Predicting Two-Year Magnesium Stent Fate from Six Months of Coupled Corrosion–Immunity Data

Biodegradable magnesium stents are meant to hold an artery open for a few months and then dissolve, sparing patients the permanent foreign body that a metal stent leaves behind. The catch is observational, not mechanical: preclinical trials of these alloys typically end at six months, while the alloy itself takes twelve to twenty-four months to fully resorb. The late phase is exactly when the reinforcing alloying elements — neodymium, zinc, zirconium — are released into tissue, and nobody currently observes what the immune system does about it. Medicine has learned this lesson repeatedly and painfully: short follow-up hides late failures, whether it is atrial fibrillation returning after a three-to-six-month observation window closes (Yang et al.), six-month ventricular remodeling failing to predict mortality (Sohal et al.), or vaccine durability left unevaluated after a median 5.3-month blinded phase (El Sahly et al.).

This proposal builds a computational bridge across that gap for biodegradable magnesium implants. The core object is a joint model that reads short, dense time series of two coupled processes — electrochemical degradation of the stent and the host's evolving immune response — and outputs a calibrated probability distribution over late outcomes (structural collapse, chronic inflammation, delayed hypersensitivity) at horizons the animal study never observed. The methodological wager is that degradation and immunity are not two separate readouts but a single coupled dynamical system, and that modeling their interaction is what makes six months of data informative about month twenty-four.

Core Claim

The empirical hook comes from Mao et al., whose double-extruded JDBM-2 alloy (yield strength 276 ± 6 MPa, elongation 34.3 ± 3.4%, and a micro-galvanic potential difference of roughly 25 mV between precipitates and matrix) kept structural integrity in rabbit arteries for a full six-month study, with rapid endothelialization within one month. Crucially, they also showed that the alloy's own corrosion products locally raise pH, which actively suppresses macrophage adhesion and blocks foreign-body giant cell formation relative to 316L stainless steel. Degradation is not merely tolerated by the immune response; it is temporarily shielded by it. That shield disappears when corrosion finishes. The immunology literature warns about exactly this kind of hidden slow dynamics: in the five-year PREVAC follow-up, Prague et al. and Wiedemann et al. found that standard acute assays declared T-cell memory gone after one year, while an eight-day in-vitro expansion assay (EpiMax) revealed a stable reservoir persisting to 60 months with no significant decline between 24 and 60 months — and that an early day-7 cytokine spike (IL-18, IP-10, TRAIL) carried information about the five-year outcome.

Our claim is threefold. First, early immune trajectories (cytokine panels, circulating metal-ion levels) and early degradation trajectories (strut thinning, coating integrity) jointly predict late implant fate better than either alone, because the coupling itself is informative. Second, no existing survival model can consume this signal: the strongest right-censoring tools in the literature — SurvivalPFN (Qi et al.), DCSM (Hou et al.), and VaDeSC (Marcinkevičs et al.) — all take static baseline covariates and all assume censoring is independent of the failure process. Third, in degrading-implant trials that assumption is false: animals are dropped because they are corroding or sick, which is informative censoring, and Noroozizadeh et al. showed in SurvHTE-Bench that once censoring grows severe (beyond about 70% in their sweep of 53 estimator variants over 52 datasets) naive imputation-style approaches degrade sharply.

Research Program

We will build a joint longitudinal–survival model — a model in which time-varying marker trajectories and a time-to-event outcome share latent variables, so the markers carry information about the event hazard. Each implant $i$ carries two latent states: a degradation state $d_i(t)$ (coating adhesion, remaining metal cross-section) and an immune state $m_i(t)$ (innate inflammation and latent sensitization). The event hazard is

$$h_i(t) = h_0(t)\,\exp\!\big(\gamma_d\, d_i(t) + \gamma_m\, m_i(t) + \gamma_{dm}\, d_i(t)\, m_i(t)\big),$$

where $h_0(t)$ is a baseline hazard and the coefficients $\gamma_d, \gamma_m, \gamma_{dm}$ are learned; the interaction term $\gamma_{dm}$ is the formal expression of the coupling and is precisely the ingredient our ablation will remove. Informative censoring is handled by giving the dropout process its own hazard that shares the same latent states, so an animal removed early because of rapid corrosion informs rather than biases the fit.

