Federating the Mechanics of Calcium Fracture: Multi-Site Prediction of Stent Expansion After Intravascular Lithotripsy

In heavily calcified arteries, a stent often fails to open fully. Intravascular lithotripsy (IVL) is a balloon catheter that fires sound pulses to crack the calcium ring before the stent goes in. Whether it works in a given patient depends on whether it produces a crack that runs through the full thickness of the calcium. Finite-element simulations show that such a "through-crack" acts like a hinge and roughly doubles lumen gain compared with a partial crack. Doctors therefore want to know, from the imaging done before treatment, which lesions will crack and expand. The patients who need this answer most, such as dialysis patients with bone-like vessels, are spread across many hospitals in small numbers. Their imaging is stored in proprietary catheter formats, and the kidney-disease history that may explain their calcification sits in separate records.

We propose a simple design and test it. Each hospital segments calcium locally at full image resolution. The hospitals then jointly train only a small shared model. It maps a handful of mechanically meaningful lesion descriptors (calcium arc, thickness, length, and a simulator-derived hinge prior) to two targets: the probability of a through-crack and the resulting stent expansion. The shared model is tiny, so it can be aggregated without a central server, with an auditable record of who contributed, at negligible cost. The central claim is falsifiable. Sharing this mechanics-shaped bottleneck should generalize to unseen hospitals at least as well as federating a full image network, and better than any single hospital's model.

Why this question, and what already exists

The motivating evidence comes from mechanics. Sogbadji et al. built 22 parametric finite-element models of stent deployment, calibrated to pre- and post-IVL optical coherence tomography (OCT, a catheter-based imaging method that produces cross-sectional images along a vessel) from 262 Disrupt CAD-III patients. They report several findings:

  • A through-crack yields up to 51% relative lumen gain in a 270° plaque, versus 16–25% for non-penetrating cracks.
  • Crack length matters far more than crack width.
  • Extensive plaques benefit more than localized ones.

They leave open whether real-world gain comes from macroscopic hinges or from diffuse micro-cracking, and the study uses no learning. Our proposal turns their mechanics into the structure of a predictor and tests the hinge hypothesis on outcome data.

The imaging-analysis tools exist but are single-center. Wang et al. classify fibrous, lipid and calcified tissue in OCT with handcrafted features and a support vector machine, trained on 35 pullbacks. Their accuracy drops to 81.15% on ex-vivo cryo-images. FiAt-Net detects fibrous caps on 24 patients. The OCTivat-Stent toolkit (Application and Evaluation of Highly Automated Software for Comprehensive Stent Analysis in Intravascular Optical Coherence Tomography) measures stent area with a concordance of 0.97 against experts across 292 pullbacks. We reuse these tools as measurement instruments. OCTivat-style analysis defines our expansion label, and Wang-style features serve as one local segmentation option. None of them addresses multiple sites or predicts the intervention outcome.

On the decentralized side, the closest work is Swarm learning for decentralized artificial intelligence in cancer histopathology. It trained image networks peer-to-peer across five colorectal cohorts. Weighted swarm learning reached a BRAF AUROC of 0.7736 on an external cohort, against 0.70–0.73 for local models and 0.7567 for the merged model. Small sites of 100–300 patients were rescued to parity with the merged model. PPBFL adds secret sharing and poisoning filters, but only on 28×28 MedMNIST images. Those images are far too coarse to show a crack edge, a limitation MedMNIST v2 itself accepts in exchange for lightweight benchmarking. TruFLaaS flags the cost of pushing large weight tensors through smart contracts.

These papers motivate our design in two ways. Decentralized training can match pooling, but it is either too coarse for fracture mechanics or too heavy for ledgers. Our design sidesteps both problems: pixels stay local, and only a model of a few thousand parameters is shared. Unlike swarm learning, we do not federate the image network at all. Whether that choice costs accuracy is the main experiment.

