Motion-Aware Wearable Surrogates for Post-Vaccination Immune Waning

Whether a vaccine keeps working is a question about months and years, not days. Neutralizing antibodies fall steeply after the acute response — one 514-person cohort saw geometric mean titers against wild-type SARS-CoV-2 drop from 17.95 at 3 months to 8.00 at 9–12 months before a booster restored them (Immune Persistence against SARS-CoV-2 after Primary and Booster Immunization in Humans) — while cellular memory plateaus for years (Long-term cellular immunity of vaccines for Zaire Ebola Virus Diseases). Yet everything we know about these trajectories comes from sparse, invasive blood draws at scheduled visits. Meanwhile, wrist-worn optical sensors can now measure heart rate, heart-rate variability, and pulse-wave shape continuously in free-living people — but only when deep models fight through motion artifacts, the centimeter-scale body movements that swamp millivolt- and millimeter-scale physiological signals.

No existing work connects these two literatures. Motion-robust wearable models stop at instantaneous cardiopulmonary metrics; immune-kinetics models ingest only intermittent lab assays. This proposal builds the missing link, and its core trick is to repurpose motion artifacts themselves: rather than treating corrupted signal as waste, we use the daily burden and structure of artifacts as an explicit noise model that tells a kinetic model how much to trust each day's wearable estimate of a person's underlying physiological baseline.

Core Claim

A motion-artifact segmenter running on a wearable can supply more than clean heart-rate numbers. If, for each day, it outputs both a physiological baseline estimate (nocturnal and sedentary resting heart rate, heart-rate variability, pulse-wave morphology) and an artifact-dependent confidence, then a state-space extension of established within-host immune-kinetics models can fuse this noisy daily stream with a handful of blood draws to predict individual immune trajectories — antibody waning and booster response — more accurately and with better calibration than sparse draws alone. The signal-level objective changes from "estimate instantaneous heart rate" to "estimate a daily baseline with quantified uncertainty," which is exactly the quantity a slow biological process leaves in autonomic and vascular signals.

Research Program

Three components, each grounded in methods the literature has already validated in isolation.

First, the edge layer. We adapt sample-wise motion-artifact segmentation in the style of Tiny-PPG, which labels corrupted samples rather than whole windows and runs in 206 ms per 30-second segment with only 19,726 parameters on an STM32 microcontroller, achieving an 87.4% DICE score on PPG-DaLiA. Our modification: the segmenter emits a per-day artifact profile (fraction and temporal clustering of corrupted samples, by activity context from the co-recorded accelerometer), and clean segments are embedded by a field-trained encoder. Here we take a side in a live dispute: A robust PPG foundation model using multimodal physiological supervision argues that pretraining on high-grade ICU waveforms with electrocardiogram anchors produces the most transferable features, but Pulse-PPG shows the opposite for deployment in the wild — pretraining on 822,247 uncurated field segments beat the ICU-anchored and PaPaGei morphology-pretrained models on 10 of 11 wearable tasks, with roughly a 6.28 BPM improvement in heart-rate MAE. Because our setting is free-living cohorts, not ICU beds, we build on the field-trained approach; this choice is itself an experimental variable we will test.

Second, the kinetic layer. Modeling COVID-19 vaccine-induced immunological memory development and its links to antibody level and infectiousness established a three-variable ordinary differential equation model — virus, humoral effector, humoral memory — whose central finding is the "six-month rule": memory needs over six months to mature before a booster can elicit high-titer recall. That model is fit to discrete lab measurements and has no interface for continuous data. We extend it into a state-space model with individual-specific parameters and an observation equation:

$$y_{i,t} = g(x_{i,t};\,\theta_i) + \varepsilon_{i,t}, \qquad \varepsilon_{i,t} \sim \mathcal{N}\!\big(0,\ \sigma^2(a_{i,t})\big)$$

where $y_{i,t}$ is person $i$'s wearable-derived baseline on day $t$, $x_{i,t}$ is the latent immune state evolving under the ODE dynamics, $\theta_i$ collects that person's waning and recall parameters, and $a_{i,t}$ is the daily artifact burden from the edge layer, which sets the observation noise. Heavy-motion days contribute less; quiet days contribute more; occasional antibody titers enter as low-noise anchors. Inference is standard filtering, so the whole pipeline runs on a laptop per cohort, not a cluster.

