When a heart has been overloaded for years, by a narrowed aortic valve or a leaking one, its main pumping chamber stretches or thickens and pumps less efficiently. If the load is removed, for example by a catheter-placed valve, the chamber can partly recover. Clinicians call this recovery reverse remodeling. They judge it almost entirely by one number, the left ventricular ejection fraction (LVEF), which is the percentage of blood the ventricle pushes out per beat, measured on echocardiography (cardiac ultrasound) before and months after treatment. A common definition labels a patient a "responder" if LVEF rises by at least 10 absolute points to above 35% (Liu et al.). That label is increasingly produced by neural networks that trace the ventricle's border in ultrasound videos. The fairness literature for medical imaging checks whether those tracings are equally accurate across sex or age. Nobody checks whether the number computed from the tracing is equally accurate across groups, or whether the yes/no recovery decision built on two such numbers is.
This note argues that the second question matters more and behaves differently. Automated LVEF error is already 5 to 7 points (Chao et al.), comparable in size to the 10-point decision threshold. Suppose that error is slightly larger or slightly biased for one group, perhaps because of body habitus, smaller ventricles, or the handheld scanners used at the bedside. Then that group's recovery labels are corrupted more often, even when segmentation overlap scores look equal. We propose to measure how fairness moves across three representations, the video, the scalar derived from it, and the threshold decision, and to test a fine-tuning objective that equalizes error on the clinical endpoint instead of on pixel overlap.
Core Claim
Our claim is that fairness does not transfer from a segmentation model's pixel accuracy to the measurement derived from it, or from that measurement to a change-based clinical decision. The reason is concrete and follows from how LVEF is computed. LVEF is a ratio of end-diastolic and end-systolic volumes, and the recovery label is a difference of two ratios measured months apart. Both operations amplify small group-specific border errors in a way that Dice, the pixel-overlap score, does not register. A model can therefore have near-equal Dice across groups and still flip recovery labels at unequal rates.
The cleanest way to see this is to assume each LVEF measurement in group $g$ carries independent error with mean $b_g$ and standard deviation $\sigma_g$. A patient whose true change is $\Delta$ points then has their responder label flipped with probability of roughly $\Phi\!\left(-|\Delta - 10|/(\sqrt{2}\,\sigma_g)\right)$ when bias is equal across timepoints. Here $\Phi$ is the standard normal cumulative distribution, and the $\sqrt{2}$ comes from differencing two independent errors. The label flip rate thus depends on each group's error variance, which no current fairness method targets, and not only on its mean bias. Bias $b_g$ cancels in the difference only if the same device and view are used at both visits. It does not cancel when the baseline is a cart-based study and follow-up is a handheld point-of-care ultrasound (POCUS) scan, a routine pattern in practice.
Relation to Prior Work
Four lines of work motivate the proposal. Chao et al. (Foundation versus Domain-Specific Models) fine-tune the Segment Anything Model (SAM) on echocardiograms. They show it holds up across hardware, with POCUS Dice of 0.857 against 0.667 for the specialized EchoNet model, and across views. They report no subgroup analysis. Xu et al.'s APPLE (Fair ultrasound diagnosis via adversarial protected attribute aware perturbations) finds the opposite side of the same coin: medical foundation models such as MedSAM segment thyroid ultrasound unequally by sex and age. APPLE also notes that biased masks distort downstream measurements such as lesion area. It fixes the problem by adversarially perturbing latent embeddings to equalize Dice, and it never evaluates the derived quantities. Our proposal takes APPLE's observation about downstream distortion as its starting point and asks whether APPLE's own remedy, Dice parity, repairs it.
Third, CareBench (When Does Multimodal Learning Help in Healthcare?) shows that feeding imaging into tabular clinical models can widen racial gaps in true positive rate, including for heart failure. Our work supplies a candidate mechanism for such gaps: image-derived features can enter a tabular model with group-dependent noise. Fourth, the reverse-remodeling prediction literature (Liu et al., XGBoost AUC 0.97 on 109 patients) and the valve-intervention cohorts (Dong et al., TAVR; Chen et al., TAVR for regurgitation) all take echo-derived LVEF or diameter as ground truth and never ask whether its error depends on who is being scanned.
