Medical imaging models are cleared by regulators as fixed sets of weights. Once deployed, they start to decay as scanners, hospitals and patient populations change. United States and European rules now let a manufacturer pre-register how a model may be updated, through a Predetermined Change Control Plan (PCCP), a document filed before launch that describes allowed retraining and the checks each update must pass. These plans do not say how much an update may change the model's behaviour, or for whom. The policy literature names this as the central gap. Manufacturers are left to choose their own deterioration thresholds, and an update can raise the overall score while quietly hurting a subgroup (Regulating radiology AI medical devices that evolve in their lifecycle; Regulating the Safety of Health-Related Artificial Intelligence). One review reports that about 78% of manufacturers have suspended algorithm iterations because re-certification is too costly (A decade of review in global regulation and research of artificial intelligence medical devices (2015–2025)).
We propose to make the cleared model itself the reference that every update is measured and trained against. The quantity a change plan should bound is the per-subgroup negative flip rate: the fraction of cases, within each pre-registered patient or site subgroup, that the cleared model got right and the updated model now gets wrong. The training mechanism that keeps updates inside this bound is knowledge distillation. Knowledge distillation trains one model, the student, to match the soft output probabilities of another, the teacher. Here the frozen cleared model is the teacher, and its distillation weight rises automatically for any subgroup that drifts toward its budget. The update is free to correct the cleared model's mistakes at a new site, but it pays a price for breaking the cleared model's correct calls. A regulator can read the price as a number.
Core Claim
Two claims are tested together. The first is that a per-subgroup negative-flip budget is a better unit of change control than aggregate accuracy on a held-out set. Aggregate accuracy hides exactly the trade-off the policy papers worry about: gains on the majority offsetting losses on a minority. Unlike raw output divergence, the flip budget does not penalize an update for fixing errors. The second claim is that enforcing the budget during training wins more improvement at the new site, at the same certified risk, than the obvious alternative. That alternative is to fine-tune freely and then accept or reject the result against the budget.
The training objective, run at each update, is
$$ \min_\theta \; \mathcal{L}_{\text{new}}(\theta) + \sum_{g} \lambda_g \, F_g(\theta), \qquad \lambda_g \leftarrow \max\big(0,\; \lambda_g + \eta\,(F_g(\theta) - \delta_g)\big). $$
Here $\theta$ are the deployed model's weights and $\mathcal{L}_{\text{new}}$ is the supervised loss on newly labelled site data. $g$ indexes subgroups pre-registered in the change plan, such as scanner vendor, site, age band or breast-density class. $F_g$ is a smooth surrogate for subgroup $g$'s negative flip rate: the average drop in the probability the student assigns to the cleared model's correct label, over cases where the cleared model was right. $\delta_g$ is the registered budget, and $\eta$ is a step size. The multipliers $\lambda_g$ are updated by dual ascent, a standard way of turning a constraint into an adaptive penalty. A subgroup that respects its budget exerts no pull. A subgroup that violates it gets distillation toward the cleared model, applied only where the cleared model was right.
Compliance is never certified on the data used to compute $F_g$ during training. It is checked on a disjoint, locked verification set with exact (non-smooth) flip counts and confidence intervals. Under this design, the training constraint helps an update pass, but it is not itself the evidence that the update passed.
Relation to Prior Work
The regulatory papers motivate the problem but supply no method. Regulating radiology AI medical devices that evolve in their lifecycle lists continual learning, uncertainty estimation and monitoring as building blocks of updatable devices, and says deterioration thresholds are left to manufacturers. It contains no algorithm or experiment. Rethinking clinical trials for medical AI with dynamic deployments of adaptive systems argues that freezing weights does not preserve safety. Regulating the Safety of Health-Related Artificial Intelligence frames subgroup bias as a safety hazard. We take from these papers the requirement that bounds be pre-specified and subgroup-resolved, and we supply a concrete bound and a way to train under it.
