PaperlayerView exact revision on arXiv

Source-linked AI summary

PaPaGei: Open Foundation Models for Optical Physiological Signals

Arvind Pillai, Dimitris Spathis, Fahim Kawsar, Mohammad Malekzadeh

arXiv:2410.20542v2cs.LGeess.SP

TL;DR

PPG machine-learning work is constrained by task-specific models and limited generalization, while prior studies often rely on restricted datasets or lack broad out-of-domain evaluation. PaPaGei pre-trains an open PPG foundation model on public data using morphology-aware learning; it outperforms baselines in at least 14 tasks, though robustness across darker skin tones remains a challenge.

  • Problem

    PPG research has lacked widely used general-purpose models, with prior work often limited by proprietary or inaccessible data, limited out-of-domain evaluation, or unavailable models.

    Source: §1 · p. 2, H EXTENDED RELATED WORK · p. 29

  • Method

    PaPaGei pre-trains PPG foundation models on public datasets and introduces self-supervised learning with morphology augmentation to capture PPG signal features.

    Source: §1 · p. 2

  • Results

    PaPaGei outperforms baselines in at least 14 tasks, with classification improvements of 4.7%-6.3% and regression improvements of 2.9%-4.9%.

    Source: §6 · p. 10

  • Takeaways & Limitations

    PaPaGei’s embeddings show predictive information across health tasks, and the models are evaluated across 10 diverse datasets.

    Source: §1 · p. 2

    Across dark skin tones, BYOL and REGLE obtain the lowest MAE for systolic and diastolic BP, and further work is needed to improve robustness.

    Source: §5.3 · p. 10

Abstract

from arXiv · show

Photoplethysmography (PPG) is the leading non-invasive technique for monitoring biosignals and cardiovascular health, with widespread adoption in both clinical settings and consumer wearable devices. While machine learning models trained on PPG signals have shown promise, they tend to be task-specific and struggle with generalization. Current research is limited by the use of single-device datasets, insufficient exploration of out-of-domain generalization, and a lack of publicly available models, which hampers reproducibility. To address these limitations, we present PaPaGei, the first open foundation model for PPG signals. The model is pre-trained on over 57,000 hours of data, comprising 20 million unlabeled PPG segments from publicly available datasets. We introduce a novel representation learning approach that leverages domain knowledge of PPG signal morphology across individuals, enabling the capture of richer representations compared to traditional contrastive learning methods. We evaluate PaPaGei against state-of-the-art time-series foundation models and self-supervised learning benchmarks across 20 tasks from 10 diverse datasets, spanning cardiovascular health, sleep disorders, pregnancy monitoring, and wellbeing assessment. Our model demonstrates superior performance, improving classification and regression metrics by 6.3% and 2.9% respectively in at least 14 tasks. Notably, PaPaGei achieves these results while being more data- and parameter-efficient, outperforming models that are 70x larger. Beyond accuracy, we examine model robustness across different skin tones, establishing a benchmark for bias evaluation in future models. PaPaGei can serve as both a feature extractor and an encoder for multimodal models, opening up new opportunities for multimodal health monitoring.

1 INTRODUCTION

PPG supports health monitoring across clinical and consumer settings, but annotation costs, signal noise, and variability complicate generalizable machine learning. PaPaGei addresses the lack of general-purpose PPG models with public-data pretraining, morphology-aware learning, and evaluation across diverse tasks.

