Pith. sign in

REVIEW 5 major objections 5 minor 23 references

Synthetic warm-starts stabilise federated learning, not accuracy

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-01 12:28 UTC pith:UZOKQRZV

load-bearing objection Solid synthetic-data and calibration work, but the headline FL claim is contradicted by the paper's own Tables 6 and 9. the 5 major comments →

arxiv 2607.19524 v1 pith:UZOKQRZV submitted 2026-07-21 cs.LG cs.AIcs.DC

SynPre-FL: Synthetic data-driven pretraining integrated Federated Learning training framework

Akarsh K Nair , Muhammad Arifur Rahman , Nicholas Shopland , Andy Burton , Jun He , Yuan Shen , David Baldwin , Emma O'Dowd
show 3 more authors
Amna Burzic Mufti Mahmud David J. Brown
This is my paper
classification cs.LG cs.AIcs.DC
keywords Federated LearningSynthetic data generationLatent diffusion modelElectronic health recordsPrivacy preserving AINon-IID heterogeneityProbability calibrationSHAP explainability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that the main obstacles to federated clinical risk prediction — data scarcity, severe client heterogeneity, class imbalance, and the lack of shareable benchmarks — can be addressed in one framework by combining synthetic EHR generation with federated optimisation. The central idea is to generate a privacy-safe synthetic cohort with a latent autoencoder–diffusion model, use it to warm-start the global model before federated rounds begin, then train with class-balanced FedProx objectives and Adam-style server aggregation, and finish with post-hoc calibration, threshold optimisation, and SHAP explanation on the global model only. The authors argue the framework's value is not a large accuracy jump — their own ablation shows removing synthetic pretraining leaves AUROC essentially unchanged — but improved training stability, reliable probability estimates, stable and clinically coherent feature attributions, and a reproducible benchmark for future work. A reader should care because this is one of the first complete pipelines to make synthetic data a principled component of FL rather than a test-time substitute, and it is explicit about what each component contributes.

Core claim

On its own terms, the paper's central claim is that a latent autoencoder–diffusion generator can produce synthetic EHR cohorts that preserve the univariate, bivariate, and multivariate structure of real clinical data, resist membership-inference and reconstruction attacks, and carry enough signal to warm-start federated training, and that this synthetic pretraining, combined with class-balanced FedProx local training and Adam-style aggregation, makes federated models more robust and scalable as client heterogeneity grows. The authors also claim that post-hoc logistic recalibration and threshold optimisation substantially improve probability reliability and F1, and that SHAP attribution on th

What carries the argument

The carrying mechanism is the synthetic pretraining loop: an autoencoder compresses 26-dimension encoded EHR features into a 32-dimension latent space, a 300-step diffusion model learns that latent distribution, and decoded samples have engineered clinical features recomputed and labels assigned by a gradient-boosting label-transfer model. The resulting synthetic cohort is used once, for 10 epochs, to initialise the global model before federated rounds; during federated training each client optimises a class-balanced binary cross-entropy with a FedProx proximal penalty, the server aggregates with an Adam-style update, and a held-out real validation set drives logistic recalibration and F1 th

Load-bearing premise

Every conclusion rests on the assumption that the machine-generated lung cancer cohort is a faithful stand-in for real hospital EHRs in the ways that matter — distributional shape, noise, client heterogeneity — since no real multi-centre EHR data were used.

What would settle it

Compare SynPre-FL against FedAvg without pretraining on a real multi-site EHR cohort (e.g., 3–5 hospitals): if the MIA AUROC on the synthetic cohort rises much above 0.50, or if the no-pretraining ablation continues to match or beat SynPre-FL on AUC/ECE across sites the way it does in Table 9, the paper's core claims about privacy preservation and warm-start stabilisation would be empirically refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X LinkedIn Reddit HN