The technical papers supply the method, and two of them disagree. Addressing catastrophic forgetting for medical domain expansion expands a mammography density model across scanners and institutions. It uses Elastic Weight Consolidation (EWC), a penalty that keeps weights important to the old task close to their old values. It also forces the network to keep the source batch-normalization statistics, the running means and variances that normalize each layer's activations. It uses no distillation. Adapt Your Teacher: Improving Knowledge Distillation for Exemplar-free Continual Learning shows the opposite need on natural images: a frozen teacher with source statistics gives distorted soft labels on new-domain data, and re-estimating only the teacher's statistics on unlabeled new data repairs distillation. On 10-task CIFAR-100, this raised average incremental accuracy from 42.52% to 45.25%.
Our design has to take a side, because the teacher is the cleared model and its outputs define what "correct before" means. We treat the conflict as an experimental factor rather than assuming either answer. Rethinking Momentum Knowledge Distillation in Online Continual Learning uses a teacher that is a moving average of the student. That suits the goal of not forgetting, but it is wrong for ours: a drifting teacher cannot be the regulatory anchor, because its reference point moves with every update. We include it as a baseline to show that difference empirically. Finally, LifeLonger: A Benchmark for Continual Disease Classification and Continual Learning in Medical Imaging: A Survey and Practical Analysis constrain the design. Replay of stored patient images is the strongest continual-learning family in medicine but is often barred by privacy rules, and the field optimizes only aggregate metrics. Our method stores no images from previous sites.
Research Program
The program has four pieces. The first is the budget itself: the exact negative-flip rate per subgroup, with exact binomial intervals, computed on a locked verification set that the manufacturer holds as part of the change plan. Alongside it we report the positive flip rate, meaning cases fixed by the update, so that net benefit per subgroup is visible.
The second is the constrained distillation update described above, in two variants that differ in where $F_g$ is computed. In the development-set variant, a small labelled set that the manufacturer retains under the plan is kept separate from the verification set. In the label-free variant, the surrogate is computed on incoming unlabelled site data. Cases where the cleared model is highly confident are treated as presumed-correct. This variant needs no retained data, but it risks protecting the cleared model's confident errors. We measure that risk directly.
The third is the batch-normalization study. We cross teacher statistics (source or re-estimated on target) with student statistics (source-locked, re-estimated, or split per domain). For each combination we measure flips on the cleared domain and gains on the new one. The aim is to state when each of the two prior papers is right.
The fourth is a budget-gain curve, sweeping $\delta_g$. This tells a manufacturer what a stricter budget costs in new-site improvement. The curve is the figure a change plan would actually cite.
Proposed Experiments
The primary testbed is the scanner-to-scanner and institution-to-institution breast-density expansion in Addressing catastrophic forgetting for medical domain expansion, using the DMIST and MGH mammography cohorts at full resolution. Breast density is a four-category clinical rating scale, and the metric is Cohen's kappa, an agreement score that corrects for chance. Subgroups are pre-registered as scanner, site, density class and age band, which is the metadata these cohorts plausibly carry. We will confirm access to that metadata before committing.
The secondary testbed is the cross-domain incremental stream of LifeLonger: A Benchmark for Continual Disease Classification, built on MedMNIST v2. Here subgroups are defined by class and source dataset, because demographics are absent. This testbed gives a public, reproducible replication, but only at 28×28 pixels, a resolution its authors warn is unsuitable for clinical claims. Each update is evaluated on three things: aggregate kappa or accuracy on both domains, the per-subgroup negative and positive flip rates on the locked set, and the fraction of updates that meet every registered budget.
Baselines
The naive reference points are two. The first is the frozen cleared model, which has zero flips by definition and sets the floor for new-site performance. The second is unconstrained fine-tuning on new-site data, which Addressing catastrophic forgetting for medical domain expansion found degrades the source domain significantly (p < 0.01). In LifeLonger, fine-tuning drops BloodMNIST class-incremental accuracy to roughly 21–46%.
