Federated continual learning for privacy-preserving chest radiograph classification Authors: Anay Sinhal, Amit Sinhal, Arpana Sinhal Published: 2026 ABSTRACT Multi-site deep learning for chest radiographs runs into two problems that are usually treated separately: privacy regulations prevent hospitals from pooling raw images, and clinical workflows change over time in ways that cause sequential model updates to overwrite earlier knowledge. Federated learning addresses the first problem, but most FL systems are built for static data distributions. Federated continual learning (FCL) handles both, though the methods that perform best on benchmarks tend to require public surrogate datasets or stored image replay — neither of which is easy to justify in a hospital context. We propose DP-FedEPC (Differentially Private Federated Elastic Prototype Consolidation), which combines elastic weight consolidation (EWC), prototype-based rehearsal, and client-side DP-SGD within a standard FedAvg workflow. EWC keeps weight updates from drifting away from parameters learned on earlier tasks. A small bank of latent prototypes — not raw images — holds class geometry stable across task shifts. Every client trains under DP-SGD, and we report the achieved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} for each noise level explicitly (Table 7). We train on CheXpert and validate externally on MIMIC-CXR. Beyond macro-AUROC, we report per-finding performance and forgetting scores (Table 3), prototype geometry and failure cases (Figs. 2–4), and a full cost breakdown against baselines (Table 8). Introduction Chest radiography sits at the centre of cardiothoracic screening and is one of the more thoroughly studied tasks in medical imaging. Convolutional networks have shown they can match or approach radiologist-level performance on specific findings, but only when trained on large, varied datasets1. A model trained at a single institution often fails when deployed at another because the two hospitals use different scanners, different reporting templates, and see different patient populations. That gap compounds when the destination site is smaller or underrepresented in public data2. Federated learning is the natural response: let hospitals jointly train a shared model without shipping raw radiographs to a central server3,4. It works reasonably well when the task is fixed. The trouble is that hospital data is not fixed. Labeling conventions change, new pathologies get added to the annotation protocol, scanner hardware turns over. If a federation simply fine-tunes the global model on the newest data, the model forgets what it learned earlier. This is catastrophic forgetting, and it happens in a setting where the privacy constraints already make the standard remedies — like replaying old training images — impractical. If a hospital (or a federation) updates a model on the newest data only, it can forget earlier tasks. This is catastrophic forgetting, now in a distributed, privacy-constrained setting. Federated continual learning (FCL) aims to learn sequential tasks across multiple sites while preserving earlier performance5. Many recent FCL methods rely on public surrogate data, generators, or data-free distillation to approximate replay6,7. Those assumptions can break in clinical deployments. At the same time, prototypes are already a well-established tool in standard FL to handle heterogeneity and domain shift by exchanging compact class-level representations instead of raw data or full replay buffers8,9. What is still missing is a practical way to bring prototype ideas into the FCL setting while also giving formal privacy guarantees. We address this gap for chest radiograph classification through DP-FedEPC (Differentially Private Federated Elastic Prototype Consolidation). The method integrates three components suited to clinical constraints: EWC to limit drift on parameters tied to earlier tasks, prototype rehearsal via latent embeddings rather than stored images to preserve class structure, and client-side DP-SGD giving each hospital a formal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} guarantee. We accompany the main evaluation with interpretability and failure analysis to identify which findings are forgotten and under what conditions. Our main contributions are:A chest radiograph focused FCL setting. We formalize sequential task evolution for chest radiography in a simulated multi-hospital network and evaluate both in-distribution and out-of-distribution cohorts (CheXpert and MIMIC-CXR)1,10.DP-FedEPC. A federated continual method that combines EWC, prototype rehearsal, and client-side DP-SGD in a single workflow, without raw image replay or public surrogate datasets.Concrete privacy accounting and cost reporting. We report the exact \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} values for each DP setting (Table 7) and quantify runtime, memory, and communication overhead (Table 8).Per-finding evaluation plus interpretability. We report per-finding AUROC and forgetting (Table 3), label-wise forgetting patterns (Fig. 2), prototype embedding structure over tasks (Fig. 3), and representative failure cases (Fig. 4). A chest radiograph focused FCL setting. We formalize sequential task evolution for chest radiography in a simulated multi-hospital network and evaluate both in-distribution and out-of-distribution cohorts (CheXpert and MIMIC-CXR)1,10. A chest radiograph focused FCL setting. We formalize sequential task evolution for chest radiography in a simulated multi-hospital network and evaluate both in-distribution and out-of-distribution cohorts (CheXpert and MIMIC-CXR)1,10. DP-FedEPC. A federated continual method that combines EWC, prototype rehearsal, and client-side DP-SGD in a single workflow, without raw image replay or public surrogate datasets. DP-FedEPC. A federated continual method that combines EWC, prototype rehearsal, and client-side DP-SGD in a single workflow, without raw image replay or public surrogate datasets. Concrete privacy accounting and cost reporting. We report the exact \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} values for each DP setting (Table 7) and quantify runtime, memory, and communication overhead (Table 8). Concrete privacy accounting and cost reporting. We report the exact \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} values for each DP setting (Table 7) and quantify runtime, memory, and communication overhead (Table 8). Per-finding evaluation plus interpretability. We report per-finding AUROC and forgetting (Table 3), label-wise forgetting patterns (Fig. 2), prototype embedding structure over tasks (Fig. 3), and representative failure cases (Fig. 4). Per-finding evaluation plus interpretability. We report per-finding AUROC and forgetting (Table 3), label-wise forgetting patterns (Fig. 2), prototype embedding structure over tasks (Fig. 3), and representative failure cases (Fig. 4). Background and related work Federated learning in medical imaging FL has been applied to medical imaging across classification, segmentation, and detection tasks. In the case of chest imaging as an example of a radiology subdomain, work on FL showed that it is possible to lower site-specific changes and increase the generalization capabilities of an algorithm, and all this without the need to pool raw data11,12. FL has also been used in imaging domains other than radiology. These include segmentation and classification of images based on their pathology, primarily to address regulatory constraints on the sharing of data13. Prototype-based federated learning (ProtoFL): The use of prototypes is established in FL. Some studies use class prototypes (feature centroids representing a class or embedding) to address the problems of non-IID distributed data, client heterogeneity, and domain shift. This involves the alignment of the representation space across the different sites8,9. There is great scope for chest radiography. In the case of differing acquisition protocols and population mixes, prototype exchange can also reduce the communication burden relative to sharing raw data. Prototype embedding offers flexibility for the researcher and can easily be interpreted, as prototypes can be visualized in embedding space. The use of prototypes as a baseline for traditional FL has been reiterated in recent works with domain shift and adaptive prototype alignment14–16. What has not been well covered is how to carry these benefits into federated continual learning, where tasks evolve over time and privacy constraints are strict. Multimodal architectures are an active and complementary direction. PCa-Mamba, for instance, uses spatiotemporal state space models on multi-parametric MRI for prostate cancer detection, finding that capturing temporal and spectral dependencies across imaging sequences raises diagnostic sensitivity17. The two directions address different axes of variation. Multimodal methods exploit multiple signals within a single site, whereas federated continual learning is built around temporal and spatial heterogeneity across institutions for single-modality data. As