1 INTRODUCTION

  • PPG annotation costs, noise, motion artifacts, and variation across sensing conditions and users complicate generalizable models, while existing datasets are often small and task-specific.The passage also identifies skin tone and body composition as sources of signal variability.Source: §1 · p. 1
  • The PPG domain lacks general-purpose foundation models, and PaPaGei is introduced as a robust pre-trained backbone for varied PPG tasks.The introduction notes that most existing work focuses on single-dataset, task-specific models.Source: §1 · p. 2
  • 57,000 hours from 20 million signals in public datasets support PaPaGei, presented as the first open foundation model pre-trained on PPG.The authors position this as large-scale pretraining for wearable and clinical health monitoring.Source: §1 · p. 2
  • PaPaGei’s self-supervised framework uses morphology augmentation to align signals with similar blood-volume changes and focus on the systolic peak and dicrotic notch.These are identified as key PPG markers.Source: §1 · p. 2
  • PaPaGei is evaluated across 20 tasks spanning cardiovascular health, sleep disorders, pregnancy monitoring, and well-being.The tasks use 10 datasets, and the authors report that embeddings contain predictive information across health conditions.Source: §1 · p. 2
  • The evaluation includes ablations and comparisons addressing morphology augmentation, contrastive learning, model size, data efficiency, and skin tone.These studies examine robustness and the effects of key components.Source: §1 · p. 2

2 RELATED WORK

Prior PPG representation studies show promise, but the literature described here lacks widely used, openly available PPG foundation models and broad out-of-domain validation. The paper situates its PPG-specific approach against generic time-series and other physiological-signal models.

2 RELATED WORK

  • Prior PPG embedding studies used proprietary or restricted datasets, often omitted out-of-domain evaluation or model release, and PaPaGei contrasts this with public training data and validation on 10 datasets.The cited work also reports PPG embeddings for diverse health tasks and genetic discovery or risk prediction.Source: §2 · p. 2
  • The PaPaGei-S overview maps computed IPA, SVRI, and SQI metrics to an encoder and specialized heads that contrast sVRI or predict IPA and SQI.The projection head contrasts based on sVRI values; two mixture-of-expert heads predict IPA and SQI.Source: Figure 2 · p. 3
  • Generic time-series foundation models may lack physiological representations, while adapting ECG or EEG models is challenging because their signal characteristics differ from PPG.The passage says knowledge transfer from time-series models might help PPG tasks, but their performance is limited compared with PPG-specific models.Source: §2 · p. 3

3 METHODS

PaPaGei introduces two self-supervised objectives for PPG: one pairs segments from the same participant, while the other pairs segments with similar morphology. The morphology-aware approach combines contrastive learning with predictions of PPG indices.

3.1 PARTICIPANT-AWARE OBJECTIVE: PAPAGEI-P

  • PAPAGEI-P defines positive pairs as distinct PPG segments from the same participant and trains their embeddings with NT-Xent.Training applies random cropping, Gaussian noise, time flipping, negation, and magnitude scaling; vanilla SimCLR instead pairs augmented versions of randomly sampled segments.Source: §3.1 · pp. 3–4

3.2 MORPHOLOGY-AWARE OBJECTIVE: PAPAGEI-S

  • The morphology module measures sVRI, IPA, and SQI to capture amplitude variation, waveform width, and signal quality.SQI is included for cases where noise or differing morphology makes IPA difficult to compute; the authors report that SQI is significantly larger in signals with a dicrotic notch (p < 0.05).Source: §3.2 · p. 4
  • PAPAGEI-S defines positive pairs by discretized sVRI labels, not participant identity, with sVRI divided into eight bins.The module outputs sVRI, IPA, and SQI labels from an input augmented with Gaussian noise and cropping.Source: §3.2 · p. 4
  • PAPAGEI-S combines contrastive learning on sVRI-based pairs with IPA and SQI prediction heads trained using mean absolute error.The overall objective weights the sVRI contrastive loss against the IPA and SQI losses.Source: §3.2 · p. 4

4 EXPERIMENTS

The experiments pre-train PaPaGei on three PPG datasets and evaluate its representations across diverse tasks using linear probes. They compare it with pretrained models, self-supervised methods, and statistical features.