Third, the cohort. We propose a prospective study of roughly 300 adults presenting for a routine booster (for example, a seasonal COVID-19 booster campaign), each wearing a wrist PPG-plus-accelerometry device for twelve months, with blood draws on days 0, 7, and 14, then months 1, 3, 6, 9, and 12 — a schedule dense enough at the start to capture the acute reactogenicity window that Long-term cellular immunity of vaccines for Zaire Ebola Virus Diseases showed is biologically informative (pro-inflammatory cytokines IL-18, IP-10, and TRAIL peaked at day 7 in rVSV recipients and preceded durable T-cell responses out to 60 months), and matched at the back end to the draw schedule of Immune Persistence. A secondary hypothesis, motivated by that day-7 cytokine result, is that a multi-day wearable signature of autonomic strain in the first two weeks post-booster predicts the magnitude of the antibody response — the cohort paper reports booster fold-rises of 3.5 to 6.8, and predicting where an individual lands in that range from a wearable alone would be a clinically useful, assay-free signal.

An equity stress test runs alongside: MMPD showed heart-rate error rising from 3.06 BPM for Fitzpatrick type 3 skin to 15.43 BPM for type 6, exceeding 25 BPM during walking. If artifact burden is systematically misestimated for darker skin tones, our noise model inherits that bias and distorts trajectory estimates exactly where equitable tracking matters most. We will quantify whether trajectory error and calibration differ by skin tone under simulated degradation matched to the MMPD figures.

Baselines

The do-nothing reference is what the field does today: interpolate. We take each participant's log-transformed antibody titers at the scheduled draws and predict held-out draws by linear interpolation and by the population-average waning curve reported in Immune Persistence (17.95 to 8.00 over roughly nine months, with the 6-month optimal booster interval). This baseline defines the information content of sparse draws alone; no reported result found on Lacuna for individual-level titer prediction from sparse draws, so we construct it ourselves on identical held-out draws. Beating it means reducing held-out log-titer error while producing calibrated uncertainty — the population curve cannot do the latter per person.

The strongest cheap competitor is the sleep-only pipeline: discard all motion-corrupted waking data, compute nightly resting baselines, and feed those into the same state-space model with uniform noise. It requires no artifact modeling at all, and because sleep data is comparatively clean it may capture most of the nocturnal signal. It is strongest precisely because it attacks our weakest assumption — that waking artifact-laden data adds information. It will be run on the same cohort, same draws, same inference code, differing only in which days and hours enter the observation stream, so any gap is attributable to the data used, not tuning.

The strongest existing end-to-end method is a two-stage pipeline built from the best published components: motion-robust signal reconstruction with EnhancePPG, which reports 3.54 BPM heart-rate MAE on the ambulatory PPG-DaLiA benchmark (a 12.2% improvement over its 4.03 BPM baseline) and runs on the same class of microcontroller we target, feeding daily features into a direct regression from wearable history to log-titer, with no ODE structure. It gets the same training data, the same draws, and an equal compute and tuning budget via nested cross-validation over its hyperparameters. Beating it means the mechanistic prior earns its keep, which matters because with a few hundred participants a direct regressor has limited data to learn waning kinetics that the ODE already encodes.

The key-ingredient ablation is the uniform-variance model: identical edge features, identical ODE, but $\sigma^2(a_{i,t})$ replaced by a constant. This isolates whether artifact-aware confidence — the one thing no existing system computes — is the source of any gain, rather than the extra data or the kinetic prior.

