Smoothing the Immunity Gap: A Turbulence-Calibrated Audit of Interpolation Bias in Exposure-Proximal Correlates

When vaccine trials ask "do antibodies protect against severe disease?", they face a quiet signal-processing problem. Antibodies are measured at a handful of visits, yet risk models need the antibody level at the exact, unobserved moment a person is exposed to the virus. Current practice fills the gap between visits by fitting a smooth parametric decay curve and reading off the interpolated value. Space physics ran into the identical trap a decade earlier: when Voyager 2's plasma data at 5 AU (24–28% missing from tracking gaps) was bridged by linear interpolation before a Fourier transform, the interpolation acted as a low-pass filter that erased high-frequency power and artificially steepened the recovered turbulence exponent, corrupting a basic physical scaling law (Spectra and correlations in the solar wind from Voyager 2 around 5 AU).

This proposal asks whether the same smoothing bias contaminates clinical risk estimates — and, more provocatively, whether part of the celebrated "humoral–cellular immunity gap" is a measurement artifact. In the ENSEMBLE trial, people with completely unquantifiable neutralizing antibodies still showed 63.1% vaccine efficacy against severe-critical COVID-19, and statistical mediation analysis attributes only 28.6% of protection to neutralizing antibodies (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial); in COVAIL the mediation fraction is 27.3% (Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost). The standard interpretation is that unmeasured cellular immunity does the rest. An alternative, untested interpretation is that interpolation smoothing attenuates the apparent predictive power of the humoral arm: smoothing a regressor toward its mean biases its fitted effect toward the null, making antibodies look weaker than they are and inflating the apparent gap. We will build the instrument to tell these apart.

Core Claim

The core claim has two parts. First, the bias is quantifiable: the parametric biphasic-decay interpolation used in exposure-proximal correlates analysis (Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine; Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost) is a low-pass filter in exactly the sense documented for Voyager data, and its effect on exposure-time titer estimates and downstream hazard ratios can be measured on ground-truth benchmarks where the answer is known. Second, the bias is correctable: the tools that fixed the heliophysics problem — estimating correlation structure directly from unevenly spaced observation pairs, without ever touching a uniform grid — can be transplanted to trial data, because vaccine cohorts offer something Voyager never had: thousands of subjects whose staggered visit times jointly cover the timeline densely even though each individual timeline is ~99% unobserved.

Weak MHD turbulence serves this proposal in two specific roles, not as subject matter. It supplies the calibration testbed — synthetic turbulence benchmarks (Synt1/Synt2) with known spectra, overlaid with the real Voyager gap pattern, on which classical pair-based estimators recover the true exponent with under 3% error (Spectra and correlations in the solar wind from Voyager 2 around 5 AU) — and it supplies the diagnostic concept of decoupling: at 5 AU the velocity and magnetic fields cascade with distinct exponents (≈ −1.5 vs. ≈ −1.75), and detecting that split required exactly the interpolation-free estimation we want to use on the two arms of immunity.

Research Program

The program has three components: calibrate the audit instrument, build the corrected estimator, and re-analyze trial-style data.

The audit instrument is a simulation harness that imposes realistic clinical sampling masks — four to six irregular visits over 220 days, the ENSEMBLE/PREVENT-19 schedule — on trajectories with known ground truth, then measures how each reconstruction method distorts three quantities: the estimated titer at a randomly chosen exposure time, the spectral content of the trajectory, and the hazard ratio a Cox model recovers when the smoothed regressor is used. The harness is first validated on the physical side: we reproduce the Fraternale result that Blackman–Tukey correlation and Rybicki–Press reconstruction recover synthetic Kolmogorov exponents to under 3% under the 1979 Voyager gap mask, and confirm that linear interpolation steepens the slope, exactly as reported. Only an instrument that reproduces the known distortion on known data is trustworthy on immune data.

The corrected estimator replaces the parametric decay prior with a spectrum-constrained, non-stationary Gaussian-process reconstruction in the Rybicki–Press tradition — minimum-variance interpolation from a covariance matrix, which is precisely a Gaussian process. The key trick is pooling. A single subject's trajectory is hopeless (a handful of points), but across thousands of participants with staggered visits, every lag is covered by many subject-pairs. Define the pooled pairwise covariance estimate

$$\hat{C}(t, t+\Delta) = \frac{1}{|P_\Delta|}\sum_{(i,t,t+\Delta)\,\in\,P_\Delta} \big(x_{i,t}-\bar{x}(t)\big)\big(x_{i,t+\Delta}-\bar{x}(t+\Delta)\big),$$

