Breath-Hold Failure as a Shared Risk Signal in Cardiac MRI Survival Models

When a patient cannot hold their breath during a cardiac MRI scan, two things happen at once. The images get worse: slices taken on different breath-holds no longer line up, and heart volumes computed from them become unreliable. And the scan reveals something about the patient: people who cannot hold their breath tend to be heavier, to have lung disease, and to be sicker overall. Survival models built on cardiac MRI treat only the first effect. They either throw the bad scans away or feed the corrupted measurements in as if they were clean. We propose to model both effects together. A patient-level hidden variable, a "respiratory reserve", is inferred from the per-slice misalignments in each scan. That variable then does two jobs. It sets how much the imaging measurements should be trusted, and it enters the hazard of several correlated clinical endpoints, such as heart failure, atrial fibrillation and death, as a shared risk factor.

The hidden variable is a frailty in the survival-analysis sense: an unobserved, patient-specific multiplier on risk that makes one person's several failure times move together. What is new here is that this frailty has a noisy, repeated, observable footprint inside the scan itself. A short-axis stack is acquired over roughly ten separate breath-holds, so each patient contributes a small cluster of breath-hold "attempts". Clustered repeated measurements tied to a shared latent risk is exactly the structure that shared frailty models handle. The aim is a model whose imaging hazard ratios stay stable across scan quality and whose risk estimates remain usable for the roughly one in six patients whose scans are currently set aside.

Core Claim

Breath-hold failure produces measurement error in imaging biomarkers, and the size of that error depends on patient risk. Two common fixes handle this badly. Discarding high-motion scans removes the highest-risk patients. Adding mean motion as one more covariate does not correct the weakened biomarker effect in exactly those patients.

A joint model fixes both problems if it has three ingredients:

  1. A shared patient latent estimated from per-breath-hold misalignments.
  2. A motion-dependent error model for the imaging biomarkers, calibrated on paired scans of the same people under controlled breathing.
  3. A shared frailty across multiple endpoints.

We expect this model to deliver hazard ratios that agree between low-motion and high-motion strata, better calibration in the high-motion stratum, and a measurable reduction in the unexplained correlation among a patient's endpoints.

The claim is mainly about inference and calibration, not raw discrimination. That framing is chosen deliberately. A flexible model given motion features may match our ranking accuracy, and the experiments are designed to tell us whether it does.

Where This Sits

The evidence that motivates the proposal comes from two lines of imaging work.

Tarroni et al. built an automated quality-control pipeline for UK Biobank. It measures inter-slice misalignment caused by inconsistent breath-holds. They found a median of 2.29 mm, with up to 16% of stacks above 3.4 mm, and they found that misalignment is associated with weight, body surface area and COPD. Their pipeline rejects or flags scans. It never asks what the misalignment means for outcomes.

The CMRxMotion challenge (results; protocol) scanned 40 healthy volunteers under four breathing regimes: full breath-hold, half breath-hold, free breathing and intensive breathing. It showed that left-ventricular end-diastolic volume (LVEDV) errors more than double under severe motion. It also showed that the best segmentation networks lose over 3 Dice points on poor-quality images. Its volunteers have no outcomes, so the challenge cannot say how this error reaches risk models. Its paired design does, however, provide what we need: a direct, same-subject measurement of how biomarker error grows with motion. We reuse it as the calibration set for the error model.

On the survival side, 4Dsurvival is the closest bridge between the two fields. It compresses dense 3D-plus-time motion of the heart wall from cine MRI into a latent code and trains it with a standard Cox partial likelihood. On 302 pulmonary hypertension patients it reaches C = 0.73, against 0.59 for volumetric clinical measures. It assumes well-gated breath-hold inputs, independent patients and a single endpoint. We keep its idea that imaging features feed a survival loss, and we change what the loss assumes about the data.

The method comes from the neural frailty literature. Zhou et al. add a cluster-level random effect to a neural Cox model and fit it by penalized quasi-likelihood with a Laplace approximation. In simulation with frailty variance 3.5, their model reached C = 74.67 while DeepSurv fell to 59.61. On 8,378 kidney recipients across 154 centres it achieved the best average C-index (59.24). Their inputs are tabular, and their frailty is a nuisance with no observable footprint. Neural Frailty Machine relaxes proportional hazards with individual frailty but explicitly leaves shared frailty open.