4.1 PRE-TRAINING

  • Pre-training uses 20 million segments from 13.5K participants across VitalDB, MIMIC-III, and MESA.The sources include surgical, ICU-monitor, and sleep-study PPG recordings.Source: §4.1 · p. 5, Table 1 · p. 5

4.2 DOWNSTREAM TASKS

  • Evaluation spans cardiovascular, sleep, pregnancy, and wellbeing tasks, with Table 2 marking unseen datasets and task types.The figure summarizes classification AUROC and regression MAE across downstream tasks.Source: Table 2 · p. 5, §4.2 · p. 6, Figure 3 · p. 6

4.3 BASELINES

  • Baselines include Chronos, MOMENT, and REGLE; SimCLR, BYOL, and TF-C; and a random forest trained on statistical features.Table 3 compares pretrained models and reports 95% confidence intervals, with the best value bolded.Source: §4.3 · p. 6, Table 3 · p. 7

4.4 LINEAR EVALUATION

  • Linear evaluation extracts representations from data resampled to 125 Hz and trains task-specific probes.Splits are subject-level; binary tasks use logistic regression and AUROC, regression uses ridge regression and MAE, and multiclass tasks use random forests and accuracy.Source: §4.4 · p. 7

5 RESULTS

PaPaGei-S leads the reported overall classification and regression comparisons, while ablations show effects of upstream data, model components, and model size. Case studies examine participant embeddings, regression predictions, and skin-tone differences in blood-pressure estimation.

5.1 OVERALL PERFORMANCE

  • PaPaGei-S leads average AUROC and MAE among SSL methods and achieves 0.67 average AUROC and 10.12 average MAE overall.The overall comparison also reports particular strengths in ICU admission, hypertension, sleep-disordered breathing, AHI, and blood-pressure tasks.Source: §5.1 · pp. 7–8, Table 4 · p. 8

5.2 ABLATION STUDIES

  • Performance improves with more pre-training data and increased label availability, with the best pre-training results using all three datasets.MESA outperforms the other individual datasets despite having the fewest participants and most segments.Source: Figure 4 · p. 7, §5.2 · p. 8, Figure 5 · p. 8
  • The full model (0.67, 10.12) outperforms individual components, while combining SQI and IPA provides greater benefits than either alone.sVRI alone (0.64, 10.35) exceeds sVRI + SQI (0.62, 10.80) and sVRI + IPA (0.64, 10.73) on the reported mean and median metrics.Source: §5.2 · p. 8, Figure 6 · p. 9
  • The 5M-parameter model outperforms larger models on all but one task, revealing a nonmonotonic scaling trend.The 139M model surpasses the 35M model but still trails the 5M model.Source: §5.2 · p. 9, Figure 6 · p. 9
  • Demographic features complement PaPaGei-S, and adding them creates a stronger model; PaPaGei-S alone beats demo + PPG in 14 of 18 tasks.The diverse-device evaluation is offered as an explanation for lower demographic-prediction performance than larger closed studies.Source: §5.2 · p. 9

5.3 CASE STUDIES

  • PaPaGei-S has the widest dispersion of SDB participant embeddings, whereas SimCLR and BYOL cluster participants more tightly.The authors suggest the wider dispersion may help distinguish participants’ medical conditions.Source: §5.3 · pp. 9–10, Figure 7 · p. 9
  • For AHI prediction, PaPaGei-S shows steeper slopes and higher R2 values than pre-trained and SSL baselines, with a wider prediction distribution that captures the left tail better.SimCLR and Chronos predictions regress more toward the mean.Source: §5.3 · p. 10, Figure 8 · p. 10
  • PaPaGei-S performs best for blood-pressure estimation across light skin tones, but BYOL and REGLE have the lowest dark-tone SBP and DBP MAE.The authors report no single best model across all skin tones and call for further work on darker skin-tone robustness.Source: §5.3 · p. 10

6 DISCUSSION & CONCLUSION

PaPaGei outperforms baselines in at least 14 tasks, with reported classification and regression gains. The discussion highlights its efficiency and identifies training-data diversity, sampling rates, and PPG morphology as areas for further work.

