PPG Foundation Models: The Complete Guide to Every Pre-Trained Model for Photoplethysmography
PPG foundation models are large pretrained networks that learn reusable pulse-wave representations for heart rate, AF screening, blood pressure proxies, and related tasks. Compare PaPaGei, Pulse-PPG, GPT-PPG, AnyPPG, SiamQuality, and more.
- Blood Pressure
- PPG
Short answer:
- A PPG foundation model is a large network pretrained on unlabeled or weakly labeled photoplethysmography so it learns a reusable pulse-wave encoder, then adapted to tasks such as heart rate, AF screening, blood pressure proxies, stress, and sleep.
- As of late 2026, the open PPG-first pack to know is PaPaGei, Pulse-PPG, GPT-PPG, AnyPPG, and SiamQuality, plus proprietary wearable-scale work from Apple and broader biosignal or time-series models such as BIOT and MOMENT.
- They differ in backbone, pretext task, data source (clinic vs field), openness, and how well they survive motion and device shift. For the full card catalog, see PPG foundation models compared. For a shared public scorecard, see PPG-Bench.
PPG foundation models learn general-purpose representations from large pulse-wave corpora, then transfer those representations with far less labeled data than training a fresh supervised model for every endpoint. That matters because labeled physiological datasets stay scarce and fragmented while PPG now shows up in monitors, rings, watches, patches, and phones.
Background reading: self-supervised learning for PPG, public PPG datasets and benchmarks, the PPG machine learning pipeline, and PPG foundation models compared.
Comparison table: major PPG foundation models
Scroll horizontally if needed
| Model | Backbone | Pretext | Scale (reported) | Open weights |
|---|---|---|---|---|
| PaPaGei | 1D ResNet CNN | Morphology / subject-aware SSL | 57,641 h / 13,517 subjects | Yes |
| Pulse-PPG | 1D ResNet | Motif-aware contrastive SSL on field data | ~55,000 h / 120 participants | Yes |
| GPT-PPG | Causal transformer (GPT) | Next-patch / generative modeling | >200M × 30 s samples; 19M-1B params | Research code |
| AnyPPG | Dual-branch PPG+ECG encoders | InfoNCE alignment to ECG | 109,909 h / 58,796 subjects | Yes |
| SiamQuality | ResNet CNN | Artifact-aware SimSiam | 600,000+ h / 36M paired segments | Partial |
| Apple wearable | EfficientNet-style 1D CNN | Participant-aware contrastive SSL | ~141K participants / ~3 years | No |
| BIOT | Unified tokenizer + transformer | Cross-biosignal SSL | Multi-biosignal corpora | Yes |
| MOMENT | Transformer encoder | Masked time-series modeling | Time Series Pile (incl. healthcare) | Yes |
| Samsung MMR | Transformer | Wavelet masked multiscale reconstruction | ~17M segments / ~32K users | Research-stage |
| PPG-Distill | Distilled student | Teacher-student distillation from FMs | e.g. 19M → ~1M class of students | Paper + code |
| Vision4PPG | Vision transformers (DINOv3, SigLIP-2 ViT-B, 86M) | Imagified PPG (STFT, phase, recurrence) + LoRA | Public vision FMs, not PPG pretraining | Paper + code |
| CEReBrO | Alternating-attention EEG encoder | EEG pretraining on TUEG, then ECG+PPG fine-tune | 3.58M / 39.95M / 85.15M | EEG paper · BP transfer |
| FairTune | Full fine-tune of PPG-GPT (19M reported; sizes 19M–1B) | Inverse-frequency weighting, GroupDRO, or adversarial debiasing for heart-rate fairness | UCSF ICU checkpoint, then DaLiA, BUT-PPG, and MIMIC | Paper |
Card inventory, including Vision4PPG, CEReBrO, and FairTune, lives on the compared topic hub. Pretraining corpora are listed under PPG foundation model pretraining datasets. Task numbers appear on research leaderboards when a paper reports them into PPG-Bench or related suites.
What makes a PPG foundation model different from a traditional PPG deep learning model?
A traditional PPG deep learning model is usually trained end-to-end for one narrow task on one labeled dataset. A heart rate model is trained for heart rate. A blood pressure model is trained for blood pressure. Performance often looks strong inside the training distribution, then drops when the device, population, or motion profile changes.
A PPG foundation model aims broader. It is pretrained on large unlabeled or weakly labeled corpora, usually across many subjects and recording contexts, and learns a reusable representation of pulse morphology, rhythm, noise structure, vascular dynamics, and subject variation. Downstream tasks become lightweight heads or fine-tuning jobs on top of that shared encoder.
