Telling Sensor Drift from Antibody Waning in Wearable Immune Tracking

When a vaccine fades, protection decays over months to years, yet almost everything we know about that decay comes from a handful of clinic visits with a needle. The pivotal mRNA-1273 trial reported 93.2% efficacy over a median follow-up of only 5.3 months Efficacy of the mRNA-1273 SARS-CoV-2 Vaccine at Completion of Blinded Phase, and by day 180 after infection or vaccination roughly half of participants in a 2009 H1N1 cohort had already lost sufficient neutralizing activity Antibody Dynamics of 2009 Influenza A (H1N1) Virus in Infected Patients and Vaccinated People in China — a collapse that quarterly blood draws can easily miss. A wearable biochemical sensor that reads antibody levels in interstitial fluid could fill this gap, but it creates a problem nobody has solved: over months of wear, the sensing interface itself degrades, and the resulting signal drift is nearly indistinguishable from the biological decay we want to measure.

This proposal builds such a sensor and, more importantly, solves the identifiability problem it creates. The plan combines three ingredients that currently live in separate literatures: covalent interfacial bonding as the enabling materials mechanism; longitudinal antibody-decay tracking as the application; and small-sample statistics as the regime in which the whole thing must actually be evaluated. The core idea is a dual-channel sensor in which a target antigen and an invariant reference receptor are both permanently co-polymerized into a hydrogel network during curing, so that the reference channel reports pure interface decay while the target channel reports interface decay plus true antibody decay. A joint statistical model then separates the two, even when the cohort is small (N < 30) and wear times are irregular.

Core Claim

The claim has a materials half and a statistical half, and each half is what makes the other one work. On the materials side: anchoring both capture antigen and reference receptor by permanent covalent bonds formed during network curing — the same strategy that lets vinyl-functionalized silica bond directly into silicone rubber during vulcanization Effects of Vinyl Functionalized Silica Particles on Thermal and Mechanical Properties of Liquid Silicone Rubber Nanocomposites and that let cellulose nanocrystals crosslink into epoxy past 50 wt.% loading without phase separation Printable, castable, nanocrystalline cellulose-epoxy composites exhibiting hierarchical nacre-like toughening — keeps receptor leaching low enough, for long enough, that the reference channel is a usable drift gauge. On the statistical side: a Bayesian joint model in the tradition of Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data, which jointly models biomarkers, treatment, and visit times, can be extended with a fourth process for the sensing interface, and the dual-channel design is exactly what makes that interface process estimable rather than hopelessly confounded with biology.

Formally, let $\theta(t)$ be the true antibody level, $\beta(t)$ the fraction of still-active receptor on the interface, and $\kappa$ a fixed reference analyte loading. The two channels measure

$$y_{\text{target}}(t) = \beta(t)\,\theta(t) + \varepsilon, \qquad y_{\text{ref}}(t) = \beta(t)\,\kappa + \varepsilon',$$

with $\varepsilon, \varepsilon'$ measurement noise. The reference channel identifies $\beta(t)$; the target channel then identifies $\theta(t)$. With a single channel, $\beta$ and $\theta$ enter only as a product and cannot be separated — which is precisely why prior wearable sensors cannot distinguish drift from biology.

Research Program

The program has three linked aims, and we state up front which direction carries which weight: covalent bonding is the mechanism, immunity decay is the application, and small samples are the evaluation regime that shapes every design choice.

