BP prediction reaching BHS Grade A with MAE 4.82/4.31 mmHg across 1.74 million segments from 3,127 patients, outperforming ECG-mediated pipelines [8]. Third, no retrieved paper reports prospective deployment, and a scoping review of 460 deep-learning PPG papers identifies insufficient validation in real-world environments as an unresolved field-wide limitation [12]. Finally, the statistical framing of transport is itself underspecified: transporting a marker "as is" implicitly assumes predictive values are transportable, and conventional prevalence adjustment shifts the assumption onto sensitivity and specificity, with both relying on strong assumptions about the stability of cause distributions across populations [5].

What would change the answer

The decisive experiment is a pre-registered, cross-corpus transport test in the PPG-Bench mould: pretrain or take frozen encoders from the two dominant corpora (overnight consumer pulse oximetry and ICU waveforms), then evaluate all of them on a third cohort's clinical endpoints — for example, MMR and PulseOx-FM embeddings applied to UK Biobank PPG waveforms with incident hypertension and MACE as outcomes, and ICU-pretrained embeddings applied to the same endpoints — with PRV and MOMENT feature sets as mandatory controls, subject-disjoint and site-disjoint splits, and label-shuffle negative controls [6] [1] [17] [9]. Reporting calibration slope and intercept alongside AUROC, plus error stratified by Fitzpatrick scale, BMI, age and device, would reveal whether the physiological-endpoint gains survive case-mix shift in the way heart-rate error demonstrably does not [11] [15]. A direct head-to-head between frozen self-supervised embeddings and the supervised ArtAgeGap model on the same UK Biobank endpoint and split would settle whether self-supervision adds anything beyond supervised waveform regression for cardiometabolic risk [9]. Finally, a device- and site-identification probe on the learned embeddings — high accuracy implying the encoder is capturing acquisition hardware rather than physiology — would provide a cheap diagnostic for the corpus-confounding that currently makes cross-paper generalization claims uninterpretable, and would test whether the "unseen domain" behaviour reported for generative cardiovascular models reflects physiology or sensor signature [7].

References

[1] Megha Thukral, Cyrus Tanade, Simon A. Lee, Juhyeon Lee, Hao Zhou, Keum San Chun, Migyeong Gwak, Viswam Nathan, Md Mahbubur Rahman, Li Zhu, Mehrab Bin Morshed, Subramaniam Venkatraman, Sharanya Arcot Desai. "Wavelet-Driven Masked Multiscale Reconstruction for PPG Foundation Models". arxiv. 2026-01-21T00:00:00.000Z. URL: https://arxiv.org/abs/2601.12215

[2] Eloy Geenjaar, Vince Calhoun, Scott Daly, Gouthaman KV, Lie Lu, Trisha Mittal, Daniel P. Darcy. "A robust PPG foundation model using multimodal physiological supervision". arxiv. 2026-06-08T00:00:00.000Z. URL: https://arxiv.org/abs/2606.07365

[3] Minxiao Wang, Runze Yan, Carol Li, Saurabh Kataria, Xiao Hu, Matthew Clark, Timothy Ruchti, Timothy G. Buchman, Sivasubramanium V Bhavani, Randall J. Lee. "Estimating Clinical Lab Test Result Trajectories from PPG using Physiological Foundation Model and Patient-Aware State Space Model -- a UNIPHY+ Approach". arxiv. 2025-09-23T00:00:00.000Z. URL: https://arxiv.org/abs/2509.16345

[4] Timothy Oladunni, Farouk Ganiyu-Adewumi. "Property-Specific Recoverability from Contact PPG to Camera rPPG under Heterogeneous Observation Conditions". arxiv. 2026-08-28T00:00:00.000Z. URL: https://arxiv.org/abs/2608.27392

[5] Mohsen Sadatsafavi, Gavin Pereira, Wenjia Chen. "Transportability of Prognostic Markers: Rethinking Common Practices through a Sufficient-Component-Cause Perspective". arxiv. 2026-09-23T00:00:00.000Z. URL: https://doi.org/10.1097/EDE.0000000000002044

[6] Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra, Sandeep Singh, Siddharth Panwar. "PPG-Bench: A Unified Benchmark for PPG Foundation Models". arxiv. 2026-10-12T00:00:00.000Z. URL: https://arxiv.org/abs/2610.13435

