MANNAS
Blog

PPGs and the FM Gold Rush

Arvasu Kulkarni · 29 October 2026 · 6 min

Introduction

The past decade, and in particular the last 5 years have seen a transformation in AI research in almost all fields. Research on models for many modalities has converged onto the same idea—training large, self-supervised (SSL) models on large unlabelled datasets to provide generalisable representations of data. These models are called foundation models1. The revolution started in text and images with BERT2, CLIP3, and GPT-34. Since, it has spread like wildfire to areas like speech recognition5,6 and OCR7, and even to financial transaction modelling8.

Researchers working with any type of health data have been early adopters of AI and deep learning methods to solve their own problems, usually by borrowing and adapting methods and trends that have worked in the traditional ML domains (text and images) to their modality. The FM trend has now caught on in many modalities. Protein design and analysis is perhaps the best example, where classical ML methods were frontrunners for decades before SSL protein language models like ESM9,10 brought a step-change in performance. Pathology11 and radiology12 are following suit.

The same thing is happening for biosignal time series data like EEG13, ECG14, EMG15, and fNIRS16. Each modality has seen the development of FMs, with the goal of better brain or cardiac health monitoring. The subject of this article though, is another modality, photoplethysmography (PPG).

PPG

Photoplethysmography signals are collected by shining optical light at a specific wavelength into the skin, and measuring the intensity of the light that passes through to the other side (of say, a fingertip) or reflects back to the sensor. How much light transmits or reflects is chiefly determined by the amount of blood it passes through. The volume of blood in a cross section of skin changes with cardiac rhythm17. The technique is entirely non-invasive and the sensors are cheap to make, which has led to extremely widespread integration of PPG sensors into wearable devices18.

watchblood vesselsPPG signal
Figure 1. A watch’s LED shines light into the wrist and a sensor beside it measures how much comes back. Each heartbeat sends a surge of blood through the small vessels under the skin, changing that amount; over time, it traces the PPG signal.

A normal PPG signal looks like Figure 1, oscillating up and down in time with your pulse. Simple signal processing and a peak detection algorithm can be used to calculate heart rate at short intervals from the signal. This is what happens when your smartwatch tells you your heart rate19. Changes in your heart rate, particularly the distribution of inter-beat intervals (IBIs) can be measured in a similar way and can be informative about stress levels, sleep quality20, and can pick up signs of irregular heart beats (arrhythmias)21. Two concurrent PPG signals, at two different wavelengths, can tell you about the oxygen concentration of your blood (SpO2)22. Beyond this, research efforts have gone into using PPG signals for blood pressure23, blood glucose24, and haemoglobin25. However, most models trained on PPG data fail to transfer well to unseen real-world datasets and new subjects18. This problem occurs across the spectrum of tasks and is a common issue in biosignal modelling in general.

Failure to generalize has led a part of the PPG research community to develop Foundation Models, following in the footsteps of ECG14, fMRI26, and EEG13. The first real PPG-FM to come out was PaPaGei27, released by a team from Cambridge and Dartmouth in early 2025. Using a contrastive objective and pretraining over 57k hours from 13k subjects, they demonstrated downstream performance superior to other generic pretrained time series foundation models, like MOMENT28 and Chronos29. This was followed by models like GPT-PPG (2025)30 which trained on more than 1.6 million hours of data and showed superiority over other self-supervised PPG models, and later AnyPPG31, PulsePPG32, and SigmaPPG33 would come out, demonstrating superiority within the FM group.

PPG-Bench

In our recent paper34 (now accepted at the NeurIPS ’26 Brain & Body FM Workshop), we set out to test the claims being made by PPG FMs and to see exactly how large the gap is between deployment ready models and the current generation. We included 26 tasks across nine open access datasets and evaluated 5 FMs, and have released the benchmark open-source, so teams developing PPG models can easily do independent evaluations of their models without setting up an internal testing system.

abc
1in3settings, no FM beats chance
5in6settings, no FM beats all three
Figure 2. The three controls each FM was tested against, (a) chance, (b) classical PRV features and (c) MOMENT, a generic time series FM.