Why not just…?

Why not just bandpass-filter the motion away? The frequency spectra collide: a walking cadence of 1–3 Hz sits exactly on top of a 60–180 BPM heart rate, so a filter that removes locomotion removes the pulse. The numbers bear this out — on MMPD, the classical POS method degraded from 9.70 BPM MAE when stationary to 17.05 BPM when walking, and leading models exceeded 25 BPM. Our setting is worse than these benchmarks, because twelve months of daily life guarantees sustained overlap regimes.

Why not just use sleep data and throw away everything corrupted? This is the most serious alternative, and we do not argue around it — we run it as a baseline with the same inference machinery. The case for expecting it to fall short is that reactogenicity and immune-related autonomic shifts are not confined to sleep (the day-7 cytokine peaks in Long-term cellular immunity were systemic), and that discarding waking hours shrinks the observation stream enough to widen confidence bands past usefulness. But if sleep-only matches our full pipeline on trajectory accuracy, the honest conclusion is that the artifact segmenter only needs to gate, not to weight, and we would report the simpler design as the winner.

Why not just do more blood draws? If budget were unlimited, denser assays would beat any wearable surrogate — they measure the thing itself. The draws are also our ground truth, so this objection defines our scope: we target interpolation between draws, individual-level deviation from the population curve, an assay-free reactogenicity readout in the first two weeks (where draw schedules in existing cohorts are too coarse to resolve daily kinetics), and scalability to cohorts — Immune Persistence itself notes that major prospective cohorts under-sample immunocompromised and elderly populations, precisely the people for whom an inexpensive continuous surrogate is most valuable.

Why not just fine-tune a foundation model to regress titers directly? With roughly 300 participants and maybe eight draws each, a fine-tuned network has on the order of thousands of supervised targets, an overfitting regime where the ODE prior — including the empirically established six-month memory maturation rule of Modeling COVID-19 vaccine-induced immunological memory — should act as a strong inductive bias. This is not an argument but a testable comparison, and it is exactly the EnhancePPG-plus-regression baseline above. If the direct regressor wins despite the small sample, the mechanistic layer should be dropped.

Proposed Experiments and Decisive Checks

Signal-level validation comes first, on PPG-DaLiA, where our edge segmenter must be competitive with published numbers before any biological claim is made: EnhancePPG reports 3.54 BPM MAE and A robust PPG foundation model reports 7.78 BPM on the same benchmark, so our per-day baseline estimates should not degrade beyond these when aggregated. We also validate the artifact profile itself against Tiny-PPG's 87.4% DICE segmentation labels on held-out subjects, since Tiny-PPG's authors note degradation on unseen subject morphology — a failure mode our per-individual recalibration in the first cohort month must address.

Trajectory-level evaluation on the cohort uses leave-one-draw-out prediction: for each participant and each scheduled draw, predict the log-titer from all other draws plus the wearable stream, and compare log-titer MAE and 90% interval coverage across the four contenders (interpolation, sleep-only, direct regression, artifact-aware state-space). Three checks decide the proposal. The ablation check: artifact-weighted noise must beat uniform noise on coverage and error at equal data and compute; if not, the central mechanism is falsified. The reactogenicity check: the first-two-week wearable strain signature must predict individual booster fold-rise within the 3.5–6.8 range reported by Immune Persistence better than chance and better than day-7 titers alone. The equity check: under MMPD-matched skin-tone degradation, calibrated trajectory error for type 6 conditions must not exceed type 3 conditions by more than a stated margin; if it does, the noise model needs a per-individual artifact-burden recalibration before any population deployment. A negative result on any check is reportable as such — the void between these two literatures is real (Modeling COVID-19 vaccine-induced immunological memory cannot ingest continuous data; Pulse-PPG has never been pointed at a slow biological state), and establishing whether autonomic baselines carry individual immune-kinetic signal at all is the first-order scientific question, whatever the answer.