For the linear, inference-oriented version we draw on classical shared frailty estimation, namely frailtySurv and the pseudo-full likelihood approach, which has asymptotic theory. ORA is a reminder that even large time-to-event models assume different endpoints are conditionally independent. Our model drops that assumption for the endpoints we study.

One cited paper constrains the design. Almost every free-breathing validation in the file used healthy volunteers. Our error model therefore borrows a healthy-volunteer calibration and must be checked internally on patients before it is trusted. That check is built into the experiments below.

The Model

For patient $i$ with $n_i$ short-axis slices, let $m_{ij}$ be the misalignment of slice $j$ relative to its neighbours, computed with a Tarroni-style registration. Let $u_i \sim N(0,1)$ be the respiratory-reserve latent. We model

$$\log m_{ij} = \mu + \gamma u_i + c_{s(i)} + \epsilon_{ij}, \qquad \lambda_{ik}(t) = \lambda_{0k}(t)\exp\big(f_k(B_i, X_i) + \alpha_k u_i + w_i\big).$$

The symbols are as follows:

  • $c_{s(i)}$ is an imaging-centre offset.
  • $\epsilon_{ij}$ is slice-level noise.
  • $\lambda_{ik}$ is the hazard of endpoint $k$.
  • $\lambda_{0k}$ is that endpoint's baseline hazard.
  • $X_i$ are clinical covariates.
  • $f_k$ is a risk function, either neural or linear.
  • $w_i$ is residual patient frailty shared across endpoints.

The true biomarker $B_i$ (for example LVEDV, ejection fraction, or a motion latent) is not observed. We see $\hat B_i = B_i + e_i$, where the error variance $\mathrm{Var}(e_i) = g(\bar m_i)$ is a function of mean misalignment. That function $g$ is fitted on CMRxMotion by comparing each volunteer's degraded-regime biomarkers with their own full breath-hold scan.

Fitting integrates out $u_i$, $w_i$ and $B_i$ by Laplace approximation, following Zhou et al. Two quantities are scientifically interesting. One is $\alpha_k$, how much respiratory reserve predicts each endpoint. The other is the drop in $\mathrm{Var}(w_i)$ when $u_i$ is included, which measures how much of the correlation among a patient's endpoints the scan's breathing footprint explains.

Research Program

The work has three stages.

Calibrating the error model. On CMRxMotion we run fixed segmentation (nnU-Net trained on the challenge training split) and compute misalignment on every series. We then fit $g$ and its uncertainty across the four regimes, using full breath-hold as reference.

Building and checking a semi-synthetic simulator. The simulator draws $u_i$, generates slice misalignments with the UK Biobank distribution shape, perturbs true biomarkers with $g$, and generates correlated multi-endpoint survival times with known $\alpha_k$ and frailty variance. This tests whether each estimator recovers the true biomarker hazard ratio and gives honest confidence intervals.

Fitting real outcomes. We fit the model to UK Biobank imaging participants with linked outcomes. Access to that linkage is an assumption of this proposal; the file reports no survival analysis on UKBB.

A transport check guards against the healthy-volunteer calibration failing in patients. Long-axis cine views are each acquired in a single breath-hold, so they are largely immune to inter-slice misalignment. The disagreement between short-axis stack volumes and long-axis-derived volumes, plotted as a function of $\bar m_i$, gives an internal estimate of $g$ in patients. If it disagrees with the CMRxMotion estimate, we refit $g$ on UKBB rather than trust the volunteers.

Baselines

The naive reference is a Cox model on clinical covariates plus short-axis biomarkers. It uses every scan and makes no motion adjustment, fitted separately per endpoint. This is what most population imaging studies do. It defines the attenuation and miscalibration we claim to remove.

The quality-control reference applies the same model after discarding stacks with misalignment above 3.4 mm, following Tarroni et al.. It represents current practice for clean analyses. We evaluate it on the full cohort by imputing the discarded patients' risk from clinical covariates alone, so it cannot win by shrinking the test set.

