Motion-Gated Asymmetric Noise Models for Recovering Pulse Waveforms Through Movement

Wrist-worn optical heart sensors, which measure pulse by shining light into the skin (photoplethysmography, or PPG), fail when the wearer moves. Arm swings and shifts in how hard the sensor presses on the skin produce signal changes far larger than the pulse itself. Walking cadence also sits in the same 1–3 Hz band as the heartbeat, so ordinary frequency filters cannot separate the two. Most current fixes either train a large network on labeled data or discard corrupted stretches, and both approaches assume little about what motion corruption looks like. This proposal starts from a physical observation. Motion corruption in PPG is lopsided. When the sensor presses harder, the capillaries under it compress and the pulse flattens in one direction. When the sensor lifts, ambient light leaks in and the baseline jumps the other way. We therefore model the corruption as a mixture of skewed, heavy-tailed noise components, and we let the direction and strength of each component's skew depend on what the accelerometer says the arm is doing.

The concrete object is a small unsupervised state-space model, fit separately to each recording. It is a probabilistic model in which a hidden, slowly varying pulse oscillator generates the multi-channel optical signal. Motion is modeled as a shared low-rank baseline plus a residual drawn from a mixture of asymmetric Laplace distributions. An asymmetric Laplace density has a sharp peak and a different exponential decay on each side. The skew parameter of each component is not fixed by hand. It is predicted from the accelerometer. If the idea is right, it yields a label-free pulse recovery method with honest uncertainty that holds up in exactly the regime where supervised networks collapse. The regime matters: on mobile video, supervised TS-CAN exceeds 25 bpm error during walking, while unsupervised POS stays at 17.05 bpm (MMPD). A disciplined unsupervised signal model therefore has headroom that deep supervision has not claimed.

Core Claim

The claim has two parts. First, the residual left after removing the pulse and a smooth motion baseline is systematically skewed, and the sign of the skew is predictable from motion context. Second, a noise model that encodes this predictable asymmetry recovers heart rate and beat timing better than models with the same parameter count that are only heavy-tailed (Laplace, Student-t) or only flexible (Gaussian mixtures). The novelty lies in the conditioning: asymmetry becomes a function of measured motion rather than a global constant. Without that conditioning, the idea reduces to transplanting an existing filter into a new domain, and a reviewer would rightly expect a symmetric heavy-tailed filter to do about as well.

Relation to Prior Work

Two papers supply the method.

  • Robust Filtering and Learning in State-Space Models via Asymmetric Laplace Distribution builds exact and fast filters, a smoother, and a single-loop parameter learner for state-space models with asymmetric Laplace measurement noise. It is tested only on synthetic systems and S&P 500 volatility, where it reports roughly 30% lower error than a Kalman filter. It uses one global asymmetry per model and a single channel. We reuse its Gaussian-approximated fast filter and single-loop estimation. We change it to a multi-channel observation, a mixture of asymmetric components, and skew parameters driven by an exogenous signal.
  • ALAMF models image residuals as a mixture of asymmetric Laplace components. It uses the Gaussian–inverse-gamma decomposition to make inference tractable and a stick-breaking prior to choose the number of components. It works on static 2D images. We borrow the decomposition and the stick-breaking prior. ALAMF names manual tuning of its skew parameter τ as a limitation, and our motion gating is a direct answer to it.

Several papers motivate the proposal by showing that motion distortion is structured rather than additive white noise.

  • WF-PPG clamps the wrist sensor at controlled pressures and shows that pressure alone flattens the dicrotic notch and widens the systolic pulse. Sub-optimal pressure raises beat-interval variability (RMSSD) error to 10.95 ms, against 0.13 ms at the fingertip reference.
  • PaPaGei uses pulse skewness as a scalar signal-quality target during pretraining. This implicitly concedes that asymmetry marks corruption. It never models a skewed noise density, and its authors note that ambulatory motion is underrepresented in its clinical corpus. - **Standardizing the Probabilistic Sources of Uncertainty** argues that forcing symmetric likelihoods onto skewed noise makes models mistake data noise for model ignorance. This is why we evaluate calibration and not only error.

  • The over-splitting literature describes how symmetric mixtures stack spurious components along a skewed tail. Hejblum et al. show this in flow cytometry. Deep asymmetric mixture model shows it in histopathology, where the Lizard Dice score is 0.6295 against 0.5095 for a classic GMM. These results predict a measurable symptom we will check: Gaussian mixtures should need more components on the same PPG residuals.