Most importantly, we evaluated the FMs against three increasingly strong controls, rather than just against each other. The latter showed that AnyPPG was generally the strongest FM, but the former showed a serious concern in overall performance. The 3 controls we tested were a) a random chance / constant baseline, i.e., the dumbest algorithm you could make with knowledge of the training data, b) a classical Pulse-Rate Variability (PRV) algorithm that uses basic deterministic features of the signal20, and c) MOMENT, a generic time series FM28. Models that cannot beat (a) are essentially useless, and any comparison between them will generally be comparing random noise; those that cannot beat (b) are not viable replacements for currently existing classical methods (and are not producing representations that add value beyond basic physiological markers), and those that cannot beat (c) are not adding value by pretraining on PPG data.

Concerningly, we found that across 68 downstream settings, 35.3% showed no meaningful FM advantage over the chance baseline, and there are very few cases where even one model outperforms all 3 controls, only 16.2%. This means that only in one out of six cases would using any PPG-FM be justified.

I would weakly speculate that it is possible that FMs are not the way to go for PPG modelling. PPGs are inherently low-information compared to other biosignal modalities like EEGs or ECGs, and simply might not have enough information available to be extracted by large self supervised models that cannot be extracted by smaller (but no less complex) supervised approaches. There are two reasons I think this may be true - a) the results from PPG-Bench are somewhat discouraging, and b) PPG as a biosignal modality, is blessed with an enormous scale of data in a way that no other modality is. Existing corpora are already quite large30,35, and real-world data collection is fairly easy as well—if FMs are the right approach, I believe most teams already have roughly (on the same OOM) the required volume of data to have made them much better than they currently are. One sign comes from Google’s LSM, a multimodal foundation model specifically for wearables36. LSM is trained on simultaneous HR, HRV (heart rate variability), EDA, skin temperature, and accelerometer data. The first two are derived from PPG and fed into the model, which may be a more stable approach compared to feeding in raw PPG signals.

It is important to note though, that the general public may not have all the information on this subject. Big players in the wearables market like Apple, Google, and Samsung may already be using PPG FMs to some extent (though there is no indication in either direction of this). Apple even released a study in 2024 detailing the development of a PPG FM and its applications in clinical markers35. Additionally, multi-modal cardiac FMs have shown promise as well31, which may be a better direction than developing PPG-only FMs.

Conclusion

The current generation of PPG FMs has made impressive advances in scale and representation learning, but it remains unclear whether these advances translate into meaningful improvements in practice. Our results suggest that the field has focused too heavily on comparisons between FMs, and not enough on whether these models offer any advantage over simpler alternatives. A model that outperforms its peers but cannot reliably beat a trivial baseline or established signal-processing methods is not a meaningful step forward.

We believe that the next phase of PPG research should be driven not by the race to build the largest or best-performing FM, but by the pursuit of reliable, generalisable and deployable solutions to real physiological problems. PPG-Bench is a step towards making those claims easier to evaluate. Ultimately, the success of PPG FMs should be measured not by how well they perform against one another, but by what they enable us to do that we could not already do as well, more efficiently, or more reliably without them.

