Reading Steal from Routine Angiograms: Physics-Inverted Shunt Parameters for Small Retrospective AVM Cohorts

Arteriovenous malformations (AVMs) are tangles of vessels that connect arteries directly to veins and skip the capillaries. Because blood takes the path of least resistance, they act as a drain: they pull flow away from neighbouring brain ("steal") and push arterial pressure into thin-walled veins. Clinicians believe steal causes seizures and venous overload causes bleeding. The retrospective studies that test this, however, describe the lesion with crude anatomical stand-ins such as "larger or smaller than 3 cm." These studies are small. A pediatric series of 89 patients found large size raised seizure odds roughly threefold, but it could not tell a high-flow lesion from a low-flow lesion of the same diameter (Liu et al.). Every hospital already holds the data that could make that distinction: digital subtraction angiography (DSA), an X-ray movie of injected contrast, sitting in imaging archives. In those movies, contrast washes through the malformation and into the veins at a measurable speed.

We propose to fit a small physical circuit model of the shunt to the contrast time-curves of each archived angiogram. This turns each patient's DSA into three or four interpretable numbers: the resistance of the malformation core, the fraction of inflow it diverts, and the resistance of its venous drainage. We then use those numbers, with their uncertainty, as covariates in small retrospective outcome models. The bet is about where to spend scarce samples. With fewer than a few hundred patients, a learned model cannot discover flow physics from pixels. Physics can compress the angiogram into a fixed, low-dimensional summary that is comparable across decades of changing injection protocols. The outcome model then only has to learn a handful of coefficients.

Core Claim

The claim has two parts, each falsifiable. First, parameters inverted from a lumped shunt circuit stay stable when only the injection protocol changes, while raw angiographic timing measures such as time-to-peak or arteriovenous transit time do not. This matters because retrospective archives mix injection rates, frame rates and catheter positions across eras and centers. Second, the inverted shunt fraction predicts seizure presentation better than size or Spetzler-Martin grade, and inverted venous outflow resistance predicts hemorrhagic presentation. This dissociation is what the steal and venous-hypertension mechanisms predict. Size alone cannot reproduce it.

The circuit is deliberately crude. The feeding artery splits into the malformation core with resistance $R_n$ and the adjacent normal cortical bed with resistance $R_c$. Both drain through venous segments, and the core's drainage has resistance $R_v$. In the parallel arrangement, the fraction of inflow captured by the shunt is

$$ \phi = \frac{R_c}{R_n + R_c}, $$

which is the quantity we call steal. The observed contrast curve in a region of interest, $d(t)$, is modelled as the injected bolus $b(t)$ convolved with that region's transit-time distribution $h(t;\theta)$. Here $\theta$ collects the circuit resistances and compliances, and the overall scaling and frame-rate effects are absorbed into nuisance terms:

$$ d(t) = \alpha \,(b * h(\cdot;\theta))(t) + \epsilon(t). $$

Treating injection rate, bolus shape and scaling explicitly as nuisances is what gives the expected protocol invariance. Raw time-to-peak mixes bolus shape with physiology. The inversion separates them.

Research Program

The method borrows its structure from Med-Real2Sim. That work trains on 3,840 simulations of a lumped heart circuit and then fits real echocardiograms by passing predicted parameters through the frozen simulator. We keep the "simulate, then invert non-invasive imaging" logic, but change two things: the circuit (a parallel shunt instead of a ventricle and valve) and the signal (DSA time-density curves instead of echo video). We also replace point estimates with posteriors. A simulation-based inference network is trained on a large synthetic library that varies $\theta$, the bolus parameters, frame rate, vessel overlap and noise. Its output is a posterior over $\theta$ for each patient, conditioned on curves from a feeder, the core and a draining vein.

The physics also has to survive model misspecification. A pure lumped model is not the true anatomy, so part of the synthetic library is generated from the 3D porous-medium description of the malformation core in Franzetti et al.. That work models the unresolvable nidus with Darcy's law and showed, in one patient, shunt flow falling from 71% to 1% of inflow as permeability was lowered during staged embolization. Here it serves as a richer simulator for stress-testing the lumped inversion.

The retrospective side has two components. We assemble a cohort of brain AVMs from the imaging archives and clinical records of two or more centers, with the West China pediatric series as the model design. We extract pre-treatment angiograms, presentation (seizure, hemorrhage) and treatment history. Inverted parameters enter logistic or survival models as measurement-error covariates, by drawing multiple imputations from each patient's posterior. Small samples then pay for parameter uncertainty honestly instead of hiding it.

A second aim uses the same covariates for confounding control. The finding in Liu et al. was that intervention within six months of symptom onset gave 71% seizure freedom, against 25% for intervention after six months. That comparison may be confounded if surgeons expedite high-flow lesions. With $\phi$ and $R_v$ added to a propensity model, in the style of Tewari et al., a previously unmeasured confounder becomes measured. We will report how much the timing effect moves.