Most models in this category share four traits:
- Scale: thousands to millions of recording hours.
- Transferability: evaluated across multiple downstream tasks, not just one.
- Pretraining before supervision: self-supervised, contrastive, generative, masked reconstruction, or multimodal alignment.
- Representation reuse: the encoder is the product, not only the final classifier.
Not every large PPG model qualifies. Some are simply large supervised models for a single endpoint. Others are generic time-series models that happen to include PPG. The useful PPG foundation models are the ones designed to learn transferable physiology from the pulse wave itself. For the learning algorithms behind these objectives, see self-supervised and contrastive learning for PPG.
Timeline: how PPG foundation models evolved
Scroll horizontally if needed
| Year | Milestone | Why it mattered |
|---|---|---|
| 2023 | BIOT and Apple's large-scale wearable SSL | Showed biosignal FMs and wearable-scale SSL were viable |
| 2024 | SiamQuality, PaPaGei, MOMENT | PPG-specific foundation modeling became explicit |
| 2025 | Pulse-PPG, GPT-PPG, multimodal PPG supervision | Field robustness, generative pretraining, physiological guidance |
| 2025-2026 | AnyPPG, wavelet MMR, Vision4PPG, CEReBrO transfer, PPG-Distill, PPG-Bench, fairness work (FairTune) | Phenome-scale profiling, vision backbones, EEG-to-PPG transfer, edge distillation, shared benchmarks |
Early papers asked whether self-supervised pretraining on physiological time series worked at all. By 2025 the question shifted to which pretext is best for PPG (contrastive morphology, masked reconstruction, autoregressive generation, ECG/respiration alignment) and how to evaluate models under shared protocols.
PaPaGei, the first open PPG foundation model
Paper: Pillai et al., "PaPaGei: Open Foundation Models for Optical Physiological Signals." ICLR 2025. arXiv: 2410.20542, DOI: 10.48550/arXiv.2410.20542, OpenReview: https://openreview.net/forum?id=kYwTmlq6Vn. Model card: /research/models/papagei-s.
PaPaGei made "PPG foundation models" a concrete category. The authors include Guanghan Ding, Shivam Pillai, Dimitris Spathis, Fahim Kawsar, Mohammad Malekzadeh, and collaborators, with affiliations including Nokia Bell Labs, Dartmouth College, the University of Cambridge, and the University of Glasgow [1].
What mattered was openness plus design intent. The model was pretrained on 57,641 hours of single-channel PPG from 13,517 participants, pooled from VitalDB (17,355 h), MIMIC-III (19,990 h), and MESA (20,296 h), for 20,751,206 10-second segments [1].
Architecturally, PaPaGei uses a ResNet-style CNN encoder with 18 convolutional blocks and a 512-dimensional projection layer. The authors favored strong local morphology extraction, then built foundation-model behavior into the pretraining objective. Instead of generic augment-and-contrast alone, the paper encodes PPG morphology across individuals, treating waveform shape as signal rather than nuisance.
PaPaGei was evaluated on 20 downstream tasks across 10 datasets, including heart rate, activity, blood pressure, stress-related outcomes, and sleep-related tasks [1]. The pretrained encoder usually beats task-specific training from scratch and often beats more generic time-series baselines. The main limitation is also clear: most of the pretraining data is clinical-grade, so everyday wearable motion and contact noise are under-represented. That gap is where Pulse-PPG enters.
Pulse-PPG, the first open field-trained wearable PPG foundation model
Paper: Saha et al., "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, 2025. arXiv: 2502.01108, DOI: 10.1145/3749494. Model card: /research/models/pulse-ppg.
Pulse-PPG attacks the obvious weakness in early PPG foundation models: overreliance on clean clinical data. The authors, including Mithun Saha, Maxwell A. Xu, Wanting Mao, Sameer Neupane, James M. Rehg, and Santosh Kumar, present it as the first open-source field-trained PPG foundation model [2]. The core dataset comes from a 100-day field study with 120 participants wearing Fossil Sport smartwatches, yielding roughly 55,000 hours of raw 100 Hz PPG in naturalistic conditions [2].
The model uses a ResNet encoder and a contrastive framework with a motif-based distance function for positives and negatives [2]. Evaluations cover stress, activity, heart rate, sleep disturbance, blood pressure, and hypertension-related tasks across lab and field settings [2]. Field pretraining often helped clinical tasks as well as field ones, which suggests robustness training can teach durable physiology rather than only artifact tolerance.
GPT-PPG, the largest published PPG pretraining corpus
Paper: Chen et al., "GPT-PPG: A GPT-based Foundation Model for Photoplethysmography Signals." arXiv: 2503.08015, DOI: 10.48550/arXiv.2503.08015. Model card: /research/models/gpt-ppg.