The first aim is the sensing substrate. We build on the dynamic covalent hydrogel of Multifunctional hydrogel sensors with dynamic covalent networks for machine learning-assisted Parkinson's disease diagnosis and encrypted human-computer interaction, whose imine bonds between oxidized carboxymethyl cellulose and gelatin methacryloyl give self-healing and drift resistance under cyclic deformation. That paper, however, is the closest existing work and it stops at physical sensing: its LSTM reads tremor and electrophysiology, with no biochemical receptor, no external cohort validation, and no treatment of small-sample risk. We keep its dynamic network for mechanical compliance on skin — rigid permanent networks like the vulcanized silica composite lack the aqueous compliance skin contact needs — but we add a second, orthogonal permanent chemistry for the receptors: a capture antigen (SARS-CoV-2 RBD to start) and an inert reference protein, both engineered with methacrylate handles so they co-polymerize into the network during curing rather than equilibrating in a reversible imine exchange. The dynamic imine literature's own reversible equilibrium is the risk here: over months of wear a dynamically tethered receptor can progressively leach or hydrolyze, which is the tension this hybrid design exists to resolve. Transdermal access to interstitial fluid follows the microneedle architecture of Fabrication of Flexible Microneedle Array Electrodes for Wearable Bio-Signal Recording.

The second aim is the joint model. We extend the three-process Bayesian joint model of Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data — which models biomarker trajectories, treatment assignment, and visit times, and proved via simulation that ignoring the visit process produces severe bias — with an explicit interface-state process $\beta(t)$ informed by the reference channel. This matters because wearable compliance is the textbook case of informative observation: people put the sensor on when they feel ill, so wear times are confounded with health status in exactly the way that paper documents for clinic visits. Prior immune-decay studies assume fixed follow-up schedules, from the 6-month BA.5/BF.7/XBB breakthrough cohort Longitudinal Analysis of Humoral and Cellular Immune Response up to 6 Months after SARS-CoV-2 BA.5/BF.7/XBB Breakthrough Infection and BA.5/BF.7-XBB Reinfection to the 5-year Ebola vaccine follow-up Long-term cellular immunity of vaccines for Zaire Ebola Virus Diseases, and none models a decaying instrument in the loop.

The third aim is a small-sample evaluation protocol. Algorithmic Accountability in Small Data: Sample-Size-Induced Bias Within Classification Metrics proves that confusion-matrix metrics are combinatorially unstable below roughly $n \ge 100$: single misclassifications cause discrete jumps, and metrics like Matthews correlation have $4n$ undefined cases. Our human pilot will sit far below that threshold, so every reported classification number (e.g., "high vs. low responder") will be Cross-Prior Smoothing-stabilized using the aggregated HR-VILAGE resource as the well-powered reference group, echoing how that paper's smoothing reduced mean squared error across all 15 metrics tested on census data. We do this with eyes open to the failure mode the smoothing authors themselves flag: if the reference group is structurally mismatched, smoothing imports bias. HR-VILAGE-3K3M is the right reference precisely because it harmonizes 66 longitudinal immunization studies spanning 3,178 subjects — but its own authors found that batch correction with ComBat suppressed genuine responder/non-responder variance, so we will stratify the reference by responder status rather than smoothing against a pooled distribution. The same resource supplies decay-shape and responder-classification priors for transfer: its benchmark shows a simple PCA-based logistic regression reached an F1 of 0.822 and an AUC of 0.748 for predicting vaccine responsiveness, outperforming more complex architectures — a caution against deep models on small immune cohorts that echoes the overfitting documented at N=6 for surface electromyography gesture classification Hand Motion Classification Using a Multi-Channel Surface Electromyography Sensor, N=8 for fMRI speech-in-noise decoding Neural bases of listening to speech in noise, and N=14 for EEG learning-style classification Design and Implementation of an EEG-Based Learning-Style Recognition Mechanism.

A smaller computational sub-aim asks where on the antibody to put the permanent tether. Non-directed conjugation randomly crosslinks into binding loops and destroys the paratope — the same chemistry-versus-function tension that Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design faces when optimizing soluble antibodies with no sensor surface in the loop. We will screen whether existing covalent docking can predict crosslink placement: CovDocker achieves a 37.2% docking success rate at RMSD < 2.0 Å on its own protein-ligand benchmark (versus 6.1% for Vina and 9.9% for Smina), but it assumes rigid receptors and knows nothing of polymer crosslinkers, so we expect it to transfer poorly; a negative result here is itself informative and would justify site-specific enzymatic anchoring at an engineered distal tag as the default route.

