When the Signal Fades, the Pulse Turns Healthy: Rarity-Weighted Waveform Recovery Under Camera Shift

Cameras can recover the full shape of the pulse from a face video, not just the heart rate. That shape includes the sharp systolic upstroke, the small dip when the aortic valve closes (the dicrotic notch), and the reflected diastolic wave. It carries clinical information about arterial stiffness, blood pressure, and arrhythmia. Current shape-recovery networks are trained almost entirely on young, healthy volunteers recorded in good conditions. We argue this creates a specific, measurable failure. A network that has seen mostly one kind of pulse learns that pulse as a prior. When the input gets noisier, as it does with phone cameras, darker skin, motion, or video compression, the network leans on that prior more heavily. Its output then drifts toward a canonical healthy beat. The patients whose pulses look unusual are the ones a shape-based tool exists to serve, and they are exactly the ones whose waveforms get erased first.

This proposal treats the three directions as one mechanism. Class imbalance appears here at the level of waveform shapes rather than diagnostic labels. Domain shift acts as the stressor that exposes it, because lower signal quality increases reliance on the imbalanced prior. Morphology recovery is the task where the damage happens. We propose to measure this rarity-by-noise interaction directly, and then to test whether two training changes remove it without hurting typical beats. The first reweights training beats by how rare their shape is, using no labels. The second adds synthetic pathological pulses.

Core Claim

Our hypothesis is that, for current morphology-recovery models, reconstruction error on rare beat shapes grows faster with falling signal-to-noise ratio than error on common shapes. A label-free reweighting of training beats by inverse shape density, which we call rarity weighting, flattens that slope. It should do so more effectively than rhythm-label class weighting or than simply adding pathological data. If the interaction does not exist, the premise is wrong and we will report that as the main finding.

The closest prior work handles pieces of this but not the combination. rPPG-SysDiaGAN addresses an imbalance inside the cardiac cycle. It uses second-derivative losses and discriminators so the short notch region is not swamped by the long systolic wave, and it reports Pearson 0.915, RMSE 0.167, and Fréchet distance 0.248 on UBFC-rPPG. That imbalance is across time within healthy beats; ours is across beats and patients, and SysDiaGAN was evaluated only on healthy cohorts without hardware or skin-tone shift. SCAMPS renders synthetic faces whose pulses are built from Gaussian windows for systolic and dicrotic peaks. By design these are healthy pulses only. MMPD documents the shift we use as a stressor: TS-CAN heart-rate error rises from 3.06 BPM on Fitzpatrick skin type 3 to 15.43 on type 6, and exceeds 25 BPM while walking. It measures heart rate only, not shape. Imbalance-aware domain adaptation for embryo assessment and LADA motivate our concern. Both show that alignment under label imbalance pulls minority samples into majority regions, which LADA calls smearing. However, both study image classification with discrete labels, not continuous waveform reconstruction where the "classes" are unlabeled shape variations.

Research Program

The first component is a shape space and a rarity weight. We train a small autoencoder on single beats of clean contact PPG from MIMIC-III and VitalDB, the clinical waveform sources used by LSCT. Each beat is first resampled to a fixed length so the embedding reflects shape rather than rate. A kernel density estimate $\hat p$ over the embeddings $z(x)$ gives each training beat $x$ a weight

$$w(x) = \frac{\hat p(z(x))^{-\alpha}}{\mathbb{E}\left[\hat p(z)^{-\alpha}\right]},$$

where $\alpha \in [0,1]$ interpolates between no reweighting ($\alpha=0$) and full inverse-density weighting ($\alpha=1$). Weights are clipped at a fixed percentile to avoid memorizing single noisy beats, the overfitting risk the file attributes to naive duplication. The reconstruction loss, including derivative terms following the attention-network approach and SysDiaGAN, is multiplied per beat by $w$. This choice is deliberately label-free. Morphological rarity such as an absent notch from stiff arteries, damped pulses, or individual reflection patterns is mostly unannotated, so rhythm labels capture only part of it.

The second component is pathology-aware synthetic pulses. We fit SCAMPS-style Gaussian-window pulse parameters to the rare tail of the clinical beat density. That tail includes ectopic beats, irregular intervals, and missing or shifted notches. We then render or inject these into synthetic training sequences, so rare shapes appear under controlled skin tone and lighting.

The third component is a direct measure of prior collapse. For a reconstructed beat $\hat y$, true beat $y$, and the population-mean beat $\bar y$ from training, we define

$$C = \rho(\hat y, \bar y) - \rho(\hat y, y),$$

where $\rho$ is Pearson correlation. Positive $C$ means the output looks more like the average training pulse than like the patient's actual pulse. We report $C$ alongside RMSE, dynamic-time-warping distance, notch-presence F1, notch timing error, and second-derivative landmark ratios. All metrics are stratified by shape-density decile and by signal-to-noise ratio.

