One Collateral Signal, Many Partial Views: A Measurement Model for Stroke Triage Under Skipped Imaging

When a large brain artery is blocked by a clot, the tissue downstream survives only as long as backup vessels, called collaterals, keep blood reaching it. How good a patient's collaterals are strongly predicts whether reopening the artery will help. Guidelines treat it as a core triage variable, and good pre-treatment collateral flow on MRI roughly doubles the chance of functional independence after clot retrieval (MRI meta-analysis). The difficulty is that no single measurement captures collaterals reliably, and none is always collected. Human visual grades agree across readers as little as 16% of the time (radiomics collateral study). Catheter angiography cannot see cross-flow in 40% of carotid-terminus occlusions (flat-detector CTA validation). Perfusion scans are skipped or acquired depending on arrival time and local protocol. Each imaging study (CT angiography, CT perfusion delay maps, multiphase or flat-detector angiography, FLAIR vessel hyperintensity, arterial spin labeling, even retinal photographs) is a partial, noisy reading of the same underlying physiological quantity.

We propose to model that quantity explicitly. The model has a small latent variable, the patient's collateral state, and a learned measurement model for each modality and each human reader. It predicts outcome from whichever subset of views happened to be collected. The outcome is allowed to depend on the imaging only through that latent state and the admission clinical variables. The missingness pattern (which scans were ordered) is deliberately kept out of the outcome predictor. This design has two consequences a generic fusion network lacks. First, predictions should transfer across hospitals whose scanning protocols differ, because the model cannot learn "this hospital orders perfusion for sick patients." Second, because several views measure one quantity, their mutual disagreement can be computed at deployment without any outcome labels. That gives a self-check on shifted populations that does not depend on disease prevalence, which is the situation where confidence-based performance estimators break down (label-free estimation under shift).

Core Claim

In hyperacute stroke, the redundancy of collateral circulation is matched by redundancy in how collaterals can be measured. A fusion model built as a latent-variable measurement model with an outcome head that ignores the missingness pattern will do three things better than general missing-modality fusion. It will keep pre-treatment discrimination when protocols shift between centers. It will turn noisy, disagreeing human collateral grades into usable supervision. Its cross-view inconsistency will track real performance loss at a new site more faithfully than confidence-based label-free estimators.

Concretely, let $z$ be a low-dimensional collateral state, $c$ the admission clinical variables (age, NIHSS, onset-to-door time, glucose, occlusion site), $x_m$ the image-derived features of modality $m$, $O$ the set of modalities actually observed, $g_r$ the grade from reader $r$ when one exists, and $y$ the 90-day functional outcome. The model factorizes as

$$ p(y, x_O, g, z \mid c) = p(z \mid c)\, \prod_{m \in O} p(x_m \mid z)\, \prod_{r} p(g_r \mid z)\; p(y \mid z, c), $$

where each $p(x_m \mid z)$ is a modality-specific encoder-decoder and each $p(g_r \mid z)$ is a reader-specific confusion model in the style of classical annotator models. The set $O$ never appears in $p(y \mid z, c)$. The implied assumption is that the choice of scans depends on $c$ and workflow, not on the outcome once $z$ and $c$ are known. That assumption is stronger than what CRL-MMNAR makes, and we test it directly rather than assume it.

Relation to Prior Work

The closest methods come from general-purpose missing-modality fusion. CRL-MMNAR embeds the missingness mask as a feature because ordering behavior carries physician intent, and reports up to 13.8% gains on ICU readmission. MedMIX combines per-modality experts with a masked softmax that simply drops absent modalities. CareBench shows that fusion benefits on EHR plus chest X-ray, already small (AUPRC 0.472 to 0.495), collapse at 75% image missingness.

These papers supply our encoders and our missing-modality baselines. The proposal departs from them in two ways. None of them posits a shared physiological quantity that the modalities measure, so none can use cross-view agreement as a check. CRL-MMNAR also argues the opposite of our design choice. We think mask features are exactly what fails to transfer when a spoke hospital skips perfusion imaging for reasons unrelated to severity. The research file itself names this concern ("conflation of recording behavior with pathology"), and we make it a testable hypothesis rather than a rhetorical point.