Two strong PPG systems constrain the design.

  • BeliefPPG already represents arbitrary non-Gaussian heart-rate beliefs over 64 bins and prunes motion peaks with a hidden Markov model. It reaches 3.57 bpm on PPG-DaLiA. Adding a parametric asymmetric distribution over heart rate would therefore add little. This is why we put the asymmetry on the time-domain waveform residual, which BeliefPPG never models.
  • Tri-Spectral PPG fuses green, red and infrared channels with a supervised U-Net trained on ECG-averaged pseudo-targets. It reaches 4.5 bpm, against 5.9 bpm for green alone. This shows that multiple wavelengths carry separable artifact information. We try to exploit that information without labels.

The Model

At time $t$, let $y_t \in \mathbb{R}^C$ be the optical channels (three wavelengths where available) and $u_t$ the accelerometer features. We write

$$ y_t = a\, s_t + B\, m_t + \varepsilon_t, \qquad \varepsilon_{t,c} \sim \sum_{k=1}^{K} \pi_k(u_t)\, \mathrm{AL}\big(0,\ \sigma_{k,c},\ \kappa_k(u_t)\big), $$

The symbols are as follows:

  • $s_t$ is a latent pulse oscillator whose phase advances at a slowly drifting heart-rate state.
  • $a$ holds the per-wavelength pulse amplitudes.
  • $m_t$ is a low-dimensional smooth motion baseline with channel loadings $B$.
  • $\mathrm{AL}(0,\sigma,\kappa)$ is an asymmetric Laplace density with scale $\sigma$ and skew $\kappa$.
  • $\pi_k(u_t)$ and $\kappa_k(u_t)$ are the component weights and skews, produced by a tiny gating function of the accelerometer. The gating function is linear or has one hidden layer.

Writing each asymmetric Laplace component as a Gaussian with an exponentially distributed scale and shift keeps each filtering step conditionally Gaussian. The fast filter of Robust Filtering and Learning therefore applies. The gating weights are updated in the same single loop as the other parameters. A stick-breaking prior on $K$, as in ALAMF, lets the model choose how many corruption modes a recording needs.

The posterior probability that each sample belongs to a wide, strongly skewed component is a free artifact-segmentation output. The posterior over heart rate comes with a credible interval.

Research Program

The work proceeds in three stages, and each can stop the project early.

Stage 1: measure the asymmetry directly. We use WF-PPG, where the fingertip channel is a near-clean reference and the load cell records true contact pressure. After beat-aligning wrist and finger signals, we compute wrist residuals and test two things: whether they are skewed, and whether the sign of the skew tracks the sign of pressure change. The prediction is that skew is negative under pressure increases and positive under release. If the residuals are roughly symmetric, or if the skew sign is unrelated to pressure, the core claim is false and we stop.

Stage 2: build and fit the model.

  • On PPG-DaLiA, with one PPG channel plus a three-axis accelerometer at 64 Hz, the model runs with $C=1$ and the accelerometer serves only as the gating input.
  • On the Tri-Spectral in-the-wild recordings, if access can be obtained, and on WF-PPG, the model runs with $C=3$.

Stage 3: plug the model into existing pipelines. The recovered waveform and its per-sample uncertainty become an alternative input to BeliefPPG's heart-rate network. This tests whether a better time-domain front end helps the strongest existing tracker, rather than competing with it on its own terms.

Baselines

Naive reference. The simplest reference is band-pass filtering at 0.5–4 Hz followed by spectral peak picking. It is run on the same windows and scored against the same ECG heart rate. This gives the floor that frequency overlap is known to cap. The file contains no reported number for this exact pipeline on these datasets, so we compute it ourselves.

Strongest heart-rate tracker. On PPG-DaLiA the strongest competitor is BeliefPPG, which reports 3.57 bpm MAE. It is the strongest because it already combines learned spectral features with physiological transition constraints, which is exactly what our state dynamics also provide. We run the released model under its published leave-one-subject-out protocol. Our standalone method uses no labels, so a fair bar for it alone is to approach 3.57 bpm and clearly beat every unsupervised and filtering baseline. Beating BeliefPPG means that BeliefPPG fed our recovered waveform gets a lower MAE than BeliefPPG on raw PPG, under an identical protocol and identical retraining budget, with the gap significant under a paired per-subject test. For context, the multimodal-supervised foundation model reports 7.78 bpm on the same benchmark (A robust PPG foundation model).