References

  1. Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. On the Opportunities and Risks of Foundation Models. arXiv preprint, arXiv:2108.07258, 2021.
  2. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT), 4171–4186, 2019.
  3. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning Transferable Visual Models From Natural Language Supervision. Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, 139, 8748–8763, 2021.
  4. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D. Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems, 33, 1877–1901, 2020.
  5. Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. Advances in Neural Information Processing Systems, 33, 12449–12460, 2020.
  6. Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., and Sutskever, I. Robust Speech Recognition via Large-Scale Weak Supervision. Proceedings of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, 202, 28492–28518, 2023.
  7. Li, M., Lv, T., Chen, J., Cui, L., Lu, Y., Florencio, D., Zhang, C., Li, Z., and Wei, F. TrOCR: Transformer-Based Optical Character Recognition with Pre-trained Models. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11), 13094–13102, 2023.
  8. Stripe. Applied AI: Delivering on growth. Stripe Sessions, 2025. https://stripe.com/sessions/2025/applied-ai-delivering-on-growth
  9. Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C. L., Ma, J., and Fergus, R. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proceedings of the National Academy of Sciences, 118(15), e2016239118, 2021.
  10. Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., dos Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Candido, S., and Rives, A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637), 1123–1130, 2023.
  11. Chen, R. J., Ding, T., Lu, M. Y., Williamson, D. F. K., Jaume, G., Song, A. H., Chen, B., Zhang, A., Shao, D., Shaban, M., Williams, M., Oldenburg, L., Weishaupt, L. L., Wang, J. J., Vaidya, A., Le, L. P., Gerber, G., Sahai, S., Williams, W., and Mahmood, F. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30, 850–862, 2024.
  12. Pérez-García, F., Sharma, H., Bond-Taylor, S., Bouzid, K., Salvatelli, V., Ilse, M., Bannur, S., Castro, D. C., Schwaighofer, A., Lungren, M. P., Wetscherek, M. T., Codella, N., Hyland, S. L., Alvarez-Valle, J., and Oktay, O. Exploring scalable medical image encoders beyond text supervision. Nature Machine Intelligence, 7, 119–130, 2025.
  13. Jiang, W.-B., Zhao, L.-M., and Lu, B.-L. Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI. International Conference on Learning Representations (ICLR), 2024.
  14. McKeen, K., Masood, S., Toma, A., Rubin, B., and Wang, B. ECG-FM: an open electrocardiogram foundation model. JAMIA Open, 8(5), ooaf122, 2025.
  15. Fasulo, M., Spacone, G., Ingolfsson, T. M., Li, Y., Benini, L., and Cossettini, A. TinyMyo: a Tiny Foundation Model for Flexible EMG Signal Processing at the Edge. arXiv preprint, arXiv:2512.15729, 2025.
  16. Subedi, A., De, S., Cavuoto, L., Schwaitzberg, S., Hackett, M., and Norfleet, J. An Interpretable Transformer-Based Foundation Model for Cross-Procedural Skill Assessment Using Raw fNIRS Signals. arXiv preprint, arXiv:2506.22476, 2025.
  17. Allen, J. Photoplethysmography and its application in clinical physiological measurement. Physiological Measurement, 28(3), R1–R39, 2007.
  18. Charlton, P. H., Allen, J., Bailón, R., Baker, S., Behar, J. A., Chen, F., Clifford, G. D., Clifton, D. A., Davies, H. J., Ding, C., Ding, X., Dunn, J., Elgendi, M., Ferdoushi, M., Franklin, D., Gil, E., Hassan, M. F., Hernesniemi, J., Hu, X., Ji, N., Khan, Y., Kontaxis, S., Korhonen, I., Kyriacou, P. A., Laguna, P., Lázaro, J., Lee, C., Levy, J., Li, Y., Liu, C., Liu, J., Lu, L., Mandic, D. P., Marozas, V., Mejía-Mejía, E., Mukkamala, R., Nitzan, M., Pereira, T., Poon, C. C. Y., Ramella-Roman, J. C., Saarinen, H., Shandhi, M. M. H., Shin, H., Stansby, G., Tamura, T., Vehkaoja, A., Wang, W. K., Zhang, Y.-T., Zhao, N., Zheng, D., and Zhu, T. The 2023 wearable photoplethysmography roadmap. Physiological Measurement, 44(11), 111001, 2023.
  19. Reiss, A., Indlekofer, I., Schmidt, P., and Van Laerhoven, K. Deep PPG: Large-Scale Heart Rate Estimation with Convolutional Neural Networks. Sensors, 19(14), 3079, 2019.