GPT-PPG is the generative extreme of the field. Rather than contrastive pretraining alone, it asks whether a decoder-only GPT architecture can model PPG the way language models model token sequences. The authors include Zhaoliang Chen, Cheng Ding, Saurabh Kataria and colleagues [3]. Their pretraining dataset contains more than 200 million 30-second PPG samples [3].
Contrastive models ask which windows belong together. GPT-PPG asks which waveform fragment should come next, which suits sequence completion, denoising, and generative downstream use. The paper evaluates atrial fibrillation detection, heart rate, respiration rate, blood pressure estimation, and qualitative denoising, and explores parameter-efficient fine-tuning and personalization [3]. The limitation is deployment cost: even smaller variants are heavier than most CNN encoders. Emory follow-on work uses GPT-PPG-scale teachers inside PPG-Distill for edge students. A related Emory paper, Vision4PPG, instead renders PPG as images and fine-tunes DINOv3 and SigLIP-2. The same lab line also publishes PPG→ECG generation (PG-LRF) plus a verifiable-reasoning position paper (PhysioReason).
AnyPPG, ECG-guided pretraining for physiologically grounded representations
Paper: Nie et al., "AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling." arXiv: 2511.01747, DOI: 10.48550/arXiv.2511.01747, code: https://github.com/PKUDigitalHealth/AnyPPG. Model card: /research/models/anyppg.
AnyPPG uses synchronized ECG as physiological supervision during pretraining. Guangkun Nie, Xiaocheng Fang, Gongzheng Tang, Yujie Xiao, Jun Li, Bo Liu, Hongyan Li, Shenda Hong and colleagues pretrain on 109,909 hours of synchronized PPG-ECG recordings from 58,796 subjects, about 40 million paired 10-second segments, aggregated from resources including MC-MED, PulseDB, MESA, HSP, CFS, MIMIC, and VitalDB-derived data [4].
A PPG encoder and an ECG encoder produce embeddings in a shared latent space. ECG acts as a teacher during pretraining; inference stays PPG-only [4]. Evaluation covers eight downstream datasets spanning wearable and clinical settings, including PPG-DaLiA, UCI-BP, BUT PPG, Gyro-Acc-PPG, WESAD, DeepBeat, Real-World PPG, and WenXinWuYang, plus phenome-wide disease detection claims spanning many circulatory and non-circulatory ICD categories [4]. Those broader screening claims need careful external validation; they do not by themselves make PPG a universal diagnostic modality.
SiamQuality, quality-aware self-supervision before the term became mainstream
Paper: Ding et al., "SiamQuality: a ConvNet-based foundation model for photoplethysmography signals." Physiological Measurement, 2024. DOI: 10.1088/1361-6579/ad6747. PMC: https://pmc.ncbi.nlm.nih.gov/articles/PMC11334241/. Model card: /research/models/siamquality.
SiamQuality addressed a practical problem early: foundation models are only as good as how they handle low-quality signal. The paper trains on over 36 million 30-second PPG pairs from hospitalized ICU patients [5]. The backbone is a ResNet CNN with a SimSiam-style strategy adapted so signal quality is not scrubbed away before learning [5].
The authors report fine-tuning on six downstream tasks, with particularly strong published numbers on respiration rate estimation and atrial fibrillation detection (improvements of 75% and 5% over prior state of the art, as highlighted in the paper) [5]. In hindsight it is an early argument for what Pulse-PPG later made explicit: robustness belongs in pretraining, not only in post-hoc filters.
BIOT, a biosignal foundation model that includes PPG
Paper: Yang et al., "BIOT: Biosignal Transformer for Cross-Data Learning in the Wild." NeurIPS 2023. arXiv: 2305.10351, DOI: 10.48550/arXiv.2305.10351.
BIOT is not PPG-first, but it belongs here because it established cross-dataset biosignal pretraining with a unified biosignal tokenization strategy across modalities such as ECG, EEG, EMG, and PPG [6]. For PPG-only transfer, specialized encoders usually win. For multimodal wearable stacks, BIOT remains a reference point.
Apple's wearable model, the strongest proprietary benchmark
Paper: Abbaspourazad et al., "Large-scale Training of Foundation Models for Wearable Biosignals." arXiv: 2312.05409, DOI: 10.48550/arXiv.2312.05409, Apple ML Research: https://machinelearning.apple.com/research/large-scale-training. Hub card: /research/models/apple-wearable-biosignals.