The dynamics of $d_i(t)$ are anchored in established materials numbers, not invented. Chen et al. give dry-bench interfacial fracture energies for polymer coatings on magnesium struts (13.6 J/m² for 5 µm PLGA on shape-optimized JDBM; 97.9 J/m² for 15 µm PLLA on AZ31), and they show compliant PLGA (1.58 GPa modulus) resists cracking far better than stiffer PLLA (2.71 GPa). Their own stated limitation is that these hold under dry conditions; we will model $G_c$ as a time-decaying quantity $G_c(t)$ under hydrolysis and enzymatic attack, calibrated against immersion data. The dynamics of $m_i(t)$ follow kinetic shapes documented in the immunology literature: acute innate spikes that resolve within weeks (Wiedemann et al.), fast-decaying humoral markers (Wang et al. found roughly half of H1N1-infected or vaccinated participants had lost protective neutralization by day 180; Li et al. documented continuous antibody decay over 12 months), and slow latent memory that only restimulation assays can see (Prague et al.).

Training follows the amortized recipe of Qi et al.: we pretrain the network offline on millions of synthetic cohorts drawn from the calibrated mechanistic simulator, including cohorts with administrative, uniform, and informative censoring patterns, so that inference on a real small cohort is a single forward pass. This is a direct reuse of their prior-data-fitting technique, extended in two ways they explicitly do not cover: time-varying longitudinal covariates instead of static ones, and dependent censoring instead of the independent-censoring assumption their method relies on.

Baselines

The naive reference point is the field's current implicit practice: declare the implant safe if it is intact at the study's end, i.e., treat six-month integrity as evidence of twenty-four-month safety. Operationally this is a Kaplan–Meier curve on the six-month data with the tail extrapolated at whatever hazard the last observed interval supports. We will run this and report its calibration at months 12 and 24 on the validation arm; any model that cannot beat it has no practical value.

The strongest existing competitor is SurvivalPFN (Qi et al.), a transformer trained on millions of synthetic survival datasets that performs Bayesian survival inference in one forward pass; on its evaluation across 81 real-world survival datasets against 21 baselines it achieved the best median rank on metrics including Brier Score and concordance index, though no single universal numeric value is reported across those datasets — we will treat rank performance, reproduced under our tuning budget, as its reference level. We will run it exactly as its authors intend, on baseline covariates available at six months (alloy type, coating parameters, final cytokine levels, imaging snapshots), with identical training data, the same compute, and the same hyperparameter search budget as our model. Beating it means better time-dependent concordance and Brier score at the 12- and 24-month horizons on held-out cohorts, with calibration intervals whose empirical coverage matches their nominal level. A second strong competitor is DCSM (Hou et al.), a mixture-of-experts deep survival model whose cluster separation on the SUPPORT cohort reached a LogRank statistic of 1067.6 versus 563.4 for Deep Survival Machines — evidence that heterogeneous risk subpopulations matter, which is exactly what we expect in implant cohorts. We will also run a standard Cox proportional-hazards regression on the same baseline covariates, and VaDeSC (Marcinkevičs et al., codebase available as VaDeSC); for VaDeSC's cluster accuracy, no reported numeric result was found on Lacuna beyond its stated superiority over DSM and SCA in accuracy and adjusted Rand index, so we will reproduce its numbers on our data ourselves.

Finally, two ablations isolate our key ingredients. One removes the coupling: the same architecture sees degradation markers only, then immune markers only, and the interaction term $\gamma_{dm}$ is set to zero in the joint version. If the coupled model does not beat all three, the central claim is wrong. The second removes the informative-censoring machinery, refitting under an independent-censoring assumption, to measure how much of the gain comes from modeling dropout honestly rather than from the longitudinal covariates.

Why not just…?