Two methods constrain our design. FairDomain (Tian et al.) and TransFair (Gheisi et al.) both improve equity by injecting demographic identity into the model, the first through attention, the second by distilling fairness from a classifier into a progression predictor. APPLE instead removes it. We treat this purge-versus-inject tension as an empirical question and test both directions. FairDomain's finding that domain shift nearly tripled race disparity in retinal segmentation is the closest analogue to the cart-to-handheld shift we study. However, that study concerns paired retinal modalities and pixel-level metrics, not derived measurements.
Research Program
The first component is an audit. We will take the fine-tuned SAM and EchoNet segmentation models described by Chao et al. and run them on echocardiography cohorts that record sex, age, and body-mass index. EchoNet-Dynamic, CAMUS, and a Mayo-style cart-plus-POCUS set are the starting points, used where their metadata permit. We will also include one partner institutional cohort with paired pre- and post-intervention studies from patients treated for aortic stenosis. From each model we compute, for every patient and every group, three things: Dice; LVEF error decomposed into mean bias and standard deviation; and, for the paired cohort, the responder-label flip rate against core-lab or expert-traced reference LVEF. We stratify by sex, age band, body-mass tertile, and their intersections, and separately by device (cart versus handheld) and by whether the two visits used the same device.
The first central analysis tests whether Dice gaps predict LVEF-error gaps and whether LVEF-error gaps predict flip-rate gaps. The second tests whether the Gaussian flip-rate formula above, fed with per-group bias and variance, reproduces the observed flip rates. If it does, the formula gives a cheap audit tool for any cohort. Body habitus is included deliberately. It is the physical pathway through which acoustic imaging quality differs by group, and conditioning on it tells us whether an apparent sex or age gap is really an image-quality gap.
The second component is a method: endpoint-aware fairness fine-tuning. SAM produces end-diastolic and end-systolic masks. From those masks we compute a differentiable LVEF using the single-plane area-length approximation, which is a smooth function of mask area and long-axis length. We fine-tune with the usual segmentation loss plus a penalty on between-group differences in both the mean and the variance of LVEF error. The penalty is weighted more heavily for patients whose reference LVEF or LVEF change lies near the clinical cut-points, because that is where errors flip decisions. The group variable can be demographic, or it can be an image-quality proxy that requires no protected attributes at deployment. That choice is the purge-versus-inject test. Adding the variance term is the ingredient no existing method has. APPLE, FairAdaBN (Xu et al.), and FairDomain equalize expected performance, not dispersion of a derived quantity.
The third component closes the cross-modality loop. We feed automated LVEF and chamber measurements from each model variant into a tabular responder predictor in the style of Liu et al. We then measure whether the true positive rate gap on the recovery outcome shrinks when upstream measurement error is equalized. This is the pathway CareBench observed but could not explain.
Baselines
The naive reference point is fine-tuned SAM trained with ordinary segmentation loss, as in Chao et al. It is the strongest existing model on exactly this measurement, with POCUS Dice 0.857, CAMUS two-chamber Dice 0.891, and LVEF mean absolute error of 5 to 7% across datasets, all reported by Chao et al.. There is no reported result on Lacuna for subgroup-stratified Dice, LVEF error, or recovery-label flip rates for any echocardiography model, so establishing those numbers is itself the audit's first output. The specialized EchoNet model serves as a second unmitigated reference, because its larger hardware brittleness may interact with group composition.
The strongest existing fairness competitor is APPLE applied to the SAM echo encoder: a perturbation generator and a sex/age discriminator trained on frozen embeddings, following Xu et al.. It is the only published fairness method built for ultrasound foundation models, and it targets Dice parity, the thing we claim is insufficient. We will also run FairDomain-style demographic attention as the "inject" competitor. All methods share the same training data, backbone, number of fine-tuning steps, and a matched hyperparameter search of equal trial count, and all are selected on the same validation split by overall Dice.