The strongest existing method on the primary testbed is EWC with source batch-normalization locking, from the same mammography paper. It reports kappa 0.60 on Scanner A and 0.67 on Scanner B, and 0.66 on Institution 1 and 0.62 on Institution 2. We will re-run it on identical splits, with the same backbone, number of epochs and tuning budget as our method. We will first check that we reproduce those kappas within noise, so that comparisons are like-for-like. Beating it means matching or exceeding its new-domain kappa while showing a lower worst-subgroup negative flip rate, with non-overlapping intervals. Alternatively, it means meeting a budget that EWC violates.
The strongest privacy-agnostic reference on the public testbed is iCaRL replay. On LifeLonger it holds 67.7% on BloodMNIST class-incremental learning and shows 21.7% forgetting in the cross-domain scenario, versus 58.3% for EWC. We run it as an upper reference that uses stored images, not as a fair competitor. Its per-subgroup flip rates are not reported anywhere in the file, so no reported result was found on Lacuna for negative flip rates under any method.
Two distillation baselines isolate the anchor choice. One is Learning without Forgetting-style distillation from the previous model, as evaluated in LifeLonger. The other is Momentum Knowledge Distillation, which uses a moving-average teacher. Both run at the same budget.
The key ablations remove one ingredient at a time. Uniform distillation weight across subgroups, with no dual updates, isolates the per-subgroup multipliers. Distillation on all cases rather than only cases the cleared model got right isolates the flip-targeted loss. The batch-normalization grid isolates the normalization choice.
Why not just…?
Why not just fine-tune freely and reject any update that fails the budget on the verification set? This is the real threat. If accept-or-reject gating achieves the same new-site gain at the same budget, our training method adds nothing, and the contribution reduces to the metric. We run gating with the same compute and with multiple seeds and learning rates, keeping the best passing candidate. We then compare acceptance rate and achieved kappa at each $\delta_g$. We expect constrained training to dominate at tight budgets, where almost no free fine-tune passes. If it does not, we will report the metric and gating protocol as the result.
Why not just monitor aggregate accuracy on a locked test set, as change plans already allow? Aggregate accuracy is exactly the quantity that the regulatory analyses say hides subgroup regressions (Regulating radiology AI medical devices that evolve in their lifecycle). The experiment shows this directly: we count updates that improve aggregate kappa while exceeding some subgroup's flip budget. If such updates are rare on these cohorts, the argument for subgroup budgets weakens, and we will say so.
Why not just use EWC with locked normalization statistics, which already works on this mammography task? It is a strong, cheap competitor, and it may already keep flips low. However, it constrains weights rather than behaviour. It gives no subgroup control. LifeLonger reports that EWC suffers a 58.3% drop in the cross-domain setting. Head-to-head flip rates will decide this. We also test EWC combined with our constraint, in case the two are complementary.
Why not just keep a small replay buffer of patient images? Replay is most accurate where it is allowed. Where it is barred, the development-set and label-free variants are the options. Comparing them with iCaRL quantifies what privacy costs under a fixed budget.
Decisive Checks
Three outcomes would decide the proposal. It succeeds if constrained distillation meets every registered subgroup budget on the locked mammography verification set, while matching EWC with locked normalization in new-scanner kappa (0.67) and beating budget-gated fine-tuning at tight budgets. It fails if gating matches it everywhere, or if the label-free variant measurably protects the cleared model's confident errors. The batch-normalization grid should say which of the two conflicting prior findings holds when the teacher is a regulatory anchor rather than a forgetting aid.
References
- Regulating radiology AI medical devices that evolve in their lifecycle
- Regulating the Safety of Health-Related Artificial Intelligence
- A decade of review in global regulation and research of artificial intelligence medical devices (2015–2025)
- Rethinking clinical trials for medical AI with dynamic deployments of adaptive systems
- Addressing catastrophic forgetting for medical domain expansion
- Adapt Your Teacher: Improving Knowledge Distillation for Exemplar-free Continual Learning
- Rethinking Momentum Knowledge Distillation in Online Continual Learning
- LifeLonger: A Benchmark for Continual Disease Classification
- Continual Learning in Medical Imaging: A Survey and Practical Analysis
- MedMNIST v2 - A large-scale lightweight benchmark for 2D and 3D biomedical image classification