6 DISCUSSION & CONCLUSION

  • PaPaGei outperforms baselines in at least 14 tasks, with classification improvements of 4.7%-6.3% and regression improvements of 2.9%-4.9%.PaPaGei-S achieves the highest R2 in seven regression tasks and the highest F1 in six of nine classification tasks.Source: §6 · p. 10
  • The 5M-parameter PaPaGei is described as data- and size-efficient for medical settings where 200M+ models are impractical due to on-device limits or cloud-inference privacy concerns.The authors also report that combining the PaPaGei-P and PaPaGei-S objectives in one model is impractical because of constraints on positive-pair labels and limited samples per label.Source: §6 · p. 10
  • Future work should diversify training data, study sampling-rate effects, explore multimodal approaches or alternative architectures, and systematically evaluate PPG features and modeling.Figure 9 presents the paper’s skin-tone analysis for blood-pressure estimation.Source: §6 · p. 10, Figure 9 · p. 10

REPRODUCIBILITY STATEMENT

The paper provides public data access, code, and methodological details to support reproducibility and future research.

REPRODUCIBILITY STATEMENT

  • Public datasets, code with user-friendly examples, and model and preprocessing details are provided to support reproducibility.Dataset descriptions and ground-truth annotations are provided in §4 and Appendix §B; architecture and hyperparameters are described in §3, §4.1, and Appendix §A.Source: REPRODUCIBILITY STATEMENT · p. 11

ETHICS STATEMENT

The authors describe privacy and misuse risks alongside their commitment to transparent, responsible development and deployment of PaPaGei.

ETHICS STATEMENT

  • The authors identify potential misuse including unauthorized monitoring, discriminatory insurance or employment practices, unfair credit scoring, and targeted marketing using health data.They advocate beneficial healthcare use and continued interdisciplinary dialogue and governance to protect privacy and prevent misuse.Source: ETHICS STATEMENT · p. 11
  • The study reports following established research ethics guidelines and declares no conflicts of interest.Source: ETHICS STATEMENT · p. 11

A TRAINING AND INFERENCE DETAILS

PaPaGei uses a shared ResNet-style encoder and aligned training settings for fair comparisons, with different self-supervised methods varying in model size and feature-extraction setup. The section also specifies pre-training, model selection, and linear-evaluation procedures.

A TRAINING AND INFERENCE DETAILS

  • The ResNet-style encoder uses 18 blocks, 1D convolutions with kernel size 3 and stride 2, and filters that start at 32 and double every four blocks.Dropout is 0.5; the 512-dimensional representation passes through two MoE blocks, each with three experts.Source: A TRAINING AND INFERENCE DETAILS · p. 17, Table 5 · p. 17, Table 7 · p. 17
  • PaPaGei and SSL baselines share batch size, learning rate, input sampling frequency, and training steps; training uses Adam for 15,000 steps at 10^-4 with batch size 128.Five pre-training iterations are run, with the best-performing model or models selected for each downstream task; the authors note broader k-fold tuning would be more robust but computationally costly.Source: A TRAINING AND INFERENCE DETAILS · p. 17
  • PAPAGEI-P and SimCLR use 5M parameters for inference, while BYOL uses its 5M-parameter online encoder and TF-C requires 10M parameters.PAPAGEI-S also uses 5M inference parameters because its expert heads are not used for feature extraction.Source: A TRAINING AND INFERENCE DETAILS · p. 18
  • Feature extraction uses projected embeddings for linear evaluation, with cross-validated grid search selecting probe hyperparameters for logistic regression, linear regression, and random forests.Moment and Chronos use their default 1024- and 768-dimensional embeddings, respectively.Source: A TRAINING AND INFERENCE DETAILS · p. 18

B DATASETS AND TASKS

The benchmark spans datasets and tasks in cardiovascular health, sleep, pregnancy, mood, activity, and stress, with some tasks designated out-of-domain. Dataset descriptions specify population, signal, and label characteristics that shape these evaluations.