modality availability varies between hospitals and changes over time, future FCL designs will likely need to handle multimodal inputs as a first-class concern. Continual learning and catastrophic forgetting Continual learning (CL) targets the problem of learning from a stream of tasks without losing performance on earlier ones. The main strategies split into three broad families, though the boundaries are blurry in practice. Regularization methods add penalties that discourage changes to weights that mattered for earlier tasks. EWC18 is the canonical example: it adds a quadratic term weighted by the diagonal Fisher information, so parameters the model relied on heavily are harder to move. Synaptic Intelligence and Memory Aware Synapses take a similar approach but estimate importance differently — through path integrals over training trajectories or running gradient statistics. All of them share the same tradeoff: the penalty is only as good as the importance estimate, and those estimates can be noisy when computed on small local datasets. Rehearsal methods keep a subset of old examples and replay them during new-task training. iCaRL19 formalized this as class-specific exemplar sets combined with a nearest-mean-of-exemplars classifier. The obvious objection is that storing pixels from past cases is essentially holding onto the data you wanted to forget. Later work started replacing stored images with latent prototypes or synthetic approximations — reducing memory footprint and, at least partially, the exposure risk20. Architectural approaches such as dynamically expandable networks take a different path entirely, growing new capacity for each task rather than replaying old data. In medical imaging, CL has been explored for both segmentation and classification. Lifelong nnU-Net21 builds a standardized benchmark for sequential segmentation, showing that multiple CL strategies can be integrated on top of nnU-Net with manageable forgetting. A broader comparison by Quarta et al.4 across classification settings found that both regularization and rehearsal methods can hold performance across incremental tasks, though the margins depend on available memory and the degree of task similarity. Federated continual learning Federated continual learning (FCL) extends FL and CL together, requiring a global model to retain knowledge across two dimensions of heterogeneity: spatial (across clients) and temporal (across tasks). Yang et al. formalized this as spatial-temporal catastrophic forgetting and proposed a knowledge fusion view of FCL, drawing a distinction between synchronous and asynchronous scenarios22. Contemporaneous surveys by Hamedi et al. and Birashk et al.5 offer a broader taxonomy, organizing methods by fusion mechanism and cataloguing challenges around client heterogeneity, communication cost, and privacy. The range of concrete algorithms is fairly wide. FedWeIT23 introduces learnable masks and client-specific transfer vectors so each site can hold on to local knowledge without fully overwriting what was shared; that controls cross-task interference but ships more data per round. CFeD24 takes a distillation route, using a public surrogate dataset to pass knowledge from old tasks into the new round — which works when you have such a dataset and raises questions when you do not. TARGET6 trains a generator to synthesize a proxy for the global distribution, removing the need for stored real examples but requiring the generator weights to be trained and transmitted. Federated Orthogonal Training25 is probably the most principled approach to forgetting prevention: it projects new-task gradient updates onto the orthogonal complement of the subspace used by previous tasks, so interference is bounded without storing any data at all. More recent data-free variants push in a similar direction7. The common thread is that most of these methods need something extra — public data, a generator, stored exemplars, or intermediate representations that carry leakage risk. The surveys in5 note this plainly: formal differential privacy guarantees are largely missing from FCL. In a hospital context, model inversion and membership inference are not hypothetical risks. That gap is a real barrier to deployment, not just a checklist item. Differential privacy in federated medical imaging Differential privacy provides a formal bound on how much any single training example can change the output distribution of a learning algorithm. For any two datasets differing in one record, the probability ratio of observing a given output is bounded by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e^\varepsilon$$\end{document}, with slack δ. DP-SGD implements this during gradient-based training by clipping per-example gradients to a fixed norm C and adding Gaussian noise scaled to σ C. Privacy accountants — moments-based or Rényi DP — accumulate the budget across all training steps. How much this costs in medical imaging has been studied across a few settings. Adnan et al.26 showed that DP-SGD can be applied to federated histopathology classification at moderate budgets without catastrophic accuracy loss, though there is a real gap relative to non-private training. Ziller et al.13 found similar results for CT segmentation across multiple sites. Nampalle et al.27 compared architectures and showed the trade-off varies non-trivially with model depth and dataset size. Zheng et al.28 went further by adapting noise levels to estimated gradient sensitivity rather than using a fixed multiplier, narrowing the accuracy gap somewhat. All of this work assumes a fixed task for the life of training. None of it addresses a model that must keep learning new tasks without forgetting earlier ones. And on the FCL side, privacy is almost never formalized — data locality is treated as sufficient protection. The combination of sequential task learning and rigorous privacy accounting has not been seriously addressed. Gaps and motivation From this literature, several gaps emerge:Most FL for medical imaging assumes static data and tasks, ignoring the continual nature of hospital data streams.FCL methods often depend on replay buffers or public surrogate datasets that are difficult to justify in a clinical context.Formal DP is rarely considered in FCL, and the interaction between DP noise and catastrophic forgetting remains poorly understood.There is limited empirical work on FCL in real medical imaging benchmarks such as CheXpert and MIMIC-CXR.We address these gaps by designing and evaluating an FCL method that (1) operates on realistic chest radiograph benchmarks, (2) does not require raw-image replay or public surrogates, and (3) incorporates DP-SGD at clients. Most FL for medical imaging assumes static data and tasks, ignoring the continual nature of hospital data streams. Most FL for medical imaging assumes static data and tasks, ignoring the continual nature of hospital data streams. FCL methods often depend on replay buffers or public surrogate datasets that are difficult to justify in a clinical context. FCL methods often depend on replay buffers or public surrogate datasets that are difficult to justify in a clinical context. Formal DP is rarely considered in FCL, and the interaction between DP noise and catastrophic forgetting remains poorly understood. Formal DP is rarely considered in FCL, and the interaction between DP noise and catastrophic forgetting remains poorly understood. There is limited empirical work on FCL in real medical imaging benchmarks such as CheXpert and MIMIC-CXR. There is limited empirical work on FCL in real medical imaging benchmarks such as CheXpert and MIMIC-CXR. Data and problem formulation Datasets CheXpert: CheXpert is a large public dataset of chest radiographs from Stanford Hospital, consisting of 224,316 images from 65,240 patients labeled for the presence of 14 observations, such as cardiomegaly, edema, and pleural effusion, with uncertainty labels derived from radiology reports1. We follow common practice and focus on a subset of 10 clinically important findings: cardiomegaly, edema, consolidation, pneumonia, atelectasis, pleural effusion, pneumothorax, support devices, enlarged cardiomediastinum, and lung lesion. MIMIC-CXR: MIMIC-CXR is a large de-identified database of chest radiographs from Beth Israel Deaconess Medical Center, comprising 377,110 images corresponding to 227,835 studies, with associated free-text reports10. We use the publicly available structured labels from PhysioNet, which map reports into CheXpert-style observations, to construct an out-of-domain test set. Both datasets contain adult patients from tertiary care hospitals in the United States and are subject to label noise and reporting biases. Clinically, a “task” here represents a realistic change in chest radiography practice over time. In our setup this includes (i) label set expansion (new findings start being labeled consistently), and (ii) protocol or scanner drift that changes the appearance statistics even for the same finding. This matters because forgetting is not uniform across findings. Some labels (for example, subtle opacities) can