Abbaspourazad and colleagues describe self-supervised pretraining on Apple Watch data curated from the Apple Heart and Movement Study, including about 141,000 participants over roughly 3 years for the PPG corpus [7]. The model uses an EfficientNet-style 1D CNN, a projection head, and a momentum-optimized contrastive setup with participant-level positive pair selection [7]. Apple reports that the embeddings encode demographics and health conditions and generalize across PPG and ECG modalities [7]. Weights and data are not public, so the paper functions as a strategic benchmark rather than a runnable baseline.
MOMENT, a general time-series foundation model that includes physiological data
Paper: Goswami et al., "MOMENT: A Family of Open Time-series Foundation Models." ICML 2024. arXiv: 2402.03885, DOI: 10.48550/arXiv.2402.03885. Related ChatPPG paper card on specialist vs generalist transfer: Generalist vs specialist time-series FMs on PPG.
MOMENT is often used as a baseline because it is a serious open time-series foundation model trained on the Time Series Pile [8]. Masked time-series modeling transfers across forecasting, classification, anomaly detection, and imputation. Dedicated PPG models usually beat it on morphology-sensitive health tasks, which is useful evidence that PPG rewards bespoke pretraining. Kataria et al. (2025) examine generalist versus specialist behavior on PPG health assessment in more detail [15].
Multimodal physiological supervision (2025 OpenReview)
Paper: "A robust PPG foundation model using multimodal physiological supervision." OpenReview 2025. https://openreview.net/forum?id=ZmGfCj1n2P.
This entry uses accompanying ECG and respiratory signals in ICU data to guide PPG pretraining, while keeping the deployed model unimodal on PPG [9]. The paper reports that with 3× fewer subjects than some state-of-the-art approaches, the model improves up to 36% in classification and 42% in regression on 14 of 15 downstream tasks [9]. Together with AnyPPG, it sharpens a recurring lesson: physiological supervision during pretraining often beats scale alone.
Emerging directions: vision backbones, EEG transfer, distillation, fairness, and benchmarks
Vision4PPG (imagified PPG + vision foundation models)
Paper: Kataria et al., "Vision4PPG: Emergent PPG Analysis Capability of Vision Foundation Models for Vital Signs like Blood Pressure." arXiv: 2510.10366, DOI: 10.48550/arXiv.2510.10366 [13]. Model card: /research/models/vision4ppg. Paper card: /research/papers/vision4ppg-emergent-ppg-vision-foundation-models-2025. Code: https://github.com/saurabh-kataria/Vision4PPG.
Vision4PPG converts 1D PPG into 2D representations (STFT magnitude, STFT phase, and recurrence plots) and fine-tunes DINOv3 and SigLIP-2 ViT-B backbones (86M parameters) with LoRA on the query, key, and value projections (r = 8, α = 16, dropout 0.05) [13]. The vision weights come from public image corpora, not from PPG pretraining. In the authors' Table 1, vision models win 9 of 14 diastolic/systolic MAE comparisons against fully tuned MOMENT-385M and PPG-GPT-345M. SigLIP-2's Aurora-Oscillometric DBP/SBP MAE is 5.32/18.57 mmHg. On the bed BCG set in that same table, MOMENT and PPG-GPT report lower error than either vision model, so the gain is not uniform [13]. Table 2 reports PPG-GPT heart-rate MAE 7.83 bpm on DaLiA, ahead of both vision models, while DINOv3 reports respiration-rate MAE 4.41 breaths/min on BIDMC and SpO2 MAE 2.19% on MIMIC-III [13].
CEReBrO transferred to ECG/PPG (edge quantization story)
Papers: Dimofte et al., "CEReBrO: Compact Encoder for Representations of Brain Oscillations Using Efficient Alternating Attention." arXiv: 2501.10885 [14]; Tóth et al., "Finetuning and Quantization of EEG-Based Foundational BioSignal Models on ECG and PPG Data for Blood Pressure Estimation." arXiv: 2502.17460 [14b]. Model card: /research/models/cerebro. Paper cards: /research/papers/cerebro-compact-encoder-brain-oscillations-2025 and /research/papers/cerebro-eeg-to-ecg-ppg-bp-finetune-quantization-2025.
CEReBrO is an EEG foundation model. The size tables report 3.58M, 39.95M, and 85.15M parameters (the abstract rounds the range to about 3.6M–85M), pretrained on more than 20,000 hours of public TUEG scalp EEG [14]. Tóth et al. fine-tune those weights on 10-second ECG and PPG windows from PulseDB, which is built from MIMIC-III and VitalDB, without a new large biosignal pretraining run [14b]. On the PulseDB MIMIC-III split, the large unfrozen model at 200 epochs reports DBP MAE 1.57 mmHg and SBP MAE 2.72 mmHg. The same setting on VitalDB reports 1.92 and 3.14 mmHg (Table II) [14b]. The abstract says dynamic INT8 quantization shrinks the smallest model from 13.73 MB to 3.83 MB. Table IV lists that small model at 13.73 MB original, 3.85 MB dynamic INT8, and 3.83 MB static INT8 [14b]. This is cross-modal transfer, not a PPG-native foundation model. The transfer PDF carries a 2025 IEEE copyright notice and does not name a conference in the text.