B DATASETS AND TASKS

  • Table 8 distinguishes out-of-domain tasks on datasets unseen during training from tasks whose test sets and labels were held out despite pre-training use.Its task-type codes identify binary classification, multiclass classification, and regression.Source: Table 8 · p. 18
  • VitalDB includes 5,866 participants with usable information from surgical cases, and its ICU label indicates whether a person was admitted to intensive care.The described cohort is mostly aged 50–70, with nearly balanced male and female representation.Source: B DATASETS AND TASKS · pp. 18–19
  • The nuMoM2B tasks classify early versus late pregnancy visits and predict gestational age, while Vital Videos contributes skin-tone data from 231 participants in Europe and Sub-Saharan Africa.Vital Videos participants are mostly Fitzpatrick skin tones 5 or 6; the pregnancy recordings are collected at 6–15 and 22–31 weeks.Source: B DATASETS AND TASKS · p. 20
  • PPG-BP provides three 2.1-second, 1000-Hz recordings per participant, which are zero-padded to 10 seconds; the SDB task groups AHI levels 1–3 as disorder and level 0 as no disorder.PPG-BP includes 219 participants, while the SDB cohort contains 146 children with polysomnography recordings longer than three hours.Source: B DATASETS AND TASKS · pp. 20–21
  • ECSMP classifies high versus low mood disturbance using the median Total Mood Disturbance cutoff, while WESAD uses questionnaire-derived valence and arousal labels and performs segment-level regression.PPG-DaLiA estimates heart rate across daily activities in 15 participants using 8-second windows; its heart rates vary from 60 to 150 beats per minute.Source: B DATASETS AND TASKS · p. 22

C REPRESENTATIVE SIGNALS FROM PRE-TRAINING DATASETS

The pre-training examples show that raw PPG amplitude and sampling rate vary across datasets, while pre-processing makes the displayed signal characteristics more consistent. The supplied classification tables identify comparison scope but do not provide cell-level results.

C REPRESENTATIVE SIGNALS FROM PRE-TRAINING DATASETS

  • Raw 10-second PPG segments from VitalDB, MESA, and MIMIC-III differ in amplitude and sampling rate.Figure 20 presents representative signals from these three datasets.Source: Figure 20 · p. 23
  • After normalization and resampling, the displayed 10-second signals from VitalDB, MESA, and MIMIC-III have more consistent characteristics.Figure 21 shows the pre-processed signals in their final form before model input.Source: Figure 21 · p. 23

D.1 MULTI-CLASS CLASSIFICATION

  • The multi-class classification comparison against pre-trained models reports feature-extraction parameters and 95% confidence intervals, with the best value bolded.The supplied caption does not state the cell values or identify a winner.Source: Table 9 · p. 24
  • The multi-class classification comparison against contrastive-learning methods reports feature-extraction parameters and 95% confidence intervals, with the best value bolded.The supplied caption does not state the cell values or identify a winner.Source: Table 10 · p. 24

D.2 F1-SCORE AND R2 EVALUATION METRICS.

The supplementary material adds metric comparisons, ablation analyses, and statistical tests. Its reported comparisons favor PaPaGei across classification and regression, while a separate permutation test distinguishes SQI values by IPA availability.

D.2 F1-SCORE AND R2 EVALUATION METRICS.