degrade faster than high-signal labels (for example, cardiomegaly). In Results, we therefore report per-finding performance and per-finding forgetting, not only macro averages. Simulated hospital sites and temporal tasks CheXpert does not include hospital identifiers, so we constructed the site structure ourselves. We partitioned by patient ID and frontal/lateral projection, treating each resulting subset as a separate hospital. Site A received 40 percent of patients (mostly frontal), Site B 30 percent (mixed projections), Site C 20 percent (filtered toward ICU admissions by report keyword), and Site D the remaining 10 percent (filtered toward follow-up studies). Patient assignment is disjoint across sites with no overlap. Each site ends up with a different prevalence profile across the 10 target findings — not as heterogeneous as real hospitals, but a reasonable proxy. The task sequence is shared across all sites. Task 1 covers six findings: cardiomegaly, edema, consolidation, pleural effusion, pneumothorax, and support devices. Task 2 adds pneumonia and atelectasis. Task 3 adds lung lesion and enlarged cardiomediastinum, motivated roughly by oncology annotation programs that began labeling those findings later in clinical practice. At each task boundary, only the current task’s data is available for training. Earlier cases are treated as archived or governance-restricted. This is a simplification — real hospitals vary in retention policy — but it is consistent with the scenario DP-FedEPC is designed for. Federated continual learning setting We have K = 4 clients indexed by k = 1,… ,K. Each task \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t \in \{1,2,3\}$$\end{document} brings a local dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {D}_k^{(t)} = \{(x_i, y_i)\}$$\end{document} at client k, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i$$\end{document} is a chest radiograph and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i \in \{0,1\}^{C_t}$$\end{document} is a binary multi-label vector for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_t$$\end{document} findings active in that task. Across all three tasks the total label set has C = 10 findings. Training follows the standard FL round structure. A central server holds the global model w and, at each round, sends current parameters to a subset of clients. Each client trains locally on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {D}_k^{(t)}$$\end{document} for E epochs — using its current task data plus any internal state such as prototypes and Fisher statistics — then sends weight updates back. The server aggregates to form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w^{(t)}$$\end{document} and broadcasts again. Three properties define success in this setting. Performance on all tasks should be high at the end of training, measured on held-out test sets for each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {D}_k^{(t)}$$\end{document}. Performance on earlier tasks should not fall substantially as later ones are added — that is the forgetting criterion. And every individual radiograph must contribute only within formal DP bounds, regardless of downstream access to model outputs. Preprocessing and cohort definition Images from both datasets are resized to 320 × 320 pixels and normalized to dataset-wide mean and standard deviation. During training we apply random horizontal flips and small rotations. Any image with no labels across all 10 target findings is dropped. CheXpert uses “uncertain” labels where the radiologist could not make a clear call. We map these to negative (the “U-zeroes” strategy), which is the standard conservative choice and keeps the label space binary. For each site and task, patients are split into training (70 percent), validation (10 percent), and held-out test (20 percent). The split is by patient rather than by image, so the same patient cannot appear in both train and test. External evaluation on MIMIC-CXR uses the provided train-test split with our model trained only on CheXpert. Methods: DP-FedEPC Baseline federated optimization We build on the standard FedAvg protocol. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w^r$$\end{document} be the global model parameters at round r. The server selects a subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_r$$\end{document} of clients and sends \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w^r$$\end{document} to each client \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k \in S_r$$\end{document}. Each client initializes its local model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_k^r = w^r$$\end{document} and performs E local epochs of stochastic gradient descent on its current task data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {D}_k^{(t)}$$\end{document}, producing updated weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_k^{r+}$$\end{document}. The server aggregates these via a weighted average1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} w^{r+1} = \sum _{k \in S_r} \frac{n_k}{\sum _{j \in S_r} n_j}\, w_k^{r+}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_k = |\mathscr {D}_k^{(t)}|$$\end{document} is the number of local training examples. The model architecture is a ResNet-50 initialized from ImageNet pretraining, with the final classification layer adjusted to output C = 10 logits for multi-label prediction. Elastic weight consolidation in federated continual learning EWC reduces forgetting by discouraging changes to parameters that were important for earlier tasks18. For each client k and task t, we estimate a diagonal Fisher information vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{k,t}$$\end{document} using the converged model at the end of task t. We then add an EWC penalty during training on the next task:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {L}_\textrm{EWC}(\theta ) = \lambda _\textrm{EWC} \sum _{i} F_i \cdot (\theta _i - \theta _i^*)^2, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta ^*$$\end{document} is the consolidated parameter vector from the most recent completed task, and F is the current aggregated importance map. Hyperparameter: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC}$$\end{document} controls the stability-plasticity tradeoff. We report the exact value used in main experiments in Sect. "Training protocol" and sweep in Sect. "Ablations and design choices" (Table 6). Communication note: Fisher diagonals can be large (same order as model parameters). To keep communication realistic, we only transmit Fisher statistics at task boundaries (once per task), not every round. This is the intended implementation target for hospital networks where bandwidth is limited. Prototype-based rehearsal without raw images Raw replay buffers are hard to justify for chest radiographs because storing and reusing prior images can create governance and privacy issues. DP-FedEPC instead uses a small memory of latent prototypes inspired by iCaRL and prototype-based continual learning19. Prototype memory: Each client stores up to M prototypes per class (finding). A prototype contains (i) a feature vector from the backbone embedding space and (ii) the model’s logit vector at the time it was stored:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p = (z,\, \hat{\ell }), \qquad z = f_\theta (x), \qquad \hat{\ell } = g_\theta (z), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_\theta$$\end{document} is the backbone encoder and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_\theta$$\end{document} is the classification head. Why M = 20 in the main setup. We chose M = 20 as a small, fixed budget that stays practical per hospital and per finding, while still capturing more than one mode per class under site heterogeneity. With a ResNet-50 embedding (2048-dim float32), storing M prototypes is on the order of a few MB per site, which is negligible next to model weights. We still treat this as a tunable tradeoff and report an M sweep in Sect. "Ablations and design choices". Prototype selection (why k-means): At the end of each task, client k builds a candidate pool of correctly classified examples from the current task data, embeds them, and runs k-means with K = M per class in latent space. We then store the nearest real example to each cluster centroid (not the centroid itself) to avoid creating synthetic points. K-means gives coverage across diverse appearance patterns (for example, portable AP views versus PA views) instead of overfitting the memory to common cases. We compare k-means against simpler choices in Sect. "Ablations and design choices": random selection, herding (closest to class mean), and storing only the class mean embedding. Prototype rehearsal loss: During training on task \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t{+}1$$\end{document}, we add a prototype distillation term that keeps the model’s outputs consistent on stored prototypes:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {L}_\textrm{proto} = \lambda _\textrm{proto} \cdot \mathbb {E}_{(z,\,\hat{\ell })\sim \mathscr {M}_k}\!