Beating the baselines means two things at once. The proposed objective must reduce the largest between-group recovery-label flip rate and the between-group standard deviation of LVEF error with non-overlapping bootstrap intervals relative to APPLE. It must also do so without raising overall LVEF error by more than one point.
Two ablations isolate the key ingredient. One keeps the endpoint penalty but drops the variance term, leaving mean only. The other replaces the endpoint penalty with an equivalent Dice-parity penalty in our own training loop, so that any gain cannot be attributed to implementation differences from APPLE.
Why not just…?
Why not just recalibrate LVEF per group after the fact? This is the most serious threat. Subtracting each group's mean bias is trivial, and if bias dominates it may close most of the gap. It cannot, however, change each group's error variance, which the flip-rate formula says governs decisions near the threshold when the same device is used at both visits. We will run per-group linear recalibration, and a version that recalibrates by image-quality stratum, as explicit baselines. If recalibration matches our method on flip-rate parity, we will report that the cheap fix suffices and that the contribution is the audit and the formula, not the training objective.
Why not just apply APPLE, since it already handles ultrasound? APPLE equalizes Dice. Our hypothesis is that Dice parity does not imply LVEF-error parity, because LVEF depends on the difference between two masks' areas, and small correlated border errors cancel in Dice but not in the ratio. The Dice-parity ablation tests this directly. If APPLE also equalizes flip rates, the hypothesis is wrong, and we would say so.
Why not just train on more handheld scans, since hardware shift seems to be the problem? Chao et al. show that foundation models already close much of the hardware gap in average Dice. The open question is whether the remaining error falls unevenly by group. More data from the same population would lower average error but need not change its distribution across body habitus or sex. The device-stratified audit will show whether subgroup gaps persist after conditioning on device. If they vanish, hardware-matched data collection is the right remedy, and we would recommend it.
Why not just predict recovery directly from tabular data and skip imaging? Liu et al.'s tabular model uses chamber diameters that are themselves echo measurements, so the measurement-noise problem moves upstream rather than disappearing. CareBench also finds that heart failure is exactly the kind of condition where imaging adds information. We keep the tabular-only predictor as a comparison in the third component.
Decisive Checks
Three results would settle the claim. The first concerns the naive SAM model on the paired pre/post cohort. If recovery-label flip rates differ across sex, age, or body-mass groups by more than the Dice gaps would predict, and the Gaussian formula with per-group bias and variance reproduces those flip rates, the central thesis holds. If flip-rate gaps are negligible, the thesis fails, and the audit becomes a reassuring negative result worth publishing for the reverse-remodeling literature. The second is whether the variance term, and not merely mean-error equalization or Dice parity, is what reduces flip-rate gaps beyond per-group recalibration. The third is whether an image-quality grouping achieves the same gains as demographic grouping, which would resolve the purge-versus-inject tension in favor of physically grounded groups that do not require protected attributes at deployment.
References
- Early prediction of clinical scores for left ventricular reverse remodeling using extreme gradient random forest, boosting, and logistic regression algorithm representations
- Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound
- Fair ultrasound diagnosis via adversarial protected attribute aware perturbations on latent embeddings
- When Does Multimodal Learning Help in Healthcare? A Benchmark on EHR and Chest X-Ray Fusion
- Effectiveness and safety of transcatheter aortic valve replacement in elderly people with severe aortic stenosis with different types of heart failure
- Clinical efficacy and safety of transcatheter aortic valve replacement for patients with severe pure native aortic regurgitation
- FairDomain: Achieving Fairness in Cross-Domain Medical Image Segmentation and Classification
- TransFair: Transferring fairness from ocular disease classification to progression prediction
- FairAdaBN: Mitigating Unfairness with Adaptive Batch Normalization and Its Application to Dermatological Disease Classification