Wrist-worn optical heart sensors measure blood volume with light. This signal is called photoplethysmography, or PPG. It fails during everyday movement, and one reason is that the sensor presses on the skin harder or softer as the arm swings. That changing contact pressure reshapes the pulse waveform. The models that clean up PPG are trained on two kinds of data: small field cohorts of 10–16 people, or large hospital recordings of patients lying still. Neither contains a controlled account of how pressure distorts the pulse. We propose to fit a small, physically parameterized model of that distortion from one existing dataset that measured pressure directly. We then use it as a simulator that turns clean hospital pulses into realistic "pressed-on-a-moving-wrist" pulses, and test whether this targeted synthetic data helps more than generic noise augmentation when only a handful of real ambulatory subjects are available.
The opportunity is narrow and testable. WF-PPG recorded 27 subjects with a load cell under the wrist sensor and a fingertip reference, across six controlled pressure levels. Sub-optimal pressure raised heart-rate-variability error from 0.13 ms to 10.95 ms, and the best pressure differed from person to person. That dataset contains exactly the paired measurements a simulator needs, yet it has only been used descriptively. A second payoff concerns sensor hardware. The same model tells us whether a pressure sensor built into the optical face of the device would be worth fabricating, and at what resolution and bandwidth.
Core Claim
The claim is that a contact-pressure distortion model fit to real paired data (real-to-sim) will improve held-out-subject heart rate estimation on small ambulatory benchmarks more than three alternatives. Those alternatives are unstructured augmentation of matched magnitude, accelerometer-conditioned networks, and plain fine-tuning of PPG foundation models. The gain should be largest when the number of real target subjects is smallest.
This differs from the closest work in identifiable ways:
- WF-PPG measured the pressure effect but built no predictive model.
- Ba et al. used a pulse estimator to keep synthesized dark-skin face video physiologically faithful. That setting is camera-based and has no contact mechanics. We borrow its principle that synthetic data must preserve the underlying pulse.
- The multimodal-supervision PPG foundation model and PaPaGei pretrain on hospital and sleep data. The PaPaGei authors themselves note that in-the-wild motion data is underrepresented. We do not replace these models; we test whether pressure-simulated data closes part of their ambulatory gap.
The file reports no differentiable tissue–optics simulator for wearables. We deliberately do not build a finite-element plus photon-transport simulator. A learned, low-dimensional model fit to real measurements is cheaper, and it is directly checkable against held-out real data.
Research Program
The distortion model. Let $x(t)$ be a clean reference pulse (the fingertip PPG in WF-PPG). Let $p$ be the measured wrist contact pressure, $\lambda \in$ {green, red, infrared} the wavelength, and $p^*_s$ a subject-specific optimal pressure, treated as a random effect because WF-PPG found no universal optimum. We fit
$$y_\lambda(t) = a_\lambda(p - p^*_s)\,\big[h_\lambda(\cdot\,;\, p - p^*_s) * x\big](t) + b_\lambda(t;\, p - p^*_s) + \epsilon(t),$$
with the following components:
- $y_\lambda$ is the observed wrist pulse.
- $a_\lambda$ is a pressure-dependent amplitude gain.
- $h_\lambda$ is a short, pressure-dependent filter applied to the pulse. It captures morphology changes such as loss of the dicrotic notch, the secondary bump after the main pulse peak.
- $b_\lambda$ is a slow baseline term.
- $\epsilon$ is residual noise whose spectrum is estimated from the fit.
Each of $a_\lambda$, $h_\lambda$ and $b_\lambda$ is a small monotone-constrained network of one scalar, the pressure offset. The model is deliberately low-capacity so that 27 subjects can constrain it. We fit it with subject-level cross-validation and hold out a set of WF-PPG subjects entirely for fidelity checks.
Dynamic pressure from motion. WF-PPG varies pressure statically, so the model does not by itself say how pressure fluctuates while walking. We drive $p(t)$ from wrist acceleration through a damped spring model of the strap. Its stiffness, damping and resting pressure are sampled from broad priors rather than fit. This step is the main remaining sim-to-real risk. We therefore report results under several prior widths, and treat sensitivity to these priors as a first-class result rather than a nuisance.
Synthetic corpus. The simulator is run on clean, subject-labeled hospital PPG, specifically MIMIC-III segments as used by the foundation-model papers. Pressure trajectories come from accelerometer traces taken from the training folds of the target benchmarks. Because labels come from the clean source, heart rate is known exactly. This mirrors the label-preserving logic of Ba et al.
Hardware specification study. On WF-PPG, where true pressure is available, we feed measured pressure to the estimator as an extra input. We then progressively degrade that input by quantization, low-pass filtering and added noise. The resulting curve of error against pressure-signal quality gives a concrete requirement for an integrated miniature piezoresistive sensor. The fabrication work in the single-side micromachined pressure sensor and the micromachined haircell flow sensor shows such structures can be made. Our study asks whether making one for a PPG face would pay off. If error barely changes when the measured pressure input is removed, we will say so.
Baselines
Naive reference. An FFT peak-picking heart rate estimator on the raw wrist channel. WildPPG reports 17.62 bpm mean absolute error (MAE) for this kind of baseline on its data. It anchors how much any learned method helps.