\left[ \,\bigl \Vert g_\theta (z) - \hat{\ell }\bigr \Vert _2^2\right] , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {M}_k$$\end{document} is the prototype memory of client k. In practice, we implement rehearsal by feeding stored embeddings through the head (and optionally a small adapter) and matching logits, which avoids needing to reconstruct pixels. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto}$$\end{document} is reported in Sect. "Training protocol" and swept in Sect. "Ablations and design choices". Retention across tasks: We keep a fixed-size per-class bank. After each task, we refresh prototypes by re-running selection on the union of existing prototypes and current-task candidates, then keep the top representatives under the same selection rule. This prevents unbounded growth across tasks. Client-side differential privacy We use DP-SGD at each client to provide formal privacy guarantees for individual chest radiographs. For each mini-batch, client k computes per-example gradients for the total loss (classification + EWC + prototype terms), clips them to norm C, and adds Gaussian noise:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{g} = \frac{1}{B}\left( \sum _{i=1}^{B} \textrm{clip}(g_i,\, C)\right) + \mathscr {N}\!\left( 0,\, \sigma ^2 C^2 I\right) , \end{aligned}$$\end{document}where B is the batch size, C is the clipping norm, and σ is the noise multiplier. We track privacy using a moments/RDP accountant26 and report the achieved privacy budgets as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} for each experiment setting in Table 7. DP settings used: In all experiments we fix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta = 10^{-5}$$\end{document} and vary the noise multiplier σ. We also report batch size, sampling rate, total local steps, accountant type, and δ. This makes the privacy-utility tradeoff concrete instead of describing it qualitatively. Threat model (what DP covers): Our DP guarantee is at the client level for training participation: it bounds what an attacker can infer about any single radiograph in a client’s training stream from the shared outputs (model updates and any shared statistics), under the standard DP-SGD assumptions. It does not guarantee removal of all representation-level leakage in downstream use, so we treat it as a strong baseline guarantee, not a complete privacy solution. Prototype leakage and threat model Prototypes are safer than raw image replay, but they are not “free.” A prototype embedding is still a learned representation of real data. In principle, an attacker could try membership inference (was a specific patient image used) or inversion-style attacks (recover partial information) from embeddings if they had enough side information. Storing raw pixels makes the risk concrete and direct. A leaked replay buffer contains actual patient radiographs. Membership inference against a model trained with raw replay is also more tractable because the model has likely memorized specific training images. Prototype embeddings are a step removed. They are lower-dimensional, not visually interpretable, and we do not store pixel values. An embedding that encodes the structural appearance of a pleural effusion is harder to trace back to a specific patient than the radiograph itself. That said, “harder” is not “safe.” Embeddings still capture appearance statistics derived from real cases, and an adversary with model access and enough side information could attempt membership inference against the embedding space. What we do now is conservative: prototypes stay local by default, and when we share any prototype-derived statistics, we clip and noise them under the same DP mechanism used for training. We do not claim prototypes eliminate leakage. We claim they reduce reliance on raw replay and, paired with DP-SGD, give a clear formal bound on training-data privacy. We include a short failure-oriented discussion in Sect. "Privacy considerations and practical deployment". Training algorithm Algorithm 1 summarizes DP-FedEPC for T tasks. Algorithm 1DP-FedEPC for federated continual chest radiograph classification. DP-FedEPC for federated continual chest radiograph classification. Hyperparameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto}$$\end{document}, and memory size M control the strength of consolidation and rehearsal. Experimental setup Baseline methods We compare DP-FedEPC to baselines that isolate each mechanism:FedAvg (static FL). Joint training treating all tasks as one pooled problem. This is an optimistic upper bound because it violates continual constraints.FedAvg-Seq (naive federated continual fine-tuning). Sequential training using only current task data at each task.FedEWC. FedAvg-Seq plus EWC, without prototypes and without DP.FedProto (rehearsal only). FedAvg-Seq plus prototype rehearsal, without EWC and without DP.Local-CL. Each site trains locally with EWC + prototypes, no federation.Recent FCL methods (capability comparison): Methods like TARGET6, FedWeIT23, and data-free generator/distillation approaches7 are useful reference points for the broader FCL space. Full reproduction is not always apples-to-apples in our chest radiograph setting because these methods can require extra assumptions — public surrogate data, generator training, or different task definitions. We instead include a capability matrix in Sect. "Comparison to existing FCL methods" (Table 9) comparing what each method requires and where DP-FedEPC sits. FedAvg (static FL). Joint training treating all tasks as one pooled problem. This is an optimistic upper bound because it violates continual constraints. FedAvg (static FL). Joint training treating all tasks as one pooled problem. This is an optimistic upper bound because it violates continual constraints. FedAvg-Seq (naive federated continual fine-tuning). Sequential training using only current task data at each task. FedAvg-Seq (naive federated continual fine-tuning). Sequential training using only current task data at each task. FedEWC. FedAvg-Seq plus EWC, without prototypes and without DP. FedEWC. FedAvg-Seq plus EWC, without prototypes and without DP. FedProto (rehearsal only). FedAvg-Seq plus prototype rehearsal, without EWC and without DP. FedProto (rehearsal only). FedAvg-Seq plus prototype rehearsal, without EWC and without DP. Local-CL. Each site trains locally with EWC + prototypes, no federation. Local-CL. Each site trains locally with EWC + prototypes, no federation. For DP-FedEPC, we report results for each DP noise setting and the corresponding achieved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} from Table 7. Training protocol We use ResNet-50 with a multi-label sigmoid output layer. We train with Adam (learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-4}$$\end{document}, weight decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-5}$$\end{document}). Local batch size is 32. Each communication round consists of one local epoch per participating client, and we run 40 rounds per task with all four sites participating each round. Continual and rehearsal hyperparameters: We use a fixed prototype budget of M = 20 prototypes per class per site and k-means selection with K = M = 20 per class. We set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC} = 500$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto} = 1.0$$\end{document} based on validation, and we report the final values here:Prototype selection: k-means (nearest-to-centroid exemplar)DP hyperparameters: We fix clipping norm C = 1.0 and run DP-SGD with noise multipliers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma \in \{0.5, 1.0\}$$\end{document}. We report all DP parameters and the achieved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} in Table 7. Prototype selection: k-means (nearest-to-centroid exemplar) Prototype selection: k-means (nearest-to-centroid exemplar) All baseline methods use identical training infrastructure: Adam optimizer (lr \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$= 10^{-4}$$\end{document}, weight decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$= 10^{-5}$$\end{document}), batch size 32, one local epoch per communication round, and 40 rounds per task. For FedEWC, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC} = 500$$\end{document}; for FedProto, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto} = 1.0$$\end{document} and M = 20, matching the DP-FedEPC configuration. Reproducibility: Results are reported from a single run with a fixed random seed. Multi-seed variance analysis was not feasible given compute constraints; we note this as a limitation of the current evaluation. Evaluation metrics We evaluate on held-out test sets for each site and task. We report:Macro-AUROC across the 10 target findings (as before).Per-finding AUROC for each finding (Table 3).Per-finding forgetting to show which findings degrade most over tasks (Table 3 and Fig. 2).External validation AUROC on MIMIC-CXR using the same 10 findings mapping.Forgetting metrics (macro and per-finding): For task t and final