Strongest multi-channel method. On multi-wavelength data the strongest competitor is the Tri-Spectral U-Net, which reports 4.5 bpm MAE and 0.831 correlation with the reference waveform. It is the direct supervised analogue of our unsupervised fusion. We train it on the same splits with the same ECG-averaged targets. Beating it means lower MAE on held-out subjects using no labels. Matching it without labels would still be a meaningful result given the MMPD evidence about supervised models breaking under walking. We also report the green-only figure of 5.9 bpm from the same paper as a reference point.

Symmetric filters. The same state-space model is run with Gaussian, Laplace, and Student-t measurement noise. This follows the comparison set in Robust Filtering and Learning, but on physiological data.

Ablation of the key ingredient. This is the most important comparison, and it has four arms that share the same pulse dynamics, the same accelerometer access, the same inference code and a matched parameter count:

  • asymmetric mixture with motion gating (the full model);
  • asymmetric mixture with skews fixed globally (no gating);
  • gated Student-t mixture (gating but no asymmetry);
  • gated Gaussian mixture with $K$ increased to match parameters.

The full model has to beat all three alternatives, or the asymmetry claim fails.

Artifact segmentation. Our component responsibilities are scored against Tiny-PPG, which reports 87.4% DICE on PPG-DaLiA, and SMoLK, which reports 0.8322 DICE on TROIKA. We only aim to be competitive here, since those models are trained on segmentation labels and ours is not.

Why not just…?

Why not just use more Gaussian components? A large enough Gaussian mixture can approximate any skewed density. In mixture models, though, the documented cost is over-splitting: spurious components along the tail, as in cytometry (Hejblum et al.) and histopathology (DAMM). In our model the extra components would also absorb pulse energy. The gated Gaussian mixture with matched parameters is in the ablation. If it matches the asymmetric model on MAE and calibration, we will report that asymmetry is a convenience rather than a necessity. The number of components each variant selects under the stick-breaking prior is a secondary test.

Why not just use a symmetric heavy-tailed filter such as Student-t? This is the most serious threat, because heavy tails alone might explain most of the gain over Kalman filtering. The distinguishing prediction is directional. A symmetric model shrinks one-sided excursions only partway, which leaves a biased baseline and shifted beat fiducials. That bias should show up as higher RMSSD error on WF-PPG under pressure changes even when heart-rate MAE is similar. If the gated Student-t mixture matches the full model on RMSSD and waveform correlation, the claim is refuted.

Why not just swap the supervised network's loss for an asymmetric one? Training the Tri-Spectral U-Net with a pinball (quantile) loss is equivalent to using an asymmetric Laplace likelihood. It costs one line of code, so it is a real alternative. It still needs ECG-derived targets. It also learns one global asymmetry, and it inherits the supervised collapse under unfamiliar motion that MMPD documents. We run it as an extra baseline on the same splits. We expect it to beat the plain U-Net in distribution and to lose to our model on held-out activities. If it does not lose there, the unsupervised framing is not worth its cost.

Why not just rely on BeliefPPG or a foundation model? BeliefPPG outputs heart rate, not waveforms. Beat timing and variability depend on the waveform, and WF-PPG shows that pressure distortion damages RMSSD in particular. PaPaGei's authors acknowledge that ambulatory motion is a gap in its training data. Our plug-in experiment directly tests whether a better waveform helps these systems.

Decisive Checks

Four results would settle the question.

  • Stage 1 residual test on WF-PPG. Skew sign must correlate with the direction of load-cell pressure change. If it does not, the project stops.
  • Gating ablation on PPG-DaLiA and multi-wavelength data. The motion-gated asymmetric model must beat both the fixed-skew version and the gated Student-t version on heart-rate MAE and on RMSSD error, per subject, under paired tests.
  • Calibration check. On PPG-DaLiA's high-intensity activities, the nominal 90% heart-rate intervals must cover the ECG truth close to 90% of the time, while symmetric variants under-cover. This is the observable form of the misattribution argument in Standardizing the Probabilistic Sources of Uncertainty.
  • Plug-in experiment. Feeding recovered waveforms to BeliefPPG must lower its MAE below its raw-input result under an identical protocol.

As a stretch test of generality, we apply the same model to MMPD. There the RGB face traces serve as three channels and face-tracker displacement replaces the accelerometer. The comparison point is POS at 17.05 bpm during walking.