Baselines

The "do nothing" reference is physically adsorbed antigen on the identical hydrogel and microneedle hardware, with raw confusion-matrix metrics and a standard linear mixed-effects model that treats wear times as ignorable. This is the naive point that current practice implicitly represents, and the materials literature predicts it fails: physical blends interact only through van der Waals forces and mechanical interlocking, and under cyclic strain or thermal cycling they suffer interfacial slippage and phase separation, which is why vinyl-functionalized silica needed direct covalent bonds to transfer load at all Effects of Vinyl Functionalized Silica Particles on Thermal and Mechanical Properties of Liquid Silicone Rubber Nanocomposites, and why physically mixed nanofillers cannot sustain high loadings the way covalently crosslinked cellulose could past 50 wt.% Printable, castable, nanocrystalline cellulose-epoxy composites exhibiting hierarchical nacre-like toughening.

The strongest materials competitor is the dynamic covalent hydrogel sensor of Multifunctional hydrogel sensors with dynamic covalent networks for machine learning-assisted Parkinson's disease diagnosis and encrypted human-computer interaction, the only paper in the corpus that pairs a dynamic covalent network with an on-device machine learning pipeline. We will run its tethering scheme — receptors held in the reversible imine network — on our own substrate, at matched receptor loading, in the same 180-day wear protocol and the same readout electronics, so the only difference is permanent versus dynamic anchoring of the bioreceptor. Beating it means measurably lower reference-channel drift and higher retained binding capacity at the 180-day horizon.

The strongest statistical competitor is the Bayesian joint model of Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data, which achieved near-zero bias in simulations where models ignoring the visit process were severely biased and overconfident. We will fit its three-process model (biomarker, treatment, visit times) without our interface process, on the same wearable data, with the same priors, MCMC budget, and tuning effort, plus the simpler linear mixed-effects alternative. Because that model has no notion of a decaying instrument, the decisive comparison is bias in the recovered antibody half-life under ground-truth decay schedules. For responder classification, the competitor is PCA-based logistic regression with HR-VILAGE pre-training, which reports F1 = 0.822 and AUC = 0.748 on that benchmark HR-VILAGE-3K3M; we beat it only if the wearable-derived decay features improve smoothed, cross-validated classification in our pilot, not raw metrics that the small-data paper shows are unstable below n = 100.

Two ablations isolate the key ingredients. The first removes the permanent covalent anchor (dynamic tether only) while keeping the dual-channel hardware and full joint model — it tests whether the statistics alone can compensate for a leaching interface. The second keeps the permanent anchor but removes the reference channel and the interface process — a single-channel sensor with the three-process model — testing whether the materials fix alone, without explicit drift modeling, suffices.

Why not just…?

Why not just keep using blood draws and ELISA? This is the most serious objection, and we concede the point on accuracy: laboratory neutralization assays are more specific than any interstitial-fluid readout, and well-run prospective cohorts such as the 12-month booster-persistence study Immune Persistence against SARS-CoV-2 after Primary and Booster Immunization in Humans: A Large-Scale Prospective Cohort Study set the standard. But the same literature shows the cost: multi-year venipuncture cohorts suffer attrition, and compressed windows are the norm — 5.3 months of median follow-up in the mRNA-1273 blinded phase Efficacy of the mRNA-1273 SARS-CoV-2 Vaccine at Completion of Blinded Phase, with the H1N1 cohort showing that half the participants' neutralization was gone within 180 days Antibody Dynamics of 2009 Influenza A (H1N1) Virus in Infected Patients and Vaccinated People in China. Our design is deliberately hybrid: sparse venous checkpoints anchor calibration, and the wearable fills the gaps between them. The decisive check is whether wearable-plus-two-draws recovers the decay inflection that two draws alone provably cannot.