Fragmentation enters as a constraint rather than as decoration. Dialysis patients carry 77.8% heavily calcified carotid plaques versus 51.0% in non-dialysis patients (Okuhara et al.). Cohorts like that one (18 patients) and the fistula series in Yamaya et al. (21 patients) are too small to learn from alone. Graph analyses of large regional records link low kidney filtration (eGFR) to cardiovascular decline (Thomas et al.). Lasko warns that missing lab values often mean the test was done elsewhere, not that it was skipped. We therefore treat renal covariates as optional inputs with explicit missingness indicators. We test whether they add anything beyond geometry, instead of assuming they do.

Core Claim

Post-IVL stent expansion is governed mainly by low-dimensional calcium geometry acting through the through-crack hinge. A shared predictor built on that geometry, trained across hospitals that each keep their images and segmentation pipelines private, should achieve the following under leave-one-hospital-out evaluation:

  • It should beat every single-hospital model.
  • It should match a model trained on pooled data.
  • It should be no worse than federating an end-to-end image network.

It should also do this at a communication and audit cost small enough for a ledger. If the end-to-end federated image network clearly wins, then expansion depends on image detail the geometry misses. The claim would be wrong, and that result would itself bear on the open mechanism question.

Research Program

The system has two tiers.

Local tier. Each hospital runs its own calcium segmentation on full-resolution pullbacks. The method can be any architecture, which avoids the uniform-architecture assumption that DCHM identifies as open. From the segmentation it extracts per-lesion descriptors $x$: maximal calcium arc, maximal and mean thickness, length, and minimal pre-treatment lumen area. On paired post-IVL frames it also records whether a through-crack occurred ($c \in \{0,1\}$). After stenting, it records expansion $y$, defined as the minimal stent area divided by a reference lumen area, measured with OCTivat-style automated analysis.

Shared tier. The shared model is a two-stage predictor:

$$p(c=1 \mid x) = \sigma(f_\theta(x, s(x))), \qquad \hat y = g_\phi(x, p(c=1\mid x), r),$$

The symbols are defined as follows:

  • $s(x)$ is the lumen gain predicted by a fast surrogate fitted once to a public grid of finite-element runs that follow the Sogbadji setup.
  • $f_\theta$ and $g_\phi$ are small monotone-constrained networks or generalized additive models, with parameters $\theta$ and $\phi$.
  • $r$ is the optional vector of renal covariates with missingness flags.
  • $\sigma$ is the logistic function.

The monotone constraints encode the mechanical findings we are confident about: longer cracks and larger arcs imply larger benefit from a through-crack. The explicit middle term $p(c=1\mid x)$ is what lets us test whether hinges mediate expansion.

Aggregation. Sites exchange $\theta$ and $\phi$ peer-to-peer, using weighted averaging by sample count as in swarm learning. Each round's aggregate is committed to a permissioned ledger. Clients hold constant-size membership receipts of the kind MPFL demonstrates: 96-byte proofs verified in about 20 ms, independent of model size. With so few parameters, the full aggregate and a robust median or interquartile filter (the outlier rule from TruFLaaS) can run in the contract itself. We will measure rather than assume this. The ledger is offered as an audit property for a regulated, multi-owner consortium. We do not expect it to improve accuracy, and it is evaluated only for overhead.

Data. The primary dataset is the Disrupt CAD-III OCT corpus used by Sogbadji et al. (262 patients), partitioned by enrolling site. If site identifiers are unavailable, we instead create a deliberately skewed partition by calcium-arc strata. This is supplemented by retrospective IVL OCT from three to five consortium hospitals, including at least one with a dialysis-heavy population. Access to the trial data through the sponsor is a dependency we flag as a risk. The fallback is to use only the hospital cohorts.

Baselines

The naive reference is a single-hospital model. It has the same two-stage structure and is trained only on one site's data, then evaluated on every other site. This is what a hospital would do alone. The Swarm-learning paper shows such models losing several AUROC points externally (0.70–0.73 versus 0.7736). We expect a larger gap here because our sites are smaller.

The mechanics-only reference uses the finite-element surrogate $s(x)$ directly as the expansion prediction, with no learning from outcomes. It tests whether simulation already contains the answer.

The pooled upper bound trains the shared model on all sites' descriptors merged centrally, with identical architecture and tuning budget.