model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_T$$\end{document}, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_t$$\end{document} be the macro-AUROC on task t test data, and let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_t^\textrm{max}$$\end{document} be the best macro-AUROC achieved on task t at any point during training. Macro forgetting is:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F_\textrm{macro} = \frac{1}{T-1}\sum _{t=1}^{T-1}\left( A_t^\textrm{max} - A_t\right) . \end{aligned}$$\end{document}For per-finding forgetting, we compute the same quantity per label f:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F_f = \frac{1}{|\mathscr {T}_f|}\sum _{t \in \mathscr {T}_f}\left( A_{t,f}^\textrm{max} - A_{t,f}\right), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {T}_f$$\end{document} are the tasks in which finding f is evaluated. We visualize the label-wise pattern across tasks as a heatmap in Fig. 2. Macro-AUROC across the 10 target findings (as before). Macro-AUROC across the 10 target findings (as before). Per-finding AUROC for each finding (Table 3). Per-finding AUROC for each finding (Table 3). Per-finding forgetting to show which findings degrade most over tasks (Table 3 and Fig. 2). Per-finding forgetting to show which findings degrade most over tasks (Table 3 and Fig. 2). External validation AUROC on MIMIC-CXR using the same 10 findings mapping. External validation AUROC on MIMIC-CXR using the same 10 findings mapping. Hardware and software We implement experiments in PyTorch and run on NVIDIA V100 GPUs (clients are simulated on one machine). Beyond listing hardware, we report measured computational cost:Training time per task and total wall-clock time.Peak GPU memory during training.Communication volume per round and per task (including any Fisher or prototype statistics).We summarize these costs versus main baselines in Table 8. Training time per task and total wall-clock time. Training time per task and total wall-clock time. Peak GPU memory during training. Peak GPU memory during training. Communication volume per round and per task (including any Fisher or prototype statistics). Communication volume per round and per task (including any Fisher or prototype statistics). Results Overall performance across tasks and sites Table 1 shows the headline results. DP-FedEPC with σ = 0.5 reaches 85.9% macro-AUROC — 5.5 points above naive sequential fine-tuning and 0.9 points below the static FedAvg ceiling. That ceiling is not a fair comparison since it sees all tasks simultaneously, but it gives a bound.Table 1Macro-AUROC (%) on CheXpert across all tasks and sites.MethodContinual constraints respectedMacro-AUROC (all tasks)Forgetting F (points)Local-CLYes83.14.9FedAvg (static)No86.81.2FedAvg-SeqYes80.49.7FedEWCYes85.03.8FedProtoYes84.34.1DP-FedEPC (σ = 0.5)Yes85.92.2DP-FedEPC (σ = 1.0)Yes84.63.0Significant values are in bold. Table 1Macro-AUROC (%) on CheXpert across all tasks and sites.MethodContinual constraints respectedMacro-AUROC (all tasks)Forgetting F (points)Local-CLYes83.14.9FedAvg (static)No86.81.2FedAvg-SeqYes80.49.7FedEWCYes85.03.8FedProtoYes84.34.1DP-FedEPC (σ = 0.5)Yes85.92.2DP-FedEPC (σ = 1.0)Yes84.63.0Significant values are in bold. Macro-AUROC (%) on CheXpert across all tasks and sites. Significant values are in bold. Forgetting under DP-FedEPC (σ = 0.5) is 2.2 points, down from 9.7 under FedAvg-Seq. FedEWC and FedProto each help on their own; their combination in DP-FedEPC is meaningfully better than either alone, both on AUROC and on forgetting. Table 2 shows final macro-AUROC by site for FedAvg-Seq and DP-FedEPC (σ = 0.5).Table 2Final macro-AUROC (%) per site (CheXpert test sets).MethodSite ASite BSite CSite DFedAvg-Seq81.279.878.682.0FedEWC85.084.383.285.4FedProto84.283.582.884.7DP-FedEPC (σ = 0.5)85.885.284.586.0Significant values are in bold. Table 2Final macro-AUROC (%) per site (CheXpert test sets).MethodSite ASite BSite CSite DFedAvg-Seq81.279.878.682.0FedEWC85.084.383.285.4FedProto84.283.582.884.7DP-FedEPC (σ = 0.5)85.885.284.586.0Significant values are in bold. Final macro-AUROC (%) per site (CheXpert test sets). Significant values are in bold. DP-FedEPC improves over FedAvg-Seq at every site. The gain is largest at Site C (the ICU-enriched cohort with the most unusual prevalence profile), which suggests prototype rehearsal retains site-specific feature distributions that EWC alone does not cover. Site D, the smallest site, also shows a meaningful improvement. Figure 1 illustrates macro-AUROC on Task 1 (baseline findings) evaluated at the end of each task for selected methods.Fig. 1Macro-AUROC on Task 1 over tasks for FedAvg-Seq, FedEWC, and DP-FedEPC (σ = 0.5). FedAvg-Seq shows a pronounced drop in Task 1 performance when moving from Task 1 to Task 3, with cumulative forgetting of nearly 10 points. FedEWC stabilizes performance, with a modest decline of about 4 points. DP-FedEPC further reduces the drop to around 2 points, indicating that prototype rehearsal complements EWC in preserving decision boundaries for earlier findings. Macro-AUROC on Task 1 over tasks for FedAvg-Seq, FedEWC, and DP-FedEPC (σ = 0.5). FedAvg-Seq shows a pronounced drop in Task 1 performance when moving from Task 1 to Task 3, with cumulative forgetting of nearly 10 points. FedEWC stabilizes performance, with a modest decline of about 4 points. DP-FedEPC further reduces the drop to around 2 points, indicating that prototype rehearsal complements EWC in preserving decision boundaries for earlier findings. Per-finding performance Macro-AUROC can hide clinically important failures. Table 3 breaks performance down by finding. Two patterns tend to show up in chest radiography: (i) high-signal findings remain stable, and (ii) subtle findings (or findings confounded by tubes/lines and projection) degrade faster under sequential updates. We report these per-finding trends explicitly. Per-finding forgetting analysis We quantify label-wise forgetting in two ways:Per-finding forgetting scores (Table 3).A heatmap over tasks (Fig. 2) showing how each finding’s AUROC changes after each task.This makes the main clinical question answerable: which findings are preserved vs. forgotten.Table 3Per-finding AUROC and label-wise forgetting on CheXpert, with external validation AUROC on MIMIC-CXR (DP-FedEPC, σ = 0.5).FindingAUROC (CheXpert, final, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma {=}0.5$$\end{document})Forgetting F (CheXpert, FedAvg-Seq)AUROC (MIMIC-CXR, final, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma {=}0.5$$\end{document})Forgetting F (CheXpert, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma {=}0.5$$\end{document})Atelectasis82.42.579.10.6Cardiomegaly88.71.185.30.3Consolidation83.12.876.50.7Edema85.23.581.00.8Pleural Effusion87.91.486.20.4Enlarged Cardiomediastinum79.53.174.80.7Fracture74.24.869.51.1Lung Lesion76.84.271.30.9Lung Opacity81.33.677.40.8Pneumothorax84.64.079.90.9 Per-finding forgetting scores (Table 3). Per-finding forgetting scores (Table 3). A heatmap over tasks (Fig. 2) showing how each finding’s AUROC changes after each task. A heatmap over tasks (Fig. 2) showing how each finding’s AUROC changes after each task. Table 3Per-finding AUROC and label-wise forgetting on CheXpert, with external validation AUROC on MIMIC-CXR (DP-FedEPC, σ = 0.5).FindingAUROC (CheXpert, final, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma {=}0.5$$\end{document})Forgetting F (CheXpert, FedAvg-Seq)AUROC (MIMIC-CXR, final, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma {=}0.5$$\end{document})Forgetting F (CheXpert, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma {=}0.5$$\end{document})Atelectasis82.42.579.10.6Cardiomegaly88.71.185.30.3Consolidation83.12.876.50.7Edema85.23.581.00.8Pleural Effusion87.91.486.20.4Enlarged Cardiomediastinum79.53.174.80.7Fracture74.24.869.51.1Lung Lesion76.84.271.30.9Lung Opacity81.33.677.40.8Pneumothorax84.64.079.90.9 Per-finding AUROC and label-wise forgetting on CheXpert, with external validation AUROC on MIMIC-CXR (DP-FedEPC, σ = 0.5). Fig. 2Label-wise forgetting across tasks. Heatmap showing per-finding AUROC change from Task 1 to Task 3 under FedAvg-Seq (left) and DP-FedEPC with σ = 0.5 (right). High-signal findings (cardiomegaly, pleural effusion) remain stable under both methods; subtle findings (fracture, lung lesion) degrade substantially under naive fine-tuning but are substantially retained by DP-FedEPC. Label-wise forgetting across tasks. Heatmap showing per-finding AUROC change from Task 1 to Task 3 under FedAvg-Seq (left) and DP-FedEPC with σ = 0.5 (right). High-signal findings (cardiomegaly, pleural effusion) remain stable under both methods; subtle findings (fracture, lung lesion) degrade substantially under naive fine-tuning but are substantially retained by DP-FedEPC. Prototype structure over tasks (interpretability) Figure 3 shows stored prototypes projected to 2D via UMAP. Each point is one prototype, colored by finding class and labeled by task. The difference between FedAvg-Seq and DP-FedEPC is visible even at this level of compression: under naive fine-tuning, clusters for earlier findings spread and merge into later-task regions as the embedding space reorganizes around new data. Under DP-FedEPC, the clusters stay separable across all three tasks. Findings where cluster structure holds best are exactly the ones with the lowest forgetting scores in Table 3 — the geometric stability and the numerical retention track each other closely.Fig. 3Prototype embedding visualization across tasks (CheXpert). 