where $P_\Delta$ collects all pairs of antibody observations separated by lag $\Delta$ across all subjects $i$, and $\bar{x}(t)$ is the population mean at calendar time $t$ since vaccination. Subtracting a time-varying mean handles the dominant non-stationarity (waning), which classical Blackman–Tukey, built for stationary processes, does not. The resulting covariance then defines the prior for per-subject exposure-time inference. The dual-channel extension treats the cellular arm as the heliosheath problem: PBMC-based T-cell assays are absent for the overwhelming majority of subjects and time points in the large Phase 3 cohorts (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial), a sparsity regime (roughly 3% available data in the heliosheath) where naive interpolation fails completely and only correlation-based reconstruction survives (Solar wind spectral analysis in heliosheath from Voyager data). Where any cellular kinetic data exist — for instance the phase 1 CD8-inducing nanoparticle vaccine trial, which deliberately induces T cells without an antibody arm (A gold nanoparticle/peptide vaccine designed to induce SARS-CoV-2-specific CD8 T cells: a double-blind, randomized, phase 1 study in Switzerland; no reported result found on Lacuna for its longitudinal kinetics) — we fit the decoupling diagnostic: separate power-law exponents for humoral and cellular channels, in direct analogy to the −1.5/−1.75 split.

Proposed Experiments

Experiment 1 (calibration, physical side) runs on the Synt1/Synt2 benchmarks and the real 1979 Voyager 2 gap mask, replicating the under-3% exponent recovery for pair-based methods and quantifying, for the first time, the distortion of biphasic-style smooth parametric fits on the same data. Success is reproduction of the published exponent structure (velocity ≈ −1.5, magnetic ≈ −1.75 in the real data) and a calibrated map of distortion versus gap length.

Experiment 2 (synthetic immune ground truth) generates antibody trajectories with a biphasic backbone plus realistic deviations — a fast early decay phase, transient anamnestic excursions at exposure, and assay noise — sampled under the actual COVAIL and PREVENT-19 visit schedules. We measure absolute bias in exposure-time titer and bias in the recovered log-hazard-ratio per 10-fold titer, for every method in the baseline set. The decisive output is a curve: hazard-ratio attenuation as a function of excursion amplitude and sampling sparsity. If trajectories are genuinely smooth biphasic, parametric interpolation is unbiased and the smoothing concern evaporates — that would itself be a useful, publishable null result, and we state it as a real possibility.

Experiment 3 (real-data reanalysis) applies the pooled, spectrum-constrained estimator to trial-style longitudinal titer data (PREVENT-19's case-cohort design with 287 breakthrough Delta cases, 8 severe, is the template; Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine). Fairness comes from held-out visits: each method predicts masked observed titers, and we compare out-of-sample predictive likelihood plus the shift in the fitted per-10-fold hazard ratio. The substantive readout is whether the humoral mediation fraction (published: 28.6% in ENSEMBLE, 27.3% in COVAIL) moves materially when smoothing-induced attenuation is removed. A small shift means the humoral–cellular gap is real biology; a large shift means it is partly an artifact of the ruler.

Experiment 4 (extreme-sparse dual channel) stress-tests reconstruction when one channel sits at ~25% missingness and its partner at >90%, using the heliosheath regime as the physical twin, with the deep generative baseline included as the only method with published evidence of frequency stability at 90% missingness (LSCD: Lomb-Scargle Conditioned Diffusion for Time series Imputation). Metrics are Spectral MAE and exponent-recovery error; for the ChannelTokenFormer asynchronous-channel alternative we would compare high-frequency bias, but no reported result found on Lacuna for the SolarWind benchmark it was evaluated on (Towards Robust Real-World Multivariate Time Series Forecasting: A Unified Framework for Dependency, Asynchrony, and Missingness), so it enters as an unquantified competitor to be run under our masks.

Baselines

The do-nothing reference point is the static snapshot correlate: a single post-vaccination draw (Day 29 in ENSEMBLE) used directly in the risk model, which is the incumbent design that produced the headline finding of 63.1% efficacy against severe-critical disease at unquantifiable antibody levels and 28.6% humoral mediation (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). It is the strongest competitor for the interpretive question, because it makes no assumptions at all about what happens between visits; we run it on identical data and Cox specifications, and our method must beat its out-of-sample prediction of held-out titers to justify any added structure.

The strongest existing method is exposure-proximal correlates with biphasic exponential interpolation, as implemented in PREVENT-19 (hazard ratio 0.13 per 10-fold Day-35 Delta nAb-ID50, 95% CI 0.03–0.57) and COVAIL (27.3% mediation; HR 0.41 per 10-fold in previously infected vs. 0.74 in naïve participants, interaction p = 0.013) (Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine; Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost). It is the incumbent because it is published, regulator-facing, and interpretable. We run it with the same priors, the same optimizer budget, and the same held-out folds; beating it means strictly lower absolute bias in exposure-time titer and recovered hazard ratio on the ground-truth simulations of Experiment 2, and higher held-out predictive likelihood on real data — and if it ties on smooth trajectories, we report that honestly.