Wavelet MMR and ProtoMM
OpenReview's Wavelet-Driven Masked Multiscale Reconstruction pretrains on about 17 million unlabeled 10-second segments from roughly 32,000 smartwatch users, reconstructing wavelet multiresolution coefficients rather than raw samples [10]. ChatPPG's related card is /research/models/samsung-mmr. ProtoMM aligns PPG and accelerometry through shared prototypes [11]; it is multimodal by design rather than a pure PPG FM.
PPG-Distill (edge deployment)
Paper: Ni et al., "PPG-Distill: Efficient Photoplethysmography Signals Analysis via Foundation Model Distillation." arXiv: 2509.19215, DOI: 10.48550/arXiv.2509.19215 [12]. Model card: /research/models/ppg-distill. Paper card: /research/papers/ppg-distill-foundation-model-distillation-neurips-ts4h-2025. Code: https://github.com/xiaoyuli0575/PPG-Distill.
PPG-Distill is operational rather than a new pretraining paradigm. Teacher-student distillation tailored to PPG (prediction-, feature-, patch-, rhythm-, and morphology-level losses) reports up to 21.8% student performance improvement, 7× faster inference, and 19× lower memory use on heart rate estimation and AF detection in the paper [12]. On DaLiA the distilled GPT-PPG-1m student reaches HR MAE 8.34 bpm; on StanfordAF it reaches AF F1 0.77 when taught by GPT-PPG-19m [12]. Wearable deployment of foundation models depends on this class of compression as much as on representation quality.
PG-LRF (physiology-guided PPG→ECG generation)
Paper: Wang et al., "PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation." arXiv: 2605.12541 [18]. Model card: /research/models/pg-lrf. Paper card: /research/papers/pg-lrf-ppg-to-ecg-neurips-2026.
PG-LRF is a generative method card, not a released PPG encoder. It couples an electro-hemodynamic simulator with latent rectified flow so ECG can be synthesized from PPG alone. On MC-MED the paper reports waveform MAE 0.71, RMSE 1.07, FID 37.79, and HR MAE 3.94 bpm, plus a Macro-AUROC of 0.633 for multi-label CVD labels on generated ECG versus 0.545 for PaPaGei-on-PPG [18]. Useful when a pipeline needs ECG-shaped features from optical hardware; no public weights were listed with the preprint.
PhysioReason (verifiable waveform reasoning)
Paper: Wang et al., "Position: Beyond Prediction: Toward Verifiable Physiological Waveform Reasoning with Foundation Models and Agentic LLMs." Emory PDF: https://www.cs.emory.edu/~jyang71/files/physioreason.pdf [19]. Hub card: /research/papers/physioreason-verifiable-waveform-reasoning-icml-2026.
PhysioReason is an ICML 2026 position paper (with a listed KDD 2026 tutorial) that argues physiological FMs should expose localized evidence, map it to physiological semantics, and support guideline-constrained decisions that can abstain. It sketches a System-1 / System-2 agent stack and a Verification Ledger; it does not publish a new encoder scoreboard [19].
FairTune (fairness under fine-tuning)
Paper: Panchumarthi et al., "FairTune: A Bias-Aware Fine-Tuning Framework Towards Fair Heart Rate Prediction from PPG." arXiv: 2509.16491, DOI: 10.48550/arXiv.2509.16491 [16]. Model card: /research/models/fairtune. Paper card: /research/papers/fairtune-bias-aware-fine-tuning-ppg-heart-rate-2025.
FairTune is a fine-tuning study of GPT-PPG, not a new encoder. The authors start from a PPG-GPT checkpoint pretrained on UCSF ICU data and adapt the 19M model, with scaling runs from 19M to 1B, on PPG-DaLiA, BUT PPG, and a MIMIC ICU PPG set cited through Charlton et al. 2022 [16]. Table 2's unbalanced fine-tune cuts heart-rate MAE from 89.54 to 17.62 bpm on DaLiA (80%), from 78.22 to 19.23 bpm on BUT-PPG (75%), and from 93.81 to 24.73 bpm on MIMIC (74%). Those cross-dataset errors sit above the same-dataset Table A1 baselines of 4.28, 9.94, and 9.76 bpm [16].