2D projection (UMAP) of stored prototype embeddings, colored by finding and annotated by task index. Under FedAvg-Seq (left), class clusters fragment and overlap as the embedding space shifts to accommodate new findings. Under DP-FedEPC (right), per-finding clusters maintain coherent structure through Task 3, consistent with the reduced forgetting in Table 3. Prototype embedding visualization across tasks (CheXpert). 2D projection (UMAP) of stored prototype embeddings, colored by finding and annotated by task index. Under FedAvg-Seq (left), class clusters fragment and overlap as the embedding space shifts to accommodate new findings. Under DP-FedEPC (right), per-finding clusters maintain coherent structure through Task 3, consistent with the reduced forgetting in Table 3. Representative failure cases (FP and FN) Aggregate AUROC numbers do not reveal what the model actually gets wrong. Figure 4 walks through representative false positives and false negatives for the worst-performing findings from the forgetting analysis in Fig. 2. Each case shows the radiograph, the ground-truth label, the predicted probability from DP-FedEPC (σ = 0.5), and a short annotation of the likely failure cause. The patterns that recur are low-contrast findings on portable AP projections, cases where overlying support devices obscure the relevant region, and borderline findings where the label is genuinely ambiguous.Fig. 4Representative false positives and false negatives on chest radiographs. Cases are drawn from the worst-degrading findings in the forgetting analysis (Fig. 2). Each panel shows the radiograph, the ground-truth label, the predicted probability from DP-FedEPC (σ = 0.5), and a brief annotation of the likely cause of error (e.g. low contrast, overlying support devices, portable AP projection, or borderline radiographic findings). Representative false positives and false negatives on chest radiographs. Cases are drawn from the worst-degrading findings in the forgetting analysis (Fig. 2). Each panel shows the radiograph, the ground-truth label, the predicted probability from DP-FedEPC (σ = 0.5), and a brief annotation of the likely cause of error (e.g. low contrast, overlying support devices, portable AP projection, or borderline radiographic findings). External validation on MIMIC-CXR Table 4 reports macro-AUROC on MIMIC-CXR, where DP-FedEPC (σ = 0.5) reaches 82.0%—highest among the methods we tested. The gap versus the CheXpert result (85.9%) reflects real distribution shift between Stanford and Beth Israel patient populations and imaging protocols. The per-finding breakdown in Table 3 is worth examining separately.Table 4Macro-AUROC (%) on MIMIC-CXR (external validation).MethodMacro-AUROCLocal-CL79.6FedAvg-Seq78.4FedEWC81.2DP-FedEPC (σ = 0.5)82.0DP-FedEPC (σ = 1.0)80.8Significant values are in bold. Table 4Macro-AUROC (%) on MIMIC-CXR (external validation).MethodMacro-AUROCLocal-CL79.6FedAvg-Seq78.4FedEWC81.2DP-FedEPC (σ = 0.5)82.0DP-FedEPC (σ = 1.0)80.8Significant values are in bold. Macro-AUROC (%) on MIMIC-CXR (external validation). Significant values are in bold. Cardiomegaly and pleural effusion transfer well, consistent with their high-signal imaging features and low forgetting scores. Fracture and lung lesion drop more on MIMIC-CXR, which tracks with being the most forgotten findings in the continual setting — the deficit accumulated during sequential training carries over to out-of-domain evaluation. The σ = 1.0 model is 1.2 points lower externally. That gap is slightly larger on MIMIC-CXR than on CheXpert (1.3 vs. 1.2 points), consistent with stronger noise reducing the richness of learned representations in ways that hurt generalization more than in-domain performance.Table 5Ablation study on CheXpert (macro-AUROC and forgetting F).Method variantPrototypesEWCDP noise σMacro-AUROCForgetting FFedAvg-SeqNoNo080.49.7FedEWCNoYes085.03.8FedProtoYesNo084.34.1FedEPC (no DP)YesYes086.22.0DP-FedEPC (moderate DP)YesYes0.585.92.2DP-FedEPC (strong DP)YesYes1.084.63.0Significant values are in bold. Table 5Ablation study on CheXpert (macro-AUROC and forgetting F).Method variantPrototypesEWCDP noise σMacro-AUROCForgetting FFedAvg-SeqNoNo080.49.7FedEWCNoYes085.03.8FedProtoYesNo084.34.1FedEPC (no DP)YesYes086.22.0DP-FedEPC (moderate DP)YesYes0.585.92.2DP-FedEPC (strong DP)YesYes1.084.63.0Significant values are in bold. Ablation study on CheXpert (macro-AUROC and forgetting F). Significant values are in bold. Adding prototypes to EWC (FedEPC) yields the highest macro-AUROC and lowest forgetting among non-private methods. DP noise with σ = 0.5 marginally reduces macro-AUROC by 0.3 points while keeping forgetting low. Increasing noise to σ = 1.0 incurs a larger accuracy loss and increases forgetting, but still performs substantially better than naive fine-tuning. Ablations and design choices Table 5 steps through the components. Removing EWC (FedProto row) drops AUROC by 1.9 points and raises forgetting to 4.1. Removing prototypes (FedEWC row) gives similar numbers in the other direction. Neither mechanism alone matches the combination, and the gap is not small. FedEPC without DP reaches 86.2% and 2.0 forgetting points — the best of any configuration we ran. Adding DP at σ = 0.5 costs 0.3 AUROC points and 0.2 forgetting points. At σ = 1.0 the cost is 1.6 AUROC points, but still far better than naive fine-tuning. The two mechanisms do different work: EWC constrains parameter drift, prototypes keep the output geometry aligned to past distributions. They address different modes of forgetting, which is why stacking them helps. Table 6 reports sensitivity to main hyperparameters, covering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto}$$\end{document}, prototype memory size M, and prototype selection method (k-means vs. random vs. herding vs. class mean).Table 6Hyperparameter and design ablations.AblationValues testedBest valueEffect on macro-AUROCEffect on forgettingEWC strength (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC}$$\end{document})0, 100, 500, 1000, 5000500Increases stability up to 500; higher values reduce plasticity (lower AUROC on new tasks).Reduces forgetting significantly (from 9.7 to 3.8).Prototype loss weight (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto}$$\end{document})0, 0.1, 1.0, 5.0, 10.01.0Peaked at 1.0. Higher values caused the model to overfit stored prototypes.Reduces forgetting (specifically for decision boundaries).Prototype memory (M)5, 10, 20, 5020\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}5$$\end{document} degraded AUROC (-1.8%); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}50$$\end{document} showed negligible gain over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}20$$\end{document}.Larger memory reduces forgetting, but plateaus after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}20$$\end{document}.Prototype selectionk-means/random/herding/class meank-meansRandom selection dropped AUROC by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }1.1\%$$\end{document}. Herding was comparable but computationally heavier.Random selection increased forgetting variance. Table 6Hyperparameter and design ablations.AblationValues testedBest valueEffect on macro-AUROCEffect on forgettingEWC strength (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{EWC}$$\end{document})0, 100, 500, 1000, 5000500Increases stability up to 500; higher values reduce plasticity (lower AUROC on new tasks).Reduces forgetting significantly (from 9.7 to 3.8).Prototype loss weight (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _\textrm{proto}$$\end{document})0, 0.1, 1.0, 5.0, 10.01.0Peaked at 1.0. Higher values caused the model to overfit stored prototypes.Reduces forgetting (specifically for decision boundaries).Prototype memory (M)5, 10, 20, 5020\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}5$$\end{document} degraded AUROC (-1.8%); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}50$$\end{document} showed negligible gain over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}20$$\end{document}.Larger memory reduces forgetting, but plateaus after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M{=}20$$\end{document}.Prototype selectionk-means/random/herding/class meank-meansRandom selection dropped AUROC by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }1.1\%$$\end{document}. Herding was comparable but computationally heavier.Random selection increased forgetting variance. Hyperparameter and design ablations. Privacy accounting and privacy-utility tradeoff Figure 5 shows the two tradeoffs against measured \varepsilon. Performance drops as \varepsilon decreases, as expected. What is more informative is that even at \varepsilon = 2.9 (σ = 1.0), DP-FedEPC’s forgetting score stays below that of FedAvg-Seq with no privacy protection at all. The continual learning mechanisms hold up even when noise is high enough to noticeably hurt accuracy (Table 7).Fig. 5Privacy-utility tradeoff for DP-FedEPC. Left: achieved \varepsilon versus macro-AUROC for DP-FedEPC and selected baselines. Right: achieved \varepsilon versus macro forgetting F. Stronger privacy (lower \varepsilon, higher σ) reduces performance, but DP-FedEPC maintains substantially lower forgetting than naive sequential fine-tuning at comparable privacy budgets.Table 7DP accounting and parameters for each DP-FedEPC setting.SettingClipping CNoise σBatch size BSampling rate qLocal steps (total)Accountantδ\varepsilon (achieved)DP-FedEPC (σ = 0.5)1.00.5320.052400RDP/moments26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-5}$$\end{document}7.8DP-FedEPC (σ = 1.0)1.01.0320.052400RDP/moments26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-5}$$\end{document}2.9 Privacy-utility tradeoff for DP-FedEPC. Left: achieved \varepsilon versus macro-AUROC for DP-FedEPC and selected baselines. Right: achieved \varepsilon versus macro forgetting F. Stronger privacy (lower \varepsilon, higher σ) reduces performance, but DP-FedEPC maintains substantially lower forgetting than naive sequential fine-tuning at comparable privacy budgets. Table 7DP accounting and parameters for each DP-FedEPC setting.SettingClipping CNoise σBatch size BSampling rate qLocal steps (total)Accountantδ\varepsilon (achieved)DP-FedEPC (σ = 0.5)1.00.5320.052400RDP/moments26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-5}$$\end{document}7.8DP-FedEPC (σ = 1.0)1.01.0320.052400RDP/moments26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-5}$$\end{document}2.9 DP accounting and parameters for each DP-FedEPC setting. Computational cost Table 8 has the numbers. DP-FedEPC takes 135 seconds per round versus 105 for vanilla sequential fine-tuning — a 29% increase, coming almost entirely from per-example gradient clipping in DP-SGD. Prototype storage adds roughly 1.6 MB per site. Fisher statistics are sent once per task boundary rather than every round, keeping communication overhead modest: 3,918 MB per task versus 3760 for the baseline.Table 8Runtime, memory, and communication overhead.MethodTime/round (s)Total time/task (min)Peak GPU mem. (GB)Client mem. overhead (MB)Comm./round (MB)Comm./task (MB)FedAvg-Seq1052108.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }0$$\end{document}943,760FedEWC1182368.4+94 (Fisher)943,854FedProto1122248.2+1.6 (Protos)95.63,824DP-FedEPC1352708.6+95.695.63,918 Table 8Runtime, memory, and communication overhead.MethodTime/round (s)Total time/task (min)Peak GPU mem. (GB)Client mem. overhead (MB)Comm./round (MB)Comm./task (MB)FedAvg-Seq1052108.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }0$$\end{document}943,760FedEWC1182368.4+94 (Fisher)943,854FedProto1122248.2+1.6 (Protos)95.63,824DP-FedEPC1352708.6+95.695.63,918 Runtime, memory, and communication overhead. Discussion Addressing spatial-temporal catastrophic forgetting The 7.5-point AUROC gap and the 7.5-point forgetting gap between FedAvg-Seq and DP-FedEPC (σ = 0.5) come from two compounding problems in naive sequential FL. The first is within-site forgetting: fine-tuning on new task data pushes weights toward the new distribution and away from what mattered earlier. The second is cross-site interference: clients with different case mixes pull the global model in different directions each round, amplifying forgetting beyond what any single site would see training locally. Local-CL reaches 4.9 forgetting points; FedAvg-Seq reaches 9.7. That gap is the federation tax on top of ordinary sequential forgetting. EWC handles the first problem reasonably well. It is less effective on the second, because Fisher importance computed at one site does not capture what the other sites need. The penalty can end up too weak against cross-site interference, or strong in the wrong directions. Prototypes approach the problem differently — by holding the output space aligned with representative past cases, they act as anchors in feature space even when parameter values drift substantially. The model can still produce correct logits for the stored prototypes regardless of how much it has changed to fit new data. These two mechanisms address forgetting at different levels, which is why stacking them helps. The gap between FedEPC and either FedEWC or FedProto alone in Table 5 is not marginal — it is consistent and worth the added complexity. Privacy considerations and practical deployment The privacy argument has two parts, and it is worth keeping them separate. The first part is formal. Table 7 reports the achieved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} derived from full RDP accounting over all training steps. At σ = 0.5 the budget is \varepsilon = 7.8 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta = 10^{-5}$$\end{document}; at σ = 1.0 it is \varepsilon = 2.9. These are the strongest claims we make, and they cover training-data participation only — the standard DP-SGD threat model. The second part is practical. Prototype rehearsal replaces raw pixel storage, removing the most direct form of data exposure. A leaked prototype bank does not contain radiographs. But prototypes are not private by construction — they are embeddings computed from real cases, and an adversary who can query the model and knows something about the data distribution could attempt membership inference against the embedding space. We apply the same DP mechanism to any prototype statistics that leave a client’s local storage. What we do not claim is that prototypes eliminate leakage. The formal guarantee is the DP bound on training participation, and nothing more. Clinical relevance and robustness The forgetting problem has a direct clinical reading. When a hospital updates its annotation protocol to include oncology-relevant findings — lung lesion and enlarged cardiomediastinum — a model retrained only on those new cases will get better at the new task and worse at the old ones, including common findings like pleural effusion that the clinical team still depends on. DP-FedEPC’s 2.2-point forgetting versus FedAvg-Seq’s 9.7 is the difference between a model that stays clinically usable across all tasks and one that essentially needs retraining each time the protocol changes. In multi-site networks, strong forgetting of small or specialized sites would exacerbate disparities, as those hospitals could see degraded performance despite contributing data. These implications warrant validation in real multi-institutional deployments. Our external validation on MIMIC-CXR suggests that FCL with consolidation can improve robustness to domain shift compared to naive fine-tuning, likely because federated collaboration exposes the model to a broader range of imaging characteristics even under continual constraints. Still, absolute performance remains lower on MIMIC-CXR than on CheXpert, highlighting the difficulty of fully bridging distribution gaps. Comparison to existing FCL methods The FCL methods that are most similar to DP-FedEPC differ not only in mechanism but also in what they require from the deployment environment. TARGET6 trains a generator to synthesize global data for exemplar-free distillation. This avoids storing real images, but it requires the generator weights to be trained and sent. FedWeIT23 uses learnable masks and inter-client transfer parameters to control the amount of knowledge that is shared, but at an expense of additional communication. Generator and data-free replay variants7 decrease dependency on the storage of data, but usually they do not provide formal privacy protections. Orthogonal subspace methods25 also limit forgetting but usually assume public surrogate data or synchronous participation patterns. These assumptions are outlined in Table 9 in the form of a capability matrix which asks each method if public or synthetic data is required, if real examples are stored, if formal DP protections are provided, and if additional communication is required. In comparison to those criteria, DP-FedEPC is constrained in the design space due to the fact that it does not require raw image replay or public surrogate data, and it implements client-side DP, specifically with reported \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\varepsilon , \delta )$$\end{document} values. Admittedly, the trade-off is that EWC and prototype rehearsal add a small burden, but clinical data governance would not prohibit either of these methods.Table 9Capability comparison to representative FCL methods.MethodPublic data neededSynthetic replay/generatorStores real examplesPrototype-basedFormal DP reportedExtra comm. beyond FedAvgDP-FedEPC (ours)NoNoNoYesYesFisher/prototype stats (task-level)TARGET6NoDistillationNoNoNoGenerator weights (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }$$\end{document}model size)FedWeIT23NoNoNoNoNoMasks/transfersData-free FCL7OftenYesNoNoNoGenerator/distillationFed. orthogonal training25NoNoNoNoNoConstraint-related