  • The multi-class comparison adds F1-score and R2 metrics, reports 95% confidence intervals, and marks the best value in bold.The supplied table caption specifies comparison against pre-trained models but does not provide the values.Source: Table 11 · p. 24

D.3 ABLATION RESULTS

  • The section reports scaling and PAPAGEI-S component analyses, and the accompanying tables describe downstream comparisons and component ablation results.Table 13 covers different PAPAGEI-S models; Table 14 is the component ablation study, but the supplied passages do not provide their cell values.Source: §D.3 · p. 24, Table 13 · p. 25, Table 14 · p. 26

D.4 STATISTICAL SIGNIFICANCE OF MODEL COMPARISON

  • Critical-difference analyses report PaPaGei as significantly better across classification and regression tasks, with Moment strong in both and Chronos and TF-C performing well for classification.The diagrams rank models by average rank; horizontal connector lines indicate no significant differences.Source: §D.4 · p. 26, Figure 22 · p. 26, Figure 23 · p. 26
  • Across 35 repeated comparisons, PaPaGei shows significant improvements in 30 AUROC comparisons and 32 MAE comparisons, with best average ranks of 0.82–0.90 for AUROC and 0.19–0.25 for MAE.Figure 24 summarizes five bootstrap-ranking experiments; colored cells mark significance versus the baseline at p < 0.05.Source: §D.4 · p. 27, Figure 24 · p. 27

D.5 STATISTICAL ASSESSMENT BETWEEN IPA AND SQI

  • Among segments where IPA is unavailable, the IPA group has higher SQI than the no-IPA group in both mean (+0.18) and median (+0.32), with p < 0.05.The permutation test uses 1,000 permutations; the authors interpret this as empirical motivation for SQI when PPG morphology is limited.Source: §D.5 · p. 27

E ADDITIONAL BASELINES: DEMOGRAPHICS & PPG MORPHOLOGY FEATURES

This block reports demographic baselines and prediction diagnostics, then situates PaPaGei’s domain-specific approach and discusses model efficiency, skin-tone robustness, and adaptation.

E ADDITIONAL BASELINES: DEMOGRAPHICS & PPG MORPHOLOGY FEATURES

  • PaPaGei outperforms the demographics-plus-PPG baseline in 14 of 18 tasks, while demographics alone outperform statistical PPG features.In NuMoM2B’s more homogeneous population, demographics add little, and PaPaGei achieves higher AUROC and MAE than supervised baselines.Source: E ADDITIONAL BASELINES: DEMOGRAPHICS &… · pp. 27–28
  • Adding demographics to PaPaGei-S improves results in several tasks, supporting complementary features rather than demographic proxies.The authors note that demographics may be unavailable and cannot alone predict real-time outcomes such as blood pressure or heart rate.Source: E ADDITIONAL BASELINES: DEMOGRAPHICS &… · p. 28

H EXTENDED RELATED WORK

  • Prior PPG models often lacked open release or out-of-domain evaluation; PaPaGei uses a PPG-specific approach and serves as both a feature extractor and multimodal encoder.The proposed encoder–LLM approach avoids requiring paired multimodal training data, but may limit end-to-end gradient propagation and interpretability.Source: H EXTENDED RELATED WORK · pp. 29, 31, 32

F ADDITIONAL PREDICTION PLOTS

G EFFECT OF SKIN TONE

I EXTENDED DISCUSSION

  • PaPaGei outperforms baselines in at least 14 of 20 tasks, with average classification improvements of 4.7%-6.3% and regression improvements of 2.9%-4.9%.PaPaGei-S performs best for cardiovascular parameters, while PaPaGei-P surpasses baseline foundation models on smoking and arousal tasks.Source: I EXTENDED DISCUSSION · p. 32
  • Ablations identify sVRI as the strongest component; the 5M-parameter model outperforms larger alternatives, and PaPaGei-S gains with more linear-probe data.The authors describe larger models as impractical for wearables given their size and cloud-inference privacy concerns.Source: I EXTENDED DISCUSSION · p. 32
  • For nuMoM2B, age range and trimester sampling contribute to distribution shift; suggested responses include domain adaptation, domain generalization, LoRA, and imbalance-aware metrics.The target population is generally aged 20–35, with gestation readings collected around the first and third trimesters.Source: I EXTENDED DISCUSSION · p. 32

Related papers

Loading 2410.20542v2…