Why not just physically adsorb the receptor and skip the covalent chemistry? Because the corpus is unambiguous that non-covalent interfaces fail under exactly our conditions: cyclic strain, moisture, and months of wear cause debonding and slippage in physical blends Effects of Vinyl Functionalized Silica Particles on Thermal and Mechanical Properties of Liquid Silicone Rubber Nanocomposites, and the hydrogel sensor paper itself needed covalent networks to eliminate signal drift during continuous deformation Multifunctional hydrogel sensors with dynamic covalent networks for machine learning-assisted Parkinson's disease diagnosis and encrypted human-computer interaction. Still, we do not argue this on citations alone — the adsorption baseline and the 180-day drift comparison are designed to settle it empirically, and if adsorbed receptors hold calibration as well as covalently anchored ones, that result would simplify the design and we would report it.

Why not just use standard mixed-effects models or G-computation without the extra processes? The joint-modeling paper answers this directly: when observation timing depends on health status, models that ignore the visit process produced severe bias and overconfidence in simulation, while the full joint model recovered near-zero bias Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data. Wearable compliance is, if anything, more health-dependent than clinic attendance. The honest residual risk is the opposite one: if a simulation shows wear times in our pilot are effectively non-informative, the extra machinery is unnecessary overhead, and our first experiment is powered to detect exactly that so we can fall back to the simpler model.

Why not just regularize the classifier (ridge, L1, dropout) instead of invoking metric smoothing? Regularization acts on model weights during training; it does nothing for the evaluation metric itself. The small-data paper proves that confusion-matrix metrics are combinatorial objects that stay jagged and sometimes undefined below the empirical stability threshold of about $n \ge 100$, across Monte-Carlo simulations of 1.45 billion samples, and that only Cross-Prior Smoothing with an informed reference reduced error across all 15 metrics tested Algorithmic Accountability in Small Data: Sample-Size-Induced Bias Within Classification Metrics. The same lesson appears on the modeling side: complex deep architectures lost to a plain PCA logistic regression on the HR-VILAGE benchmark HR-VILAGE-3K3M, and small-cohort deep learning has repeatedly overfit — six-subject EMG, eight-subject fMRI, fourteen-subject EEG Hand Motion Classification Using a Multi-Channel Surface Electromyography Sensor, Neural bases of listening to speech in noise, Design and Implementation of an EEG-Based Learning-Style Recognition Mechanism. We will report both raw and smoothed metrics side by side so the reader can see the gap rather than take our word for it.

Proposed Experiments

The first experiment is an in-vitro decay-recovery assay on the bench. All three sensor variants (adsorbed, dynamically tethered, permanently co-polymerized) run in parallel for 180 days at 37 °C in an interstitial-fluid mimic under a daily cyclic-strain schedule, with a programmed antibody concentration profile that reproduces published humoral kinetics: an exponential wane in which about half the neutralizing activity is lost by day 180, matching the H1N1 cohort Antibody Dynamics of 2009 Influenza A (H1N1) Virus in Infected Patients and Vaccinated People in China. Primary metrics are reference-channel drift in percent signal change per week, retained binding capacity at day 180, and absolute error in the recovered antibody half-life. Success for the covalent design is a reference-channel drift low enough that half-life recovery error stays within ±20% while the adsorbed variant exceeds it by a wide margin; if the dynamic tether matches the permanent anchor, the simpler chemistry wins and we say so.

The second experiment is a simulation study of identifiability and small-sample behavior. We generate synthetic cohorts at N = 6, 14, 30, and 100 — spanning the small-study sizes documented across the corpus — with informative wear-time processes and a drifting interface, then fit three models: the linear mixed-effects model ignoring wear and drift, the three-process joint model without the interface process, and our four-process model. We measure bias and interval coverage of the estimated decay constant. The expected pattern mirrors the joint-modeling paper's finding that omitting the visit process biases estimates Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data: the three-process model should remain biased by whatever drift the reference channel cannot remove in the single-channel ablation, while the four-process model approaches zero bias. This experiment also tells us the minimum cohort size at which the wearable approach yields stable estimates at all.