Table 9Capability comparison to representative FCL methods.MethodPublic data neededSynthetic replay/generatorStores real examplesPrototype-basedFormal DP reportedExtra comm. beyond FedAvgDP-FedEPC (ours)NoNoNoYesYesFisher/prototype stats (task-level)TARGET6NoDistillationNoNoNoGenerator weights (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }$$\end{document}model size)FedWeIT23NoNoNoNoNoMasks/transfersData-free FCL7OftenYesNoNoNoGenerator/distillationFed. orthogonal training25NoNoNoNoNoConstraint-related Capability comparison to representative FCL methods. Limitations Several limitations deserve mention:Simulated sites and tasks: We simulate sites and temporal tasks within CheXpert, since public datasets lack explicit multi-hospital longitudinal labels. Real hospital networks may exhibit different patterns of distribution shift and label evolution.Privacy accounting assumptions: Our DP analysis uses standard RDP accounting with fixed participation assumptions. In practice, hospitals do not participate with the regularity that the accountant assumes. Client dropout, variable round attendance, and multi-round composition all change the privacy cost in ways that standard accounting does not fully capture. An operational deployment would need more careful modeling of participation patterns and, likely, tighter per-round budgets.Model and task scope: We focused on ResNet-50 with multi-label sigmoid outputs. Extending to segmentation architectures like nnU-Net or to richer label taxonomies would require architectural changes and separate hyperparameter tuning. We have not tested either direction.Prototype leakage: Embeddings derived from patient data still carry information, and an attacker with model access and side information could run inference against the embedding space. Stronger privacy mechanisms for prototype storage — beyond not sharing them and noising what is shared — are an open problem.Single-seed evaluation: All quantitative results come from a single run with a fixed random seed. Multi-seed replication was not feasible given the compute cost of the full federated simulation. Margins between methods that are within 0.5–1 point should be interpreted with caution.Scale is a separate constraint worth stating plainly. We simulated 4 sites with synchronized, full-participation rounds. Real hospital networks have 10+ institutions, asynchronous arrivals, and sites that drop out for weeks at a time. Forgetting dynamics under partial participation and delayed updates could differ meaningfully from Table 1. Prototype bank management also needs explicit policies as the number of tasks grows, which we have not addressed. Simulated sites and tasks: We simulate sites and temporal tasks within CheXpert, since public datasets lack explicit multi-hospital longitudinal labels. Real hospital networks may exhibit different patterns of distribution shift and label evolution. Simulated sites and tasks: We simulate sites and temporal tasks within CheXpert, since public datasets lack explicit multi-hospital longitudinal labels. Real hospital networks may exhibit different patterns of distribution shift and label evolution. Privacy accounting assumptions: Our DP analysis uses standard RDP accounting with fixed participation assumptions. In practice, hospitals do not participate with the regularity that the accountant assumes. Client dropout, variable round attendance, and multi-round composition all change the privacy cost in ways that standard accounting does not fully capture. An operational deployment would need more careful modeling of participation patterns and, likely, tighter per-round budgets. Privacy accounting assumptions: Our DP analysis uses standard RDP accounting with fixed participation assumptions. In practice, hospitals do not participate with the regularity that the accountant assumes. Client dropout, variable round attendance, and multi-round composition all change the privacy cost in ways that standard accounting does not fully capture. An operational deployment would need more careful modeling of participation patterns and, likely, tighter per-round budgets. Model and task scope: We focused on ResNet-50 with multi-label sigmoid outputs. Extending to segmentation architectures like nnU-Net or to richer label taxonomies would require architectural changes and separate hyperparameter tuning. We have not tested either direction. Model and task scope: We focused on ResNet-50 with multi-label sigmoid outputs. Extending to segmentation architectures like nnU-Net or to richer label taxonomies would require architectural changes and separate hyperparameter tuning. We have not tested either direction. Prototype leakage: Embeddings derived from patient data still carry information, and an attacker with model access and side information could run inference against the embedding space. Stronger privacy mechanisms for prototype storage — beyond not sharing them and noising what is shared — are an open problem. Prototype leakage: Embeddings derived from patient data still carry information, and an attacker with model access and side information could run inference against the embedding space. Stronger privacy mechanisms for prototype storage — beyond not sharing them and noising what is shared — are an open problem. Single-seed evaluation: All quantitative results come from a single run with a fixed random seed. Multi-seed replication was not feasible given the compute cost of the full federated simulation. Margins between methods that are within 0.5–1 point should be interpreted with caution. Single-seed evaluation: All quantitative results come from a single run with a fixed random seed. Multi-seed replication was not feasible given the compute cost of the full federated simulation. Margins between methods that are within 0.5–1 point should be interpreted with caution. Future directions Adaptive noise allocation is the most natural extension. We use a fixed σ across all clients, rounds, and parameter groups. Sensitivity-aware mechanisms that assign more noise to high-risk updates and less to stable, low-information ones could narrow the accuracy gap at equivalent privacy budgets. Zheng et al.28 showed this is feasible in static FL; extending it to the continual setting is non-trivial but tractable. Robustness against adversarial clients is a separate concern. DP-SGD limits what honest clients reveal but does not protect against a site that deliberately poisons its updates. Combining DP-FedEPC with robust aggregation rules — median-based or geometric median approaches — would make the system more realistic for open or semi-trusted federations. Two further directions seem promising:Evaluating FCL under real multi-institutional collaborations spanning different regions and equipment vendors, using federated infrastructures already piloted for other imaging tasks.Extending the framework to foundation models or self-supervised pretraining, where FCL and DP constraints interact with representation learning at scale. Evaluating FCL under real multi-institutional collaborations spanning different regions and equipment vendors, using federated infrastructures already piloted for other imaging tasks. Evaluating FCL under real multi-institutional collaborations spanning different regions and equipment vendors, using federated infrastructures already piloted for other imaging tasks. Extending the framework to foundation models or self-supervised pretraining, where FCL and DP constraints interact with representation learning at scale. Extending the framework to foundation models or self-supervised pretraining, where FCL and DP constraints interact with representation learning at scale. Conclusion We investigated federated continual learning for chest radiograph classification under realistic privacy constraints, using a simulated multi-hospital federated setting. Using CheXpert and MIMIC-CXR as benchmarks, we formalized an FCL setting where hospitals act as federated clients with temporally evolving tasks and limited access to past data. To address catastrophic forgetting across sites and tasks while preserving privacy, we proposed DP-FedEPC, which combines elastic weight consolidation, prototype-based rehearsal, and client-side differential privacy within a standard FedAvg framework. DP-FedEPC reduces forgetting from 9.7 to 2.2 forgetting points under moderate noise, improves macro-AUROC across all sites and tasks compared to naive sequential fine-tuning, and maintains this advantage on MIMIC-CXR. These results hold under DP noise levels with achieved \varepsilon values consistent with health data applications in the literature. The results support the view that combining regularization, prototype rehearsal, and client-side DP can make FCL workable for hospital imaging networks that must update models sequentially under data governance constraints. The evaluation is limited to a simulated setting, and extending these findings to real deployments will require testing under asynchronous participation, larger hospital counts, and the operational privacy accounting that production environments demand.