The strongest competitor is end-to-end federated image learning in the style of weighted swarm learning: a 3D convolutional network over full-resolution pre-IVL pullbacks, aggregated peer-to-peer and predicting expansion directly. It gets the same sites, the same labels, an equal hyperparameter-search budget (same number of configurations per method) and an equal GPU-hour budget per site. It is the strongest because it can, in principle, use image cues the descriptors discard, including the micro-cracking that Sogbadji et al. leave unresolved. For this specific task, there is no reported result on Lacuna for multi-site or federated prediction of post-IVL stent expansion. The nearest reported decentralized figure is the weighted swarm learning BRAF AUROC of 0.7736 versus 0.7567 merged (Saldanha et al.), in a different domain.

Our method beats this competitor if one of two things holds under leave-one-hospital-out evaluation. Either it is statistically non-inferior in expansion error and under-expansion AUROC, with a non-inferiority margin of 0.02 AUROC fixed in advance, while costing orders of magnitude less communication. Or it is strictly better at the smallest sites.

Two ablations isolate the key ingredients. The first removes the through-crack mediator, so expansion is predicted from $x$ alone. The second removes the simulator prior $s(x)$. A third variant drops the renal covariates. Finally, ledger aggregation is compared with an ordinary trusted server running identical averaging. Accuracy should be identical, so this comparison reports only latency, storage and cost.

Why not just…?

Why not just pool the data centrally, since the trial data is already pooled? For Disrupt CAD-III this is a real point, and the pooled model is in fact our upper bound. The case for federation rests on post-approval practice. There, dialysis and other high-calcium patients are scattered across hospitals and imaging vendors, which is exactly the setting where centralization meets legal barriers (Weissler et al.). If the hospital cohorts add nothing beyond the pooled trial model on the dialysis subgroup, we will report that federation is unnecessary for this question.

Why not just run standard federated averaging with a trusted coordinator? For accuracy, it is equivalent, and we say so. The ledger ablation should show identical predictions. What the ledger adds is a verifiable record of which hospital's updates entered which model version, useful for device-regulatory audit. It is cheap here only because the shared model is tiny. If overhead measurements show non-trivial latency even at a few thousand parameters, we will recommend the plain server.

Why not just fit a logistic regression on calcium arc and thickness at each large center? This is the most serious threat, because the descriptors are few and mechanically meaningful, so a high-volume center may already have enough data. The small-site rescue experiment settles it. We subsample each site to 20, 40 and 80 lesions and compare local regression with the federated model. If local regression at 80 lesions matches federation externally, the benefit is confined to small and dialysis-heavy centers, and we will scope the claim to them.

Why not just trust the finite-element simulator? The simulator is calibrated on idealized geometries and does not model micro-cracking or patient-specific wall properties. The mechanics-only baseline measures how far it falls short of observed expansion. If it is close to the learned model, the main contribution reduces to validating the simulator on outcomes, which is still worth reporting.

Proposed Experiments

The primary evaluation is leave-one-hospital-out. For each held-out site we report:

  • through-crack AUROC;
  • mean absolute error and R² for expansion;
  • under-expansion AUROC against a core-lab threshold fixed before analysis.

Confidence intervals come from patient-level bootstrap.

Three checks are decisive.

Check 1: shared model versus federated image network. We compare the shared mechanics model with the end-to-end federated image network. A clear win for the image network, beyond the non-inferiority margin, falsifies the core claim. We would then inspect saliency on its held-out errors for cues outside the calcium descriptors.

Check 2: mediation by through-cracks. We test how much of the explained variance in expansion passes through predicted through-crack probability. A large share supports the hinge account of Sogbadji et al. in real patients. A small share, with the image network winning, points toward diffuse micro-cracking.

Check 3: small-site rescue. These are the subsampling curves described above, reported separately for dialysis and non-dialysis lesions.

Secondary analyses cover two further questions. The first is whether renal covariates with missingness flags improve calibration in dialysis patients without hurting others. The second is overhead: per-round ledger latency, storage, and receipt verification time, compared against the image-network federation's per-round transfer volume. As a protocol sanity check only, the aggregation and filtering code will first be exercised on OCTMNIST from MedMNIST v2 with injected label-flipping, a setting where PPBFL reports robust accuracy. This confirms the poisoning filter behaves as expected before it touches clinical data.