The third experiment is a small human pilot: thirty adults followed for six months after SARS-CoV-2 boosting, wearing the sensor continuously with two venous checkpoint draws for ELISA calibration, in the spirit of the hybrid regimen comparison that showed a four-dose schedule could match a five-dose one Antibody Persistence of Human Diploid Cell Rabies Vaccine Administrated Using the Four-Versus Five-Dose Essen Intramuscular Regimen in Post-Exposure Prophylaxis: A Prospective Cohort Study Among the Chinese Population. Outcomes are agreement between wearable-recovered and ELISA-measured titers, and responder classification reported with both raw and Cross-Prior Smoothed metrics against the HR-VILAGE reference, compared to the PCA logistic regression benchmark of F1 = 0.822 and AUC = 0.748 HR-VILAGE-3K3M. A subgroup analysis on any immunocompromised participants follows the correlate-of-risk logic validated in people with HIV Neutralizing and binding antibodies are a correlate of risk of COVID-19 in the CoVPN 3008 study in people with HIV; no reported result found on Lacuna for quantitative correlate thresholds in that study, so the subgroup analysis is exploratory and reported only with smoothed metrics and confidence intervals.

The closing decisive check is the one an editor should hold us to: in the single-channel ablation, the recovered decay curve must visibly diverge from the ELISA ground truth in a direction consistent with the measured reference-channel drift, while the dual-channel covalent design tracks it. That single comparison — same hardware, same cohort, same model, differing only in the covalent anchor and reference channel — is what would establish that the confound is real, that covalent interfacial engineering is what defuses it, and that longitudinal immune tracking at pilot scale is statistically meaningful rather than a metric artifact of small samples.

References

  1. Efficacy of the mRNA-1273 SARS-CoV-2 Vaccine at Completion of Blinded Phase
  2. Antibody Dynamics of 2009 Influenza A (H1N1) Virus in Infected Patients and Vaccinated People in China
  3. Effects of Vinyl Functionalized Silica Particles on Thermal and Mechanical Properties of Liquid Silicone Rubber Nanocomposites
  4. Printable, castable, nanocrystalline cellulose-epoxy composites exhibiting hierarchical nacre-like toughening
  5. Estimating Optimal Dynamic Treatment Regimes Using Irregularly Observed Data: A Target Trial Emulation and Bayesian Joint Modeling Approach
  6. Multifunctional hydrogel sensors with dynamic covalent networks for machine learning-assisted Parkinson's disease diagnosis and encrypted human-computer interaction
  7. Fabrication of Flexible Microneedle Array Electrodes for Wearable Bio-Signal Recording
  8. Longitudinal Analysis of Humoral and Cellular Immune Response up to 6 Months after SARS-CoV-2 BA.5/BF.7/XBB Breakthrough Infection and BA.5/BF.7-XBB Reinfection
  9. Long-term cellular immunity of vaccines for Zaire Ebola Virus Diseases
  10. Algorithmic Accountability in Small Data: Sample-Size-Induced Bias Within Classification Metrics
  11. HR-VILAGE-3K3M: A Human Respiratory Viral Immunization Longitudinal Gene Expression Dataset for Systems Immunity
  12. Hand Motion Classification Using a Multi-Channel Surface Electromyography Sensor
  13. Neural bases of listening to speech in noise
  14. Design and Implementation of an EEG-Based Learning-Style Recognition Mechanism
  15. Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design
  16. CovDocker: Benchmarking Covalent Drug Design with Tasks, Datasets, and Solutions
  17. Immune Persistence against SARS-CoV-2 after Primary and Booster Immunization in Humans: A Large-Scale Prospective Cohort Study
  18. Antibody Persistence of Human Diploid Cell Rabies Vaccine Administrated Using the Four-Versus Five-Dose Essen Intramuscular Regimen in Post-Exposure Prophylaxis: A Prospective Cohort Study Among the Chinese Population
  19. Neutralizing and binding antibodies are a correlate of risk of COVID-19 in the CoVPN 3008 study in people with HIV