  20. Mejía-Mejía, E., May, J. M., Torres, R., and Kyriacou, P. A. Pulse rate variability in cardiovascular health: a review on its applications and relationship with heart rate variability. Physiological Measurement, 41(7), 07TR01, 2020.
  21. Perez, M. V., Mahaffey, K. W., Hedlin, H., Rumsfeld, J. S., Garcia, A., Ferris, T., Balasubramanian, V., Russo, A. M., Rajmane, A., Cheung, L., Hung, G., Lee, J., Kowey, P., Talati, N., Nag, D., Gummidipundi, S. E., Beatty, A., Hills, M. T., Desai, S., Granger, C. B., Desai, M., and Turakhia, M. P. Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation. New England Journal of Medicine, 381(20), 1909–1917, 2019.
  22. Chan, E. D., Chan, M. M., and Chan, M. M. Pulse oximetry: understanding its basic principles facilitates appreciation of its limitations. Respiratory Medicine, 107(6), 789–799, 2013.
  23. Liang, Y., Chen, Z., Liu, G., and Elgendi, M. A new, short-recorded photoplethysmogram dataset for blood pressure monitoring in China. Scientific Data, 5, 180020, 2018.
  24. Monte-Moreno, E. Non-invasive estimate of blood glucose and blood pressure from a photoplethysmograph by means of machine learning techniques. Artificial Intelligence in Medicine, 53(2), 127–138, 2011.
  25. Loh, H. W., Xu, S., Faust, O., Ooi, C. P., Barua, P. D., Chakraborty, S., Tan, R. S., Molinari, F., and Acharya, U. R. Application of photoplethysmography signals for healthcare systems: An in-depth review. Computer Methods and Programs in Biomedicine, 216, 106677, 2022.
  26. Ortega Caro, J., de Oliveira Fonseca, A. H., Rizvi, S. A., Rosati, M., Averill, C., Cross, J. L., Mittal, P., Zappala, E., Dhodapkar, R. M., Abdallah, C. G., and van Dijk, D. BrainLM: A foundation model for brain activity recordings. International Conference on Learning Representations (ICLR), 2024.
  27. Pillai, A., Spathis, D., Kawsar, F., and Malekzadeh, M. PaPaGei: Open Foundation Models for Optical Physiological Signals. International Conference on Learning Representations (ICLR), 2025.
  28. Goswami, M., Szafer, K., Choudhry, A., Cai, Y., Li, S., and Dubrawski, A. MOMENT: A Family of Open Time-series Foundation Models. Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, 235, 16115–16152, 2024.
  29. Ansari, A. F., Stella, L., Turkmen, A. C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S. S., Pineda Arango, S., Kapoor, S., Zschiegner, J., Maddix, D. C., Wang, H., Mahoney, M. W., Torkkola, K., Wilson, A. G., Bohlke-Schneider, M., and Wang, Y. Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research, 2024.
  30. Chen, Z., Ding, C., Kataria, S., Yan, R., Wang, M., Lee, R., and Hu, X. GPT-PPG: A GPT-based Foundation Model for Photoplethysmography Signals. arXiv preprint, arXiv:2503.08015, 2025.
  31. Nie, G., Fang, X., Tang, G., Xiao, Y., Li, J., Liu, B., Li, H., and Hong, S. AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD '26), 11717–11727, 2026.
  32. Saha, M., Xu, M. A., Mao, W., Neupane, S., Rehg, J. M., and Kumar, S. Pulse-PPG: An Open-Source Field-Trained PPG Foundation Model for Wearable Applications Across Lab and Field Settings. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(3), Article 126, 2025.
  33. Guo, Z., Chen, T., Jiao, Y., Pan, Y., Hu, X., and Ferrario, M. SIGMA-PPG: Statistical-prior Informed Generative Masking Architecture for PPG Foundation Model. arXiv preprint, arXiv:2601.21031 [cs.LG], 2026.
  34. Kulkarni, A., Jain, M., Mishra, A. R., Singh, S., and Panwar, S. PPG-Bench: A Unified Benchmark for PPG Foundation Models. NeurIPS 2026 Brain & Body Foundation Models Workshop, 2026. https://github.com/mannas-ai/ppg-bench
  35. Abbaspourazad, S., Elachqar, O., Miller, A. C., Emrani, S., Nallasamy, U., and Shapiro, I. Large-scale Training of Foundation Models for Wearable Biosignals. International Conference on Learning Representations (ICLR), 2024.
  36. Narayanswamy, G., Liu, X., Ayush, K., Yang, Y., Xu, X., Liao, S., Garrison, J., Tailor, S., Sunshine, J., Liu, Y., Althoff, T., Narayanan, S., Kohli, P., Zhan, J., Malhotra, M., Patel, S., Abdel-Ghaffar, S., and McDuff, D. Scaling Wearable Foundation Models. International Conference on Learning Representations (ICLR), 2025.
PPGs and the FM Gold Rush - MANNAS