Baselines

The naive reference point is the established color-projection method POS followed by a 0.5–3.0 Hz bandpass filter. MMPD found POS more robust than poorly generalizing deep models, which makes it a fair floor. However, the morphology papers in the file note that this filter rounds peaks and flattens the notch. We expect POS to show low collapse only because it carries no learned prior, and poor landmark accuracy overall. It needs no tuning beyond the standard filter band.

The strongest existing method is rPPG-SysDiaGAN, with its Swin-based U-Net generator and time, frequency, and second-derivative discriminators. Its reported UBFC-rPPG reconstruction figures (Pearson 0.915, RMSE 0.167, Fréchet distance 0.248) are the published target we must first reproduce within tolerance on that dataset. We train it on the same data mix as our model, with the same number of tuning trials and the same compute budget. Our method is SysDiaGAN's architecture plus rarity weighting, so the comparison isolates the training change. Beating it means a statistically significant reduction in worst-decile notch timing error and collapse index at the lowest two noise levels. At the same time, common-decile error must stay within 5% of baseline and heart-rate error must not get worse.

The second strong competitor is the codebook model LSCT, which anchors outputs to a learned dictionary of valid waveform templates. Our hypothesis is that a frequency-allocated codebook under-represents rare shapes. No reported numerical result was found on Lacuna for LSCT on MIMIC-III, VitalDB, or OML, so we will retrain it on our splits. We will also test a variant whose codebook is fit with the same rarity weights. If the weighted codebook matches our method, that supports the mechanism rather than undermining it.

The key ablations hold architecture fixed and vary only the training signal. They include no weighting, rhythm-label class weighting, rarity weighting, synthetic rare pulses alone, and synthetic plus rarity weighting. A sweep over $\alpha$ checks that gains rise and then saturate or reverse, rather than appearing only at one tuned value.

Why not just…?

Why not just class-weight or focal-weight by rhythm label? This is the obvious move, and we run it. Its limitation is coverage, not mechanism. Rhythm labels such as atrial fibrillation or premature beats tag a minority of unusual shapes, while stiffness-related and individual morphology variants carry no label. The ablation that compares label weighting with rarity weighting, stratified by labeled versus unlabeled rare beats, will show whether label weighting fixes only the beats it can see. If it performs equally well on unlabeled rare beats, rarity weighting adds nothing and we will say so.

Why not just add pathological data to training without reweighting? This is the most serious threat. Including MIMIC-III and VitalDB beats in the unweighted baseline is already part of our design for exactly this reason. Even with that data present, the shape density stays heavily concentrated, and the noise-driven prior reliance we hypothesize should persist. If the unweighted model with pathological data shows no rarity-by-noise interaction, the problem was missing data rather than imbalance. That outcome would redirect the work toward data collection.

Why not just extend SCAMPS with pathological pulse templates? This is plausible and is our second component, tested on its own. Gaussian-window templates are a simplification. The canine femoral PPG study notes that models with too few wave components miss reflection dynamics, so synthetic rare shapes may help on the forms we can parameterize but not on others. The synthetic-only arm against the rarity-weighted arm, evaluated on real rare beats, settles which matters more.

Why not just use a stronger morphology model, such as SysDiaGAN's derivative discriminators or a codebook? Both are included as baselines. Neither changes how training beats are counted, and a codebook with frequency-based allocation may make collapse worse by snapping noisy rare beats onto common templates. The rarity-weighted codebook variant tests this directly.

Proposed Experiments

The first testbed is semi-synthetic, because no public video dataset pairs diverse skin tones with pathological pulses. We take clean contact PPG beats from MIMIC-III and VitalDB and corrupt them with a noise model calibrated on MMPD. The noise is the residual between POS output and finger PPG, estimated separately for each skin type, lighting condition, and motion level. We add a low-pass filter matched to compression. This yields known ground-truth morphology for rare beats at controlled noise levels that mimic the domain shift from laboratory to phone. We split by patient and, separately, train on MIMIC and test on VitalDB to add an institutional shift.

The decisive check comes first and does not depend on our method. We fit the baseline's error as a function of shape-density decile and noise level. The premise requires a positive interaction: the rare-versus-common error gap and the collapse index $C$ must grow as signal-to-noise falls, with confidence intervals from patient-level bootstrap. The method then succeeds only if it significantly reduces that interaction slope.

The second testbed is real video. We train on PURE, UBFC-rPPG, and SCAMPS, with or without pathological synthetics, and test on MMPD by skin type and motion against finger-PPG morphology. MMPD subjects are healthy, so rarity here reflects individual shape differences. We check whether the method lowers collapse on skin types 5–6 without raising heart-rate error relative to the reported TS-CAN figures. If the ambulatory cardiology cohort can be accessed, it provides the real pathological test. That cohort has 143 patients, a 71% hypertension rate, and atrial fibrillation, ectopy, and paced rhythms, and its hybrid camera model reports a systolic blood pressure MAE of 11.35 mmHg. We would test whether reduced collapse carries through to lower blood-pressure error and better arrhythmia-beat fidelity, and we will note that the cohort is 80% White as a remaining limitation.