The strongest ablation isolates this proposal's key ingredient: the pooled, mean-subtracted, non-stationary covariance kernel. The ablation replaces that kernel with a stationary one or with a per-subject fit, reducing the method to the classical Blackman–Tukey / Rybicki–Press machinery that already achieves under-3% exponent error under the Voyager gap mask (Spectra and correlations in the solar wind from Voyager 2 around 5 AU). If the classical transplant alone matches the full method on immune-style masks, the extra machinery was unnecessary and the paper becomes a clean transplant story; if it fails at clinical sparsity while the pooled version succeeds, the pooling and non-stationarity treatment are the contribution.

Finally, the strongest modern baseline is LSCD, the Lomb-Scargle-conditioned diffusion imputer, which reaches MAE 0.211 versus CSDI's 0.219 at 10% missingness on the PhysioNet ICU benchmark and holds lead-frequency recovery up to 90% missingness, at the cost of roughly 45% additional training time (LSCD: Lomb-Scargle Conditioned Diffusion for Time series Imputation). It is the strongest competitor at high missingness and it is grid-based, so it can be run on our clinical masks directly. Under equal compute and tuning budget (at least 100 trials per algorithm, following the TSI-Bench protocol of 28 algorithms across 34,804 experiments, which found no single algorithm dominates all missingness patterns; Du et al.), it is beaten only if our estimator achieves lower hazard-ratio bias at clinical sparsity; on point-error metrics like its published PhysioNet MAE it may well win, which is why our primary metric is downstream risk-estimate fidelity, not point MAE.

Why not just…?

Why not just interpolate onto a uniform grid and run the FFT? Because the distortion this causes is precisely the documented result we are auditing. Linear interpolation across gaps longer than 30 minutes damps power above $10^{-3}$ Hz and over-estimates spectral slopes in Voyager data (Spectra and correlations in the solar wind from Voyager 2 around 5 AU); higher-order splines fail in the opposite direction, oscillating across wide gaps and manufacturing ghost high-frequency power (Wickramasinghe et al.); and linear interpolation of oscillatory content destroys phase information outright (Demirel et al.). Both escape routes — over-smoothing and over-fitting — distort in opposite directions, which is why the field's honest options avoid the grid entirely.

Why not just fit biphasic exponential decay curves, as immunologists already do? That is the incumbent, and for the mean trajectory it is probably adequate — a four-parameter decay fit through five points is close to saturated, and if the true trajectories are smooth, its bias is small. The evidence that something is missing is indirect but consistent: biphasic humoral curves explain only ~27–29% of protection in mediation analyses, cannot account for 63.1% efficacy at unquantifiable titers, and predict differently in naïve versus previously infected participants (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial; Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost). Our Experiments 2 and 3 are designed precisely to measure when the parametric prior is harmless and when it attenuates risk estimates; if it is harmless, the immunity gap stands as biology, and we will say so.

Why not just use Blackman–Tukey or Rybicki–Press exactly as published, with no new machinery? Largely, we do — the signal-processing core is theirs, and we claim no new estimator for stationary gapped data. What the published methods do not provide is the transplant: they assume stationarity, which waning antibody titers violate; they were validated on per-series data with 24–28% missingness, not thousands of series each ~99% unobserved; and they output a global spectrum, not the per-subject exposure-time inference and downstream hazard-ratio transmission that a correlates-of-risk analysis needs. The pooled cross-subject covariance and the non-stationary mean subtraction are the additions, and the ablation above is built to check whether those additions actually matter.

Why not just use a modern deep imputer (LSCD, CSDI, TimesNet)? This is a real threat, not a straw man: LSCD holds frequency content to 90% missingness (LSCD: Lomb-Scargle Conditioned Diffusion for Time series Imputation), and TSI-Bench's finding that no algorithm dominates every missingness pattern (Du et al.) means our regime could favor them. Three things temper the threat: the diffusion backbone remains tied to a discrete time grid, which is exactly the structure clinical sampling violates; the spectral consistency loss adds ~45% training cost for a benefit that our clinical sparsity pattern may not exercise; and TSI-Bench also documents that generative imputers can be hundreds of times slower at inference than simple baselines. Still, we run LSCD under our masks at equal compute, and if it beats our estimator on hazard-ratio bias, the audit's conclusion survives even as the recommended tool changes.

Decisive Checks

The proposal succeeds or fails on three numbers. First, on the physical calibration: our harness must reproduce the published under-3% exponent error for pair-based methods and the interpolation-induced slope steepening on Voyager-masked synthetic data, or it is not a valid instrument. Second, on synthetic immune trajectories: a measured, mask-dependent curve of hazard-ratio attenuation, with the parametric incumbent's bias crossing our estimator's at some excursion amplitude — the location of that crossing is the result, in either direction. Third, on real trial-style data: whether removing smoothing-induced attenuation shifts the humoral mediation fraction away from ~28%, and by how much. The proposal is honest about the null branch: if smoothing bias is negligible at clinical sampling densities, the humoral–cellular gap is genuine unmeasured biology, the current parametric practice is vindicated, and the audit will have shown that with an instrument calibrated on the heliosphere — a defensible use of turbulence scaling laws in either case.