Benchmarking Open-Source PPG Foundation Models for Biological Age Prediction

N. Brag · 2026

Read Benchmarking Open-Source PPG Foundation Models for Biological Age Prediction by N. Brag (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-14030.txt.

Introduction

Brag benchmarks open-source foundation models that predict biological age from PPG pulse signals. A task-trained model built on 212,231 UK Biobank subjects fails on a new clinical group, with predicted ages squeezed into a narrow band. A general model with no age training does better on the same patients. On 906 surgical patients, Pulse-PPG gives the lowest error of the three models tested. Adding basic demographic data improves results further, and the age gap tracks diastolic blood pressure.

Catalog record

CREATORN. Brag
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25964
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.14030

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Benchmarking Open-Source PPG Foundation Models for Biological Age Prediction N. Brag Abstract A model trained on 212,231 subjects to predict vascular age from PPG (AI-PPG Age, trained on UK Biobank finger PPG) fails when applied to a different clinical population—its predictions collapse to a narrow range (38–67 years) regardless of the patient’s true age. A general-purpose foundation model trained on no age-related objective at all achieves lower error on the same data. We document this reversal, explain its cause, and show that open-source PPG embeddings reliably encode biological aging information across acquisition contexts. We benchmark three open-source PPG models—Pulse-PPG, PaPaGei-S, and AI-PPG Age—on 906 surgical patients from PulseDB (VitalDB clinical finger PPG, 125 Hz, 10-second segments), using frozen embeddings with Ridge regression and 5-fold stratified cross-validation. Pulse-PPG achieves MAE = 9.28 years (R 2 = 0.388), outperforming both the task-specific AI-PPG Age in linear probe mode (MAE = 9.72) and the combination of heart rate variability with all demographic features (MAE = 9.59). Fusing Pulse-PPG embeddings with demographics yields MAE = 8.22 years (R 2 = 0.517, r r = 0.725). The age-adjusted PPG age gap correlates significantly with diastolic blood pressure ( r r = − - 0.188, p p = 1.2 × \times 10 -8 ), consistent with Apple’s finding that PPG morphology captures vascular aging beyond chronological age. The performance gap with Apple’s PpgAge (MAE = 2.43 years) is attributable to dataset scale ( ∼ \sim 236 × \times fewer subjects) and population differences, not model architecture: our learning curve shows no plateau at 906 subjects. All code is publicly available. Keywords: photoplethysmography; biological age; foundation models; PPG embeddings; cardiovascular aging; digital biomarker 1 Introduction Biological age—a measure of physiological deterioration independent of chronological age—is a powerful predictor of mortality, cardiovascular disease, and age-related conditions ( Jylhävä et al. 2017 ) . Traditional biological age clocks rely on blood biomarkers ( Levine et al. 2018 ) , DNA methylation ( Horvath 2013 ) , or clinical assessments, limiting their use to periodic clinical visits. The photoplethysmogram (PPG), a ubiquitous optical signal captured by pulse oximeters and consumer smartwatches, encodes information about arterial stiffness, cardiac output, and autonomic function—all of which change with aging ( Elgendi 2012 ) . In October 2025, Apple introduced PpgAge, a deep learning model trained on 213,593 Apple Health Study participants that predicted chronological age with MAE = 2.43 years from wrist PPG and demonstrated that the PPG age gap (predicted - chronological age) predicted incident cardiovascular disease (HR = 1.46) ( Miller et al. 2025 ) . PpgAge’s reliance on proprietary data (Apple Health Study) and closed-source models, however, limits reproducibility and broader adoption. Two questions remain unanswered: (1) Can open-source PPG foundation models replicate this capability? (2) Does the PPG age signal persist in clinical finger PPG, beyond Apple Watch wrist PPG? Several open-source PPG foundation models have recently emerged. Pulse-PPG ( Saha et al. 2025 ) is a ResNet-based model pre-trained via contrastive learning on wearable PPG data, while PaPaGei ( Pillai et al. 2025 ) uses a mixture-of-experts ResNet architecture pre-trained on 20 million PPG segments from public datasets. Both produce general-purpose PPG embeddings that can be used for