Why not just run the animal study for 24 months? This is the real gold standard, and we do not claim a model replaces it — indeed our validation design includes a 24-month truth arm precisely because extrapolation claims are worthless without ground truth. But extension is often infeasible or slow: cohorts decay, costs double, and, as the goat suture-anchor study of Pi et al. illustrates, large-animal orthopedic windows are routinely truncated before terminal resorption. The model's practical role is triage (which animals and sites justify extension) and, later, translation to human trials where two-year data cannot precede regulatory decisions. If the 24-month arm shows six-month trajectories carry no signal beyond what a Kaplan–Meier tail gives, we will report that as the answer.

Why not just use SurvivalPFN off the shelf? It is the strongest baseline and we run it as such. But it was designed for static covariates and explicitly assumes independent censoring (Qi et al.); feeding it a flattened six-month summary discards the trajectory shape that, in the immunology evidence, carries the long-horizon signal — a day-7 cytokine spike predicting five-year memory (Wiedemann et al.) is a statement about kinetics, not levels. The decisive check is head-to-head under informative censoring; if SurvivalPFN matches us there, the coupling hypothesis fails and we will say so.

Why not just fit a Cox model on baseline biomarkers? Cox regression assumes proportional hazards and homogeneous risk across the population. The benchmark evidence says both fail in this regime: DCSM's LogRank of 1067.6 versus 563.4 for a simpler deep model on SUPPORT (Hou et al.) shows the population structure a Cox model cannot see, and SurvHTE-Bench shows simple methods breaking down as censoring passes 70% (Noroozizadeh et al.). We still run Cox as a sanity baseline because in small preclinical cohorts a simple model can win by not overfitting; that risk is exactly why we borrow the amortized pretraining strategy rather than training from scratch on rabbits.

Why not just rely on standard in vitro tests — immersion corrosion and peel tests? Chen et al. themselves demonstrated the failure of this approach: their 3D simulations show delamination initiating at the luminal interface during balloon expansion, a phenomenon invisible to static or 2D testing, and they list the in vivo chemical milieu — hydrolysis, swelling, enzymatic attack — as an open problem their dry bench cannot anticipate. Static tests can still contribute: they anchor the initial conditions of our degradation simulator, which is how we use them.

Why not just use standard ex vivo immune assays at each sacrifice? Because they generate false negatives about the long run. In PREVAC, direct intracellular cytokine staining suggested T-cell memory had vanished after one year, while the EpiMax expansion assay revealed stable reservoirs to 60 months (Prague et al.). Our study design therefore includes restimulation assays against alloy degradation products (Nd, Zr) at both the 6- and 24-month endpoints, so the model is at least trained on markers that can see latent sensitization.

Proposed Experiments and Decisive Checks

The evaluation has three stages. First, a simulation benchmark: the mechanistic simulator, calibrated to the published materials numbers (276 MPa yield strength, ~25 mV galvanic differential from Mao et al.; 13.6–97.9 J/m² interface energies and 12.3% versus 4.36% high-plastic-strain fractions for standard versus shape-optimized designs from Chen et al.) and the immune kinetic shapes from the vaccinology cohorts, generates synthetic cohorts observed only to month six with censoring swept from administrative to informative, including the >70% regime that SurvHTE-Bench identifies as decisive (Noroozizadeh et al.). We also verify no regression on standard static tasks by evaluating on SurvHTE-Bench and the SurvivalPFN evaluation suite. Second, a pilot rabbit study in the spirit of Mao et al.: JDBM-2 and AZ31 stents with PLGA and PLLA coatings, serial blood draws for cytokine panels (Luminex-style, as in Wiedemann et al.) — careful: the correct id is Wiedemann et al. — plus serum magnesium and neodymium ion levels, and serial imaging of strut integrity, all through month six; a randomly selected subset continues to month twenty-four with full histology and EpiMax-style restimulation assays, providing the held-out truth arm. Third, the decisive checks, stated in advance: the coupled model must beat SurvivalPFN, DCSM, Cox, and the Kaplan–Meier extrapolation on time-dependent concordance index and Brier score at the 12- and 24-month horizons; the coupled model must beat both single-channel ablations, confirming that the interaction term earns its place; the decoupled-censoring ablation must underperform specifically in the informative-censoring regimes; and the 24-month predictions must be calibrated, with prediction intervals achieving their nominal coverage on the truth arm. Failure on any of these is a reportable negative result about how much six months of magnesium-stent data can actually say about two years.