The abstract says inverse-frequency weighting and GroupDRO reduce fairness gaps without compromising accuracy. Table 3, which reports medians, is more mixed. On DaLiA to BUT-PPG, GroupDRO lowers the absolute male–female MAE gap from 2.36 to 0.68 bpm, and inverse-frequency weighting reaches MAE 13.44 bpm. On DaLiA to MIMIC the unbalanced gap is 0.80 bpm, and inverse-frequency weighting, GroupDRO, and adversarial debiasing all report higher gaps (1.06, 1.20, and 1.16 bpm) [16]. The experiments use binary gender only. Related reading: PPG skin tone bias and accuracy.
Paper: "PPG-Bench: A Unified Benchmark for PPG Foundation Models." arXiv: 2610.13435. Hub card: /research/papers/ppg-bench-unified-benchmark-ppg-foundation-models-2026.
PPG-Bench compares several public PPG foundation models under a shared protocol. When authors disagree about who "wins," start here and on the leaderboards rather than cherry-picking single-task tables from individual papers.
Pretraining methods: the main schools of thought
1. Contrastive learning
Common in PaPaGei, Pulse-PPG, Apple's wearable model, and several multimodal variants. Related segments are pulled together; unrelated segments are pushed apart. What counts as related differs by paper: augmentations of the same window, same-person segments, morphologically similar pulses, or synchronized ECG/respiration context. Pair selection quality decides whether the method learns physiology or artifacts.
2. Masked reconstruction
MOMENT and wavelet-driven PPG models hide part of the input and reconstruct it. For PPG, reconstructing in a transformed domain (for example wavelet coefficients) often fits multiscale physiology better than reconstructing raw samples alone.
3. Autoregressive generation
GPT-PPG predicts the next patch from prior context. That learns sequence continuity and generative structure, at higher inference cost.
4. Physiological guidance and multimodal alignment
AnyPPG and the multimodal-supervision OpenReview paper use ECG, respiration, or other synchronized signals during pretraining while keeping inference on PPG. That injects mechanistic structure into the encoder without requiring a second sensor at deployment.
What downstream tasks are these models actually good at?
Recurring evaluation endpoints across papers:
- Heart rate estimation
- Respiration rate estimation
- Atrial fibrillation and arrhythmia detection
- Blood pressure estimation and hypertension classification
- Stress classification
- Activity classification
- Sleep quality, disturbance, or stage-related tasks
- Signal denoising or waveform reconstruction
- Phenotype and disease screening (read these claims carefully)
The strongest and most reproducible wins so far are on tasks where PPG is already known to be informative: heart rate, pulse-interval variability, blood pressure proxies, AF screening, and stress-related autonomic inference. Broader disease-screening claims, especially outside cardiovascular domains, need external validation. For clinical framing of AF screening, see PPG-based atrial fibrillation screening. This guide is educational research coverage, not medical advice and not a device recommendation.
What this means for wearable product teams
Historically, most wearable vendors shipped a handful of hand-tuned pipelines: one for heart rate, one for HRV, maybe SpO2 or stress. Foundation models make it realistic to treat the PPG encoder as a platform asset.
Practical implications:
- New downstream features can launch with less labeled data.
- Personalization becomes fine-tuning a strong prior rather than training from scratch.
- Cross-device transfer improves only if the pretraining set is large and diverse enough.
- Robustness to field noise can be learned upstream instead of patched later.
- Edge budgets still matter: distillation (PPG-Distill) and INT8 quantization of the CEReBrO ECG/PPG transfer are part of making these models smaller.
Better sensors still matter. Clinical-grade and carefully engineered wearable hardware create richer pretraining corpora; richer corpora produce better models. Teams that can collect synchronized, longitudinal, multi-context PPG, ideally with paired ECG, blood pressure, respiration, and outcome labels, will shape the next generation of wearable health models.
Where the field is headed
Three trends look especially important.
First, wearable-native pretraining will keep growing. Pulse-PPG and smartwatch-scale masked models show that field data is a distinct domain, not only a noisier version of clinical data.
Second, multimodal physiological supervision will likely become standard. ECG-guided and respiration-guided objectives ground PPG encoders without requiring multimodal sensors at deployment. Vision backbones on imagified PPG (Vision4PPG) and EEG encoders fine-tuned on ECG and PPG (CEReBrO) add other transfer paths.
Third, the bottleneck is shifting from inventing yet another encoder to evaluation, fairness, and clinical validity. Shared benchmarks such as PPG-Bench, fairness studies such as FairTune, and prospective, device-diverse replication matter more than another leaderboard cherry-pick.