The strongest competitor for prediction is a flexible neural Cox model, DeepSurv-style. It receives the same biomarkers and clinical covariates plus mean, maximum and per-slice misalignment, with free interactions. In principle it can learn "trust LVEDV less when motion is high" from data alone. It gets the same training split, the same architecture family as our $f_k$, and the same hyperparameter search budget.

The strongest correlated-survival competitor is the Zhou et al. neural frailty Cox model. We fit it with a patient-level frailty across endpoints and centre as an additional cluster, using the same inputs and budget. It captures clustering but has no observable footprint for the frailty and no error model.

Where an imaging-motion pathway is available, we also train a 4Dsurvival-style model, a denoising autoencoder with Cox loss. Its reported reference is C = 0.73 on the pulmonary hypertension cohort (Bello et al.). On UKBB, no reported result was found on Lacuna for any of these methods.

The ablations isolate the key ingredients. The first removes the error model ($g \equiv 0$). The second removes the latent from the hazard ($\alpha_k = 0$). The third replaces the per-slice measurement model with a plug-in mean misalignment.

Beating the baselines means three things. First, the high-motion and low-motion strata must yield biomarker hazard ratios whose confidence intervals overlap, where the naive model's do not. Second, calibration slope in the high-motion stratum must be closer to 1 than for the flexible neural Cox model. Third, overall C-index must be no worse than the strongest baseline. A tie on the C-index is acceptable; a loss is not.

Why not just…?

Why not just discard motion-corrupted scans with automated quality control? Because the discarded scans are not a random subset. Tarroni et al. tie misalignment to weight, body surface area and COPD, so discarding up to 16% of stacks removes patients systematically. In the simulator, where truth is known, we will measure the resulting bias in the biomarker hazard ratio and in cohort-level risk estimates. In UKBB we report calibration for the patients this strategy cannot score from imaging.

Why not just add mean misalignment as a covariate in a flexible survival network? This is the real threat. A network with motion-by-biomarker interactions can learn to downweight unreliable biomarkers, and it may match our C-index. The proposal's case against it is narrower. With about ten slices per patient, a plug-in mean is itself noisy, and a covariate cannot undo the weakening of the biomarker coefficient, so its hazard ratios remain biased even if its ranking accuracy holds. The simulator settles bias and interval coverage directly. Subsampling slices (using 4, 6 or all of them) tests whether the latent's shrinkage matters when measurements are few.

If the flexible baseline matches us on stratum-specific hazard ratios and calibration as well, we will report that the latent structure is unnecessary for prediction. The surviving contribution would then be the endpoint-correlation decomposition.

Why not just use motion-robust segmentation and then a standard survival model? CMRxMotion shows that even the leading networks lose over 3 Dice points and that LVEDV error doubles under severe motion. Robust segmentation shrinks the error but does not remove it, and it ignores the prognostic content of the breathing failure itself. We include this route as a preprocessing choice shared by every baseline, so that what remains is whatever residual error and signal it leaves behind.

Why not just fit a classical shared frailty model on clinical covariates? It captures endpoint correlation but leaves the frailty with no observable footprint, and it cannot use imaging. The linear version of our model, fitted in the frailtySurv style, serves as a test. If the latent $u_i$ explains none of the residual frailty variance there, the central scientific claim fails, regardless of how neural components perform.

Decisive Checks

Four results matter most.

The simulator must show that the joint model recovers the true biomarker hazard ratio with close to nominal 95% interval coverage. In the same setting, the naive and covariate-adjusted models must be measurably attenuated in the high-motion stratum.

On CMRxMotion, the fitted $g$ must be clearly increasing in misalignment. If biomarker error does not track per-slice misalignment, the error-model ingredient has no basis.

On UKBB, the long-axis transport check must agree with the CMRxMotion estimate of $g$ within its uncertainty, or the model must be refitted on UKBB.

Finally, adding $u_i$ must reduce the estimated shared frailty variance across heart failure, atrial fibrillation and death. It must also produce consistent hazard ratios across the 3.4 mm motion threshold, while keeping C-index within the confidence band of the best baseline.

If the stratified hazard ratios disagree only in the naive model and agree under the simple covariate baseline as well, the honest conclusion is that breath-hold failure matters as a confounder but does not need a latent model. That would still be a useful finding for anyone running population survival studies on cardiac MRI.