Strongest existing methods. We use three, each tied to the benchmark where it was reported:
- Multimodal-supervision foundation model. On PPG-DaLiA, this 1D ResNet-26 pretrained on 56,000 hours of MIMIC-III with ECG and respiration anchors reports 7.78 bpm MAE, against a previous best of 10.5 bpm, in A robust PPG foundation model using multimodal physiological supervision. It is the strongest competitor because it already targets robustness to noisy PPG at scale.
- Temp-ResNet. On WildPPG, it reports 8.37 bpm and is the reference there.
- Tri-Spectral U-Net. On the Tri-Spectral dataset, Tri-Spectral PPG reports 4.5 bpm and a waveform correlation of 0.831.
We re-run each under our own leave-subject-out splits rather than relying only on reported numbers, since split protocols are not documented in the file. Each baseline gets the same target-subject budget, the same tuning budget (an equal number of hyperparameter trials selected on inner-fold validation subjects), and comparable compute. Our method is applied as an augmentation on top of the same backbones, so the comparison isolates the data, not the architecture. Beating a baseline means a lower held-out-subject MAE with a subject-level bootstrap 95% interval that excludes zero difference, at the small target sizes we care about (2–8 subjects).
Accelerometer-conditioned network. The same backbone receives the three-axis accelerometer as extra input channels. This is the standard way to give a model motion information.
Key ablation. We keep the full pipeline but replace the fitted pressure model with an unstructured perturbation: random gain, random short filter and random baseline. These are matched to the simulator's output in per-wavelength variance and spectrum. If this matches our result, the physics adds nothing beyond "more noisy data." A second ablation shuffles the pressure–distortion pairing, breaking the link between motion and distortion while keeping marginal statistics.
Why not just…?
Why not just fine-tune a PPG foundation model on the target cohort? This is the most serious competitor, and we use it as both a baseline and a backbone. Its pretraining is hospital-weighted, and PaPaGei's authors flag in-the-wild motion as underrepresented. Still, 7.78 bpm on PPG-DaLiA shows fine-tuning already absorbs much of the shift. The learning-curve experiment settles this question directly. If fine-tuning alone matches fine-tuning plus pressure simulation at every target size, the simulator is not worth its complexity for heart rate. In that case we would report that outcome, and restrict any remaining claim to waveform morphology and RMSSD, where pressure effects are largest in WF-PPG.
Why not just feed the accelerometer to the network? This is a real threat, because our dynamic pressure is itself computed from acceleration. The argument for simulation is about data efficiency, not information content. With 10–16 subjects, a network must discover the acceleration-to-distortion mapping from scratch, whereas the simulator supplies it from 27 other people with measured pressure. We therefore expect the accelerometer baseline to close the gap as target subjects increase. If it matches us even at 2–4 subjects, the prediction fails.
Why not just add recorded or random noise to clean signals? The matched-magnitude ablation answers exactly this question. The evidence for structure mattering is suggestive but indirect. In WildPPG, adding a physically meaningful input (temperature) gave only a small gain, 8.62 to 8.37 bpm. Wi-GATr shows that physical structure built into the model buys sample efficiency in wireless channel simulation. Neither result proves structured augmentation beats unstructured noise for PPG, so the ablation is decisive.
Why not just collect more subjects? WildPPG needed 13.5 hours per subject to reach 16 people. A simulator fit once to existing data costs little to test. A negative result here would also tell future data collectors that pressure logging is the measurement to add.
Proposed Experiments
Simulator fidelity. We use the held-out WF-PPG subjects to compare:
- simulated against real wrist pulses at each pressure level, by per-wavelength waveform correlation and dicrotic-notch presence;
- RMSSD error computed on simulated pulses against the reported 10.95 ms error under sub-optimal pressure;
- a two-sample classifier distinguishing real from simulated beats, where accuracy near chance indicates a small sim-to-real gap.
Main test. We run leave-subject-out heart rate estimation on PPG-DaLiA (15 subjects), the WildPPG wrist channel (16 subjects) and the Tri-Spectral data (10 subjects), with 1, 2, 4, 8 and all available target training subjects. We report MAE with subject-bootstrap intervals and across-fold variance, because at these sizes the spread across folds matters as much as the mean.
Pre-registered predictions. We make three:
- Our method beats the matched-noise ablation and the accelerometer baseline at 2–8 subjects.
- The advantage shrinks as target subjects grow.
- Results degrade gracefully, rather than collapsing, as the spring-model priors widen.
Hardware curve. The oracle-pressure degradation study on WF-PPG supplies the specification for, or the case against, an integrated pressure sensor.
Limitations. All cited wearable cohorts are restricted to Fitzpatrick skin types 1–3, and our simulator does not model skin pigment. We therefore make no fairness claim and report this as an explicit gap.
References
- WF-PPG: A Wrist-finger Dual-Channel Dataset for Studying the Impact of Contact Pressure on PPG Morphology
- Overcoming Difficulty in Obtaining Dark-skinned Subjects for Remote-PPG by Synthetic Augmentation
- A robust PPG foundation model using multimodal physiological supervision
- PaPaGei: Open Foundation Models for Optical Physiological Signals
- A Single-Side Micromachined MPa-Scale High-Temperature Pressure Sensor
- Bio-Inspired Micromachined Volumetric Flow Sensor with a Big Dynamic Range for Intravenous Systems
- WildPPG: A Real-World PPG Dataset of Long Continuous Recordings
- Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring
- Differentiable and Learnable Wireless Simulation with Geometric Transformers