Stroke-specific work motivates the problem but does not solve it. The CTA outcome model trains a 3D ResNet on complete cases, combines its output with clinical variables by logistic regression, and never models collaterals explicitly. The perfusion radiomics study predicts a three-class collateral grade at 72% accuracy from one modality, supervised by labels whose reliability it shows to be poor. The retinal multifractal and delay-volume studies each add one more proxy, analyzed alone with regression. Taken together, these studies are the evidence that collateral status has many imperfect readouts, and our model exploits that fact.

The mild-stroke thrombectomy cohort supplies the decision setting. For NIHSS below 6, thrombectomy showed no clear benefit (adjusted OR 0.93 for mRS 0–1). The authors note that quantitative collateral imaging was missing, which is precisely the variable our latent is meant to estimate.

Research Program

The build has three layers.

The first is a set of modality encoders. We reuse a pretrained 3D CNN backbone as in the CTA outcome study for CTA and NCCT. For CT perfusion we compute Tmax and delay-volume summaries, using the hemisphere-mirroring trick from the radiomics study to normalize against the patient's own healthy side. Circle of Willis completeness comes from the segmentations released with ISLES 2024. Each encoder feeds only into its own term $p(x_m \mid z)$. This sidesteps the problem that high-dimensional image tensors swamp low-dimensional tabular inputs under concatenation, because no image tensor reaches the outcome head directly.

The second layer is the reader model. Collateral grades from several readers are treated as extra noisy views. We will collect three independent grades per case on a subset, and use recorded clinical grades where available. This lets the model learn each reader's bias instead of averaging labels that agree 16% of the time.

The third layer is the monitor. At inference we compute a cross-view residual: how well the posterior over $z$ from each observed view agrees with the posterior from the others. We aggregate this residual per site and per month as a label-free drift signal.

Training maximizes the marginal likelihood above with amortized variational inference over $z$, with modalities randomly dropped during training. The latent is kept at two to four dimensions so that it cannot quietly become a general image embedding. We check this by probing whether $z$ predicts scanner vendor or center better than chance.

Baselines

The naive reference point is a pre-treatment clinical logistic regression on $c$ alone, with multiple imputation by chained equations (MICE) for missing tabular fields. The published points of comparison come from other cohorts and cannot be scored on our data. Traditional admission scores reach THRIVE AUC about 0.653 and HIAT about 0.658, and admission-time XGBoost reaches AUC 0.824 (dynamic thrombectomy prediction). The frequently quoted AUC 0.86 for a clinical-plus-treatment logistic regression (CTA outcome model) uses post-treatment variables such as reperfusion grade that do not exist at triage. We will report it only as a ceiling reference, not as the bar to clear. Its CTA-only arm, AUC 0.70, is the image-only reference.

The strongest existing competitor is general missing-modality fusion, run in two configurations. The first is MedMIX-style masked expert fusion. The second is the same encoders with a CRL-MMNAR-style mask embedding and cross-fitted debiasing, which is the direct counterpart to our choice to keep the mask out. Its reported degradation on MIMIC-IV-MM, AUROC 0.7119 to 0.6914 as missingness rose from 0.1 to 1.0, gives a sense of the robustness to expect. For ISLES 2024 there is no reported downstream outcome-prediction result on Lacuna, so every number on that dataset will be produced by us under one protocol.

Fairness will be enforced as follows. All models share identical encoders, identical train/test splits, the same leave-one-center-out folds, and the same hyperparameter search budget (a fixed number of trials each, selected on inner folds only).

Two ablations isolate our key ingredients. The first removes the latent bottleneck, feeding the same encoders into a direct outcome head. The second adds the mask to the outcome head of our model and leaves everything else unchanged.

Beating the competitors means one of two things. Within-center, our model matches or exceeds them on AUC and Brier score. Or, under protocol shift, our AUC drop is smaller with bootstrap confidence intervals that separate. A tie in-distribution combined with a clearly smaller shift penalty counts as a win for the central claim.

Why not just…?

Why not just record a visual collateral grade and add it to a clinical regression? This is the cheapest serious alternative and we include it as a baseline wherever grades exist. Its weaknesses are documented. Inter-reader agreement can fall to 16% (radiomics collateral study), catheter-based grading fails in 40% of carotid-terminus cases (flat-detector CTA validation), and the grade is often simply absent. Our reader-model arm tests whether modeling reader bias recovers signal that raw grades lose. If the simple grade-plus-regression matches our model under shift, the practical recommendation becomes "standardize grading," and we will say so.