References
- Pillai S, Ding G, Spathis D, Kawsar F, Malekzadeh M, et al. PaPaGei: Open Foundation Models for Optical Physiological Signals. ICLR 2025. arXiv:2410.20542. DOI: https://doi.org/10.48550/arXiv.2410.20542. OpenReview: https://openreview.net/forum?id=kYwTmlq6Vn
- Saha M, Xu MA, Mao W, Neupane S, Rehg JM, Kumar S. Pulse-PPG: An Open-Source Field-Trained PPG Foundation Model for Wearable Applications Across Lab and Field Settings. Proc ACM Interact Mob Wearable Ubiquitous Technol. 2025;9(3). arXiv:2502.01108. DOI: https://doi.org/10.1145/3749494. arXiv DOI: https://doi.org/10.48550/arXiv.2502.01108
- Chen Z, Ding C, Kataria S, et al. GPT-PPG: A GPT-based Foundation Model for Photoplethysmography Signals. arXiv:2503.08015. DOI: https://doi.org/10.48550/arXiv.2503.08015
- Nie G, Fang X, Tang G, Xiao Y, Li J, Liu B, Li H, Hong S. AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling. arXiv:2511.01747. DOI: https://doi.org/10.48550/arXiv.2511.01747. Code: https://github.com/PKUDigitalHealth/AnyPPG
- Ding C, Guo Z, Chen Z, Lee RJ, Rudin C, Hu X. SiamQuality: a ConvNet-based foundation model for photoplethysmography signals. Physiological Measurement. 2024. DOI: https://doi.org/10.1088/1361-6579/ad6747. Full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC11334241/
- Yang C, Westover MB, Sun J. BIOT: Biosignal Transformer for Cross-Data Learning in the Wild. NeurIPS 2023. arXiv:2305.10351. DOI: https://doi.org/10.48550/arXiv.2305.10351. OpenReview: https://openreview.net/forum?id=c2LZyTyddi
- Abbaspourazad S, Elachqar O, Miller AC, et al. Large-scale Training of Foundation Models for Wearable Biosignals. arXiv:2312.05409. DOI: https://doi.org/10.48550/arXiv.2312.05409. Apple ML Research: https://machinelearning.apple.com/research/large-scale-training
- Goswami M, Szafer K, Choudhry A, Cai Y, Li S, Dubrawski A. MOMENT: A Family of Open Time-series Foundation Models. ICML 2024. arXiv:2402.03885. DOI: https://doi.org/10.48550/arXiv.2402.03885
- A robust PPG foundation model using multimodal physiological supervision. OpenReview, 2025. https://openreview.net/forum?id=ZmGfCj1n2P
- Wavelet-Driven Masked Multiscale Reconstruction for PPG Foundation Models. OpenReview, 2025. https://openreview.net/forum?id=tVu1zfdbhu
- Leveraging Shared Prototypes for a Multimodal Pulse Motion Foundation Model. OpenReview, 2025. https://openreview.net/forum?id=CtcKEAojoE
- Ni J, Kataria S, Tang S, Yang C, Hu X, Jin W. PPG-Distill: Efficient Photoplethysmography Signals Analysis via Foundation Model Distillation. NeurIPS 2025 TS4H Spotlight. arXiv:2509.19215. DOI: https://doi.org/10.48550/arXiv.2509.19215. ChatPPG: https://chatppg.com/research/models/ppg-distill
- Kataria S, Ermis A, Panchumarthi LY, Wang M, Hu X. Vision4PPG: Emergent PPG Analysis Capability of Vision Foundation Models for Vital Signs like Blood Pressure. arXiv:2510.10366. DOI: https://doi.org/10.48550/arXiv.2510.10366. ChatPPG: https://chatppg.com/research/models/vision4ppg
- Dimofte A, Bucagu GA, Ingolfsson TM, et al. CEReBrO: Compact Encoder for Representations of Brain Oscillations Using Efficient Alternating Attention. arXiv:2501.10885. DOI: https://doi.org/10.48550/arXiv.2501.10885. ChatPPG: https://chatppg.com/research/models/cerebro 14b. Tóth B, Senti D, Ingolfsson TM, et al. Finetuning and Quantization of EEG-Based Foundational BioSignal Models on ECG and PPG Data for Blood Pressure Estimation. arXiv:2502.17460. DOI: https://doi.org/10.48550/arXiv.2502.17460. ChatPPG: https://chatppg.com/research/papers/cerebro-eeg-to-ecg-ppg-bp-finetune-quantization-2025
- Kataria S, et al. Generalist vs Specialist Time Series Foundation Models: Investigating Potential Emergent Behaviors in Assessing Human Health Using PPG Signals. 2025. ChatPPG: https://chatppg.com/research/papers/generalist-vs-specialist-tsfm-ppg-kataria-2025
- Panchumarthi LY, Kataria S, Wu Y, Hu X, Fedorov A, Kwak HG. FairTune: A Bias-Aware Fine-Tuning Framework Towards Fair Heart Rate Prediction from PPG. arXiv:2509.16491. DOI: https://doi.org/10.48550/arXiv.2509.16491. ChatPPG: https://chatppg.com/research/models/fairtune
- PPG-Bench: A Unified Benchmark for PPG Foundation Models. arXiv:2610.13435. ChatPPG: https://chatppg.com/research/papers/ppg-bench-unified-benchmark-ppg-foundation-models-2026