As an out-of-domain check, we refit the venous-resistance parameter in sacral dural fistulae, where venous congestion is the proposed mechanism and outcome is graded on the modified Aminoff–Logue scale (Duan et al.). With 36 patients in that series, this can only be descriptive.

Baselines

The naive reference is the standard clinical model: logistic regression on age, hemorrhage history, Spetzler-Martin grade and categorical size, as fitted by Liu et al.. They report odds ratios (size OR 3.063) but no discrimination metric. There is no reported result found on Lacuna for AUC of seizure prediction in brain AVMs, so we re-fit this model on our cohort under the identical resampling scheme.

The strongest simple competitor is the same regression augmented with raw DSA timing features measured from the same curves: time-to-peak in the core and the draining vein, arteriovenous transit time, and area-under-curve ratios. This is the threat that matters most, because it uses identical data and needs no physics.

The strongest published method for small vascular cohorts with flow features is the pipeline of Zhang et al.. It extracts hemodynamic features from patient-specific CFD and compares logistic regression, random forest, XGBoost and nearest neighbours under repeated 5-fold cross-validation and 0.632+ bootstrap. Its logistic regression reached AUC above 0.75 for flow-diverter occlusion in 51 aneurysms, a different task that we cite only as the scale of discrimination such cohorts support. We reuse its resampling protocol exactly and run all four learners on each feature set.

For outcome modelling under small samples, Transfer Survival Boosting is the reference transfer method. It reweights patients from large source cohorts toward the target cohort and reached C-index 0.677 against 0.500 for target-only Cox on the MAGGIC "Euro" study. We run it for time-to-seizure-freedom when a second center can serve as the source.

The key ablation replaces the posterior over $\theta$ with the inference network's point estimate. A second ablation trains a 1D convolutional network directly on the curves, with the same synthetic pretraining but no circuit. Together these isolate whether gains come from the physics bottleneck or merely from synthetic pretraining.

All models get the same patients, folds, covariates and a fixed hyperparameter budget. Feature sets are frozen before outcome labels are unblinded. "Beating" the timing baseline means a pre-registered AUC gain whose bootstrap 95% interval excludes zero on the pooled cohort, plus no loss on the held-out center.

Why not just…?

Why not just use raw DSA timing measures in a regression? This may work as well, and the experiment is built to find out. Our argument is that raw timings are confounded by injection rate and frame rate, which vary across a retrospective archive. If the protocol-invariance test below shows raw timings to be equally stable, and they match the inverted parameters for outcome prediction, we will report that the physics adds nothing beyond normalization. The contribution would then shrink to a validated set of standardized timing features.

Why not run full 3D computational fluid dynamics per patient, as in the aneurysm work? The core's vessels are below CT and MR resolution, so no valid mesh exists. Franzetti et al. had to substitute a porous medium and calibrate its permeability by hand against DSA in a single patient. The graph surrogate of Garnier et al. is fast, but it is trained on saccular aneurysms and underestimates velocity by about 10% out of distribution, an error that would bias steal estimates in high-flow shunts. We use the 3D porous model only as a misspecification stress test.

Why not train a deep network end-to-end on angiograms and outcomes? With a few hundred patients, image networks learn scanner and protocol artifacts rather than pathology, as the chest X-ray analysis in this study found. The circuit-free 1D convolutional ablation tests the closest feasible version directly.

Why not pool a much larger registry? Rare shunts cap sample size. One high-volume center accumulated 36 sacral fistulae in eight years, with a 33.3% initial misdiagnosis rate (Duan et al.). Code-based pooling therefore adds label noise faster than signal.

Proposed Experiments and Decisive Checks

Parameter recovery comes first. On held-out synthetic curves, both lumped and porous-generated, we report posterior calibration (coverage of 90% intervals) and error on $\phi$. If the posterior on $\phi$ is too wide to separate the steal regimes of interest under realistic noise, the program stops here.

Protocol invariance is the second check. Many archived exams contain repeat injections or multiple projections in one session. We compute intraclass correlation of inverted $\phi$ and $R_v$ across same-session runs and compare it with the correlation for raw time-to-peak. The prediction is that inverted parameters are markedly more stable when injection settings differ.

Third, patients treated with staged embolization give a within-patient test: inferred $\phi$ should fall monotonically across stages. This mirrors the 71%-to-1% trajectory simulated by Franzetti et al. and needs no outcome labels.

Fourth is the mechanistic outcome test. Under the pre-registered protocol, the inverted shunt fraction should add discrimination for seizure presentation over the clinical and raw-timing baselines. Venous resistance should add discrimination for hemorrhagic presentation, and the cross-associations should be weak. Failure of this dissociation would argue that the parameters are just re-encoding size.

Finally, we report the early-versus-late intervention effect for seizure freedom with and without physiological covariates in the propensity model, using matching and overlap weighting. A material shrinkage of the 71%-versus-25% contrast would show that hemodynamic severity was confounding the "golden window."