downstream tasks via linear probing. A critical question for clinical translation is whether foundation model embeddings capture aging information beyond what can be extracted from simple heart rate (HR) and heart rate variability (HRV) features. Apple reported that HR/HRV alone achieved MAE = 6.1 years on their large dataset (using 30-day aggregated statistics from thousands of segments per subject), compared to 2.43 years for their full model. If foundation model embeddings merely encode HR/HRV, their added complexity would be unjustified. In this work, we benchmark Pulse-PPG and PaPaGei-S for chronological age prediction on PulseDB ( Wang et al. 2023 ) , a public dataset of 5,361 subjects with clinical finger PPG from VitalDB surgical patients and MIMIC-III ICU patients. We evaluate frozen embeddings with Ridge regression, compare against HR/HRV and demographic baselines, and test whether the PPG age gap associates with cardiovascular risk factors, an analysis motivated by the key clinical finding from PpgAge. 2 Methods 2.1 Dataset We use PPG segments from the VitalDB subset of PulseDB ( Wang et al. 2023 ) , comprising clinical finger PPG recordings from surgical patients at Seoul National University Hospital. PulseDB provides pre-segmented 10-second windows at 125 Hz (1,250 samples per segment) with per-segment annotations including age, sex, height, weight, BMI, and intra-operative systolic and diastolic blood pressure (SBP, DBP). From 943 downloaded subject files, 906 were successfully loaded (37 excluded: 29 due to file corruption or malformed h5py object references, 8 due to age < 18). We restrict our analysis to the VitalDB subset, as MIMIC-III subjects in PulseDB lack the complete per-segment demographic annotations (height, weight, BMI, and blood pressure) required for our demographic feature experiments. We use up to 50 segments per subject, sampled evenly across available segments, yielding 41,104 total segments. The age distribution spans 18–92 years (mean 60.9, SD 15.6), with 56.2% male subjects. 2.2 Foundation Models Pulse-PPG ( Saha et al. 2025 ) : A 12-block ResNet1D (128 base filters, kernel size 11, max pooling) pre-trained via contrastive learning on wearable PPG from a 100-day field study with 120 participants. Expects 50 Hz input; we resample from 125 Hz using scipy’s signal.resample. Produces 512-dimensional embeddings directly from the forward pass. PaPaGei-S ( Pillai et al. 2025 ) : A ResNet1D with Mixture-of-Experts (18 blocks, 32 base filters, 3 experts) pre-trained on 20M PPG segments from public datasets using morphology-augmented contrastive learning. Operates at native 125 Hz. Produces 512-dimensional embeddings. AI-PPG Age ( Nie et al. 2025 ) : A 1D CNN with Squeeze-and-Excitation blocks trained end-to-end on 212,231 UK Biobank subjects for direct vascular age regression using a distribution-aware loss function. Input is a single averaged beat waveform (resampled to 100 samples); output is a scalar age prediction. We evaluate this model in two modes: (1) zero-shot , using direct age predictions without any adaptation; and (2) linear probe , extracting 192-dimensional pre-dense layer activations via a forward hook on the final convolutional stage and training Ridge regression on these embeddings. Beat templates are extracted by detecting systolic peaks in each 10-second segment, isolating peak-to-peak waveforms, resampling each to 100 samples, and averaging across all valid beats in the segment. All three models are used as frozen feature extractors—no fine-tuning is performed. Input segments for Pulse-PPG and PaPaGei-S are z-score normalized per segment before embedding extraction. 2.3 HR/HRV Feature Extraction To contextualize foundation model performance, we extract traditional heart rate and heart rate variability features from each 10-second PPG segment. Systolic peaks are detected using scipy’s find_peaks with minimum inter-peak distance of 0.4 s (corresponding to 150 BPM maximum) and minimum prominence of 0.1. Inter-beat intervals (IBI) are computed from successive peak locations and filtered to the physiologically plausible range of 0.3–2.0 s (30–200 BPM). Segments with fewer than 4 detected peaks or fewer than 3 valid IBIs are excluded. From valid segments, we extract 7 features: mean heart rate (HR), HR standard deviation, HR range, mean IBI, standard deviation of normal-to-normal intervals (SDNN), root mean square of successive differences (RMSSD), and the proportion of successive

Read the complete text