Why not just use MedMIX or CRL-MMNAR off the shelf? These are the real threat for in-distribution accuracy, and they may well tie us there. Where they should fail is when a hospital changes which scans it orders. A model that reads the mask learns local ordering habits. Our protocol-shift experiment trains where perfusion imaging is skipped for early arrivals and tests where it is universal, and the mask-in ablation isolates this single difference. If mask-reading models do not degrade under such a shift, our transportability argument is wrong and the mask should be used.

Why not just train one collateral classifier per modality and average their outputs? This is a legitimate threat. A well-calibrated ensemble with averaged per-modality predictions may capture most of the accuracy gain. It lacks two things: a shared scale that lets views check each other, and a way to absorb reader noise. We therefore run it as a baseline and also compute an ensemble-disagreement monitor from it. If ensemble disagreement predicts site-level performance loss as well as our cross-view residual does, the measurement model's contribution shrinks to label handling. The proposal would then be reframed around that narrower contribution.

Proposed Experiments

The primary public testbed is ISLES 2024. It provides NCCT, CTA and CTP, occlusion masks, Circle of Willis segmentations, clinical variables and 90-day mRS for 245 cases from Munich and Zurich, which makes a genuine two-center leave-one-center-out split possible. The task is pre-treatment prediction of 90-day mRS ≤ 2. We are candid that 245 cases limits statistical power. Confidence intervals will be bootstrapped. We will seek a partner thrombectomy cohort of the size used in the CTA outcome study, 591 patients, and a subset annotated by three readers.

The first decisive check is protocol shift. We impose non-random CTP removal at the training center following an arrival-time rule, test on the other center with full imaging, and then reverse the direction. Our model should lose less AUC than the mask-reading ablation.

The second checks whether the latent means anything. On the annotated subset, $z$ should rank patients consistently with consensus collateral grades, and it should predict follow-up infarct growth on the 2–9 day MRI better than any single modality. A latent that tracks center or scanner instead would falsify the design.

The third tests the label-free monitor. Across held-out centers and simulated shifts in LVO and good-collateral prevalence, we correlate the site-level cross-view residual with the realized AUC and calibration change, and compare against confidence-based estimators. Those estimators already show error rising from roughly 0.007 in-distribution to roughly 0.05 out-of-distribution, and they fail under prevalence shift (label-free estimation). The redundancy argument predicts that cross-view agreement is insensitive to prevalence changes but sensitive to acquisition shifts, and the experiment tests exactly that separation.

The fourth, a clinical-utility check, restricts attention to mild-stroke LVO cases. There we measure whether the estimated collateral state separates patients with good outcomes from those without, within each treatment arm. This is hypothesis-generating only, since, as the mild-stroke cohort itself stresses, retrospective data cannot establish treatment benefit.

Everything here is retrospective. The honest endpoint of this proposal is a model and monitor ready for a silent prospective run, as laid out in the implementation roadmap, not a claim of clinical validation.

References

  1. MRI for collateral assessment pre-thrombectomy and association with outcome: a systematic review and meta-analysis
  2. A deep learning approach to predict collateral flow in stroke patients using radiomic features from perfusion images
  3. Validation of collateral scoring on flat-detector multiphase CT angiography in patients with acute ischemic stroke
  4. Label-free estimation of clinically relevant performance metrics under distribution shifts
  5. Causal Representation Learning from Multimodal EHRs under Non-Random Modality Missingness
  6. GRASP: Generic Framework for Health Status Representation Learning Based on Incorporating Knowledge from Similar Patients
  7. When Does Multimodal Learning Help in Healthcare? A Benchmark on EHR and Chest X-Ray Fusion
  8. Deep learning for prediction of post-thrombectomy outcomes based on admission CT angiography in large vessel occlusion stroke
  9. Retinal vessel multifractals predict pial collateral status in patients with acute ischemic stroke
  10. The combination of baseline magnetic resonance perfusion-weighted imaging-derived tissue volume with severely prolonged arterial-tissue delay and diffusion-weighted imaging lesion volume is predictive of MCA-M1 recanalization in patients treated with endovascular thrombectomy
  11. Mechanical thrombectomy for mild stroke with anterior circulation large vessel occlusion: a multicenter cohort study
  12. Dynamic Prediction of Mechanical Thrombectomy Outcome for Acute Ischemic Stroke Patients Using Machine Learning
  13. ISLES'24 -- A Real-World Longitudinal Multimodal Stroke Dataset
  14. A roadmap to implementing machine learning in healthcare: from concept to practice