[7] Zehua Chen, Yuyang Miao, Liyuan Wang, Luyun Fan, Danilo P. Mandic, Jun Zhu. "Versatile Cardiovascular Signal Generation with a Unified Diffusion Transformer". arxiv. 2025-08-22T00:00:00.000Z. URL: https://arxiv.org/abs/2505.22306

[8] Bo Wu, Haoling Wang, Zhuodiao Kuang, Kateryna Shapovalenko. "Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines". arxiv. 2026-07-28T00:00:00.000Z. URL: https://arxiv.org/abs/2607.23406

[9] Xu, L., Zhang, L., Ray, A., Omarov, M., Malik, R., Beaufort, N., Dichgans, M., Bragazzi, N. L., Georgakis, M. K.. "A photoplethysmography-based aging clock reveals genetic determinants of arterial aging". medRxiv. 2025-11-17T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.15.25340310

[10] Felipe M. Dias, Diego A. C. Cardenas, Marcelo A. F. Toledo, Filipe A. C. Oliveira, Estela Ribeiro, Jose E. Krieger, Marco A. Gutierrez. "Exploring the limitations of blood pressure estimation using the photoplethysmography signal". arxiv. 2025-04-22T00:00:00.000Z. URL: https://doi.org/10.1088/1361-6579/adcb86

[11] Oyarzun, R., Hernandez, P.. "The dynamics of arterial pressure itself predict intraoperative hypotension beyond its current value: an interpretable additive model validated in 3,069 external patients under a selection-bias-resistant protocol". medRxiv. 2026-08-31T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.26.26361468

[12] Guangkun Nie, Jiabao Zhu, Gongzheng Tang, Deyun Zhang, Shijia Geng, Qinghao Zhao, Shenda Hong. "A Scoping Review of Deep Learning Methods for Photoplethysmography Data". arxiv. 2026-05-06T00:00:00.000Z. URL: https://arxiv.org/abs/2401.12783

[13] Zhongli Wu, Zhuangzhi Gao, He Zhao, Feixiang Zhou, Fu Wang, Jinru Ding, Yuankai Wang, Hongyi Qin, Gregory Y. H. Lip, Bil Kirmani, Yalin Zheng. "P2E-VQ: ECG-linked representation augmentation for PPG via discrete patch retrieval". arxiv. 2026-08-18T00:00:00.000Z. URL: https://arxiv.org/abs/2608.14656

[14] Wentao Li, Yonghu He, Zirong Yu, Kun Gao, Qing Liu, Yali Zheng. "Collaborative Learning-Enhanced Lightweight Models for Predicting Arterial Blood Pressure Waveform in a Large-scale Perioperative Dataset". arxiv. 2026-01-27T00:00:00.000Z. URL: https://arxiv.org/abs/2508.11669

[15] Kostrna, Jason, Oparina, Ekaterina, Palacios, Cristina, Rodriguez, Andres J, Pei, JunZhu, Ajmal, Ajmal, Ramella-Roman, Jessica C. "PPG-Based Heart Rate Accuracy in Diverse Populations: Investigating Inequities Across Body Composition and Skin Tones". arxiv. 2026-02-02T00:00:00.000Z. URL: https://arxiv.org/abs/2601.22377

[16] Suyeong Lee, Hochang Lee, Seokyong Sheem, Daekyum Kim. "RAE-PPG: Duration-Grounded Retain-and-Extend Pretraining for PPG Foundation Models". arxiv. 2026-09-30T00:00:00.000Z. URL: https://arxiv.org/abs/2609.36794

[17] Kohn, S., Lutsker, G., Diament, A., Shilo, S., Gabet, A., Sasson, G., Wolf, G., Wolf, A., Godneva, A., Weinberger, A., Rossman, H., Segal, E.. "A foundation model of wearable pulse oximetry reveals physiological signatures of health and cardiometabolic risk". medRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.07.01.26356992

How this survey was made

This survey answers one question: “How have self-supervised foundation models built on raw photoplethysmography waveforms been evaluated for generalization across external cohorts and downstream physiological or cardiometabolic prediction tasks?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 17 papers whose work this text draws on, chosen from the 162 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 21, 2026 and covered work from the previous two years.