- Wang X, et al. PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation. arXiv:2605.12541. DOI: https://doi.org/10.48550/arXiv.2605.12541. ChatPPG: https://chatppg.com/research/models/pg-lrf
- Wang X, et al. Position: Beyond Prediction: Toward Verifiable Physiological Waveform Reasoning with Foundation Models and Agentic LLMs. ICML 2026 position (Emory PDF). https://www.cs.emory.edu/~jyang71/files/physioreason.pdf. ChatPPG: https://chatppg.com/research/papers/physioreason-verifiable-waveform-reasoning-icml-2026
Frequently Asked Questions
- What is a PPG foundation model?
- A PPG foundation model is a neural network pretrained on large unlabeled or weakly labeled photoplethysmography corpora so it learns reusable pulse-wave representations. Those encoders are then fine-tuned or probed for tasks such as heart rate estimation, atrial fibrillation screening, blood pressure proxies, stress classification, and sleep-related endpoints.
- Which is the first open PPG foundation model?
- PaPaGei is generally treated as the first open foundation model built specifically for PPG. It was pretrained on more than 57,000 hours of public PPG from VitalDB, MIMIC-III, and MESA and released in the ICLR 2025 cycle (arXiv:2410.20542).
- Which PPG foundation model has the largest published training corpus?
- Among PPG-specific published models, GPT-PPG reports the largest corpus: more than 200 million 30-second PPG samples (arXiv:2503.08015). Generative transformer variants are heavier to deploy than compact CNN encoders.
- Which open model is best suited to real-world wearable PPG?
- Pulse-PPG is the clearest open model pretrained for field wearables. It was trained on roughly 55,000 hours of smartwatch PPG from a 100-day study and reports stronger transfer when the pretraining distribution includes real-world noise (arXiv:2502.01108).
- Why use ECG-guided pretraining for PPG?
- ECG records electrical timing and rhythm more directly; PPG records peripheral hemodynamics. Aligning PPG to synchronized ECG during pretraining, as AnyPPG does, can pull optical embeddings toward cardiac structure even when only PPG is used at inference (arXiv:2511.01747).
- Are generic time-series foundation models enough for PPG?
- They are useful baselines. Models such as MOMENT transfer across many domains, but PPG-specialized encoders usually win on morphology-sensitive pulse tasks because they see pulse shape, device noise, and physiological artifact structure during pretraining.
- Where can I compare open PPG foundation models side by side?
- Use the ChatPPG research hub catalog at /research/topics/ppg-foundation-models-compared for the full ItemList of model cards, and PPG-Bench (arXiv:2610.13435) for a unified public benchmark of several open PPG foundation models.
Related articles
State of PPG, October 2026: the first 38 SOTA stars
October 2026 in the ChatPPG research hub: 574 papers, 147 leaderboards and the first 38 SOTA stars. AnyPPG holds six; no star changed hands yet.
What Does PPG Stand For? A Complete Guide to Photoplethysmography
PPG stands for photoplethysmography — the optical technique that powers heart rate sensors in smartwatches and clinical pulse oximeters. Learn exactly how it works.
What Is Photoplethysmography (PPG)? The Complete Technical Guide
A comprehensive technical guide to photoplethysmography (PPG) explaining how PPG sensors work, the physics of light-tissue interaction, transmission vs reflectance modes, signal components, and modern applications in wearables and clinical devices.
Photoplethysmography Sensor: Types, Design, and Selection Guide
Everything you need to know about photoplethysmography sensors: LED wavelengths, photodetector types, analog front-ends, form factors, and how to choose the right sensor for your application.
Monte Carlo and Photon Diffusion Models for PPG Simulation: A Technical Guide
**Understanding how photons travel through biological tissue is fundamental to designing, optimizing, and interpreting photoplethysmographic (PPG) sen...
Devices in this article
Research records with the datasets, papers and published validation results for each device.