If this is right

  • If synthetic pretraining stabilises non-IID federated optimisation, then hospitals with small, skewed local datasets can obtain robust global initialisations without sharing raw patient data, reducing early-round client drift in large federations.
  • If the diffusion–autoencoder generator preserves multivariate structure and resists membership inference (MIA AUROC ≈ 0.50, NN distance ratio ≈ 19), the synthetic cohort can be freely shared as a benchmark, enabling reproducible comparisons of federated algorithms.
  • If post-hoc calibration and threshold optimisation consistently improve F1 (e.g., from 0.55 to 0.68) and reduce expected calibration error (0.027 → 0.018), then clinical decision rules are better aligned with the 0.38 operating point under class imbalance.
  • If SHAP rankings remain stable across 5-, 10-, and 15-client federations, then the global model's reasoning — top features age, gender, IHD, smoking severity — is coherent and auditable at different fragmentation levels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's own ablation suggests the headline 'synthetic pretraining improves accuracy' is not what carries the framework; a reader might infer that the framework's transferable contribution is the modular pipeline (generator + warm start + calibration + SHAP) as a benchmarking template, rather than the pretraining step per se.
  • Since the generator is trained on a single machine-generated simulator dataset (not real hospital EHRs), an inference is that the method's privacy and stability claims would need re-testing on real multi-centre data, where feature missingness, coding drift, and site-specific confounders could change both MIA behaviour and warm-start value.
  • One testable extension: use the same synthetic-pretraining warm start with a centralised (non-federated) baseline on several public EHR benchmarks; if AUC gains remain near zero, the paper's contribution is best understood as a reproducibility/calibration framework rather than an accuracy method.
  • Because TRTS scores are consistently higher than TSTR scores, one inference is that the synthetic distribution is smoother/less noisy than real data; applying the same generator to noisy longitudinal EHRs might reveal whether that smoothness is a privacy benefit or a bias source.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes SynPre-FL, a framework that combines synthetic EHR generation (latent autoencoder + diffusion), task-aware label assignment via a gradient-boosting model trained on real data, synthetic pretraining of the global FL model, FedProx-style local updates, and FedAdam-style server optimization. It also applies Platt calibration and F1-threshold selection, and reports KernelSHAP explainability on the global model. The central claim, stated in the abstract, is that SynPre-FL 'consistently improves robustness and scalability over baseline methods, especially under severe non-IID fragmentation.' All experiments are conducted on a single Synthea-simulated lung-cancer EHR dataset, with 5, 10, and 15 non-IID clients, 10 FL rounds, and repeated-seed analyses for calibration and ablation.

Significance. If the central claim were supported, SynPre-FL would offer a practical, privacy-safe initialization strategy for federated clinical risk models. The paper has several commendable elements: a structured synthetic-data evaluation suite (JSD, correlation matrices, propensity-score overlap, membership-inference attacks, nearest-neighbor distance), a transparent ablation design, and repeated-seed statistical reporting. These are useful contributions to reproducibility in synthetic-EHR research. However, the evidence presented does not support the headline claim about synthetic pretraining. The paper's own Table 9 and Section 9.7 show that removing synthetic pretraining leaves AUROC essentially unchanged and is not statistically significant, while the statistically significant gains come from calibration and threshold optimization—components that are orthogonal to the synthetic-FL contribution. The abstract and contributions therefore overstate what the experiments demonstrate.

major comments (5)
  1. [§9.4, Table 6] The abstract's claim that SynPre-FL 'consistently improves robustness and scalability over baseline methods' is contradicted by Table 6. At 5 clients, SynPre-FL achieves AUC 0.8746, below FedAvg (0.8925) and FedProx+FedAdam (0.8908). At 10 and 15 clients, the gains over FedProx+FedAdam are only +0.0056 and +0.0049 AUC, within the repeated-seed standard deviation of about 0.005 reported later in Table 9. No error bars or significance tests are reported for these headline FL comparisons. The stated consistency is not supported by the table.
  2. [§9.7, Table 9] The ablation study directly undercuts the synthetic-pretraining contribution. Full SynPre-FL gives AUROC 0.8943±0.0050, while the 'no synthetic pretraining' variant gives 0.8946±0.0051; Section 9.7 states the difference is not statistically significant. In contrast, removing calibration increases ECE from 0.0159 to 0.1355, and removing threshold optimization reduces F1 from 0.6789 to 0.6172. The statistically significant benefits of the full pipeline are thus attributable to post-hoc calibration and threshold selection, not to the synthetic pretraining that is the paper's distinctive contribution.
  3. [§4.6, Eq. (6)] The synthetic labels are assigned by thresholding a gradient-boosting model h_lab that is trained on the real data and real outcome labels. The synthetic pretraining signal is therefore distilled from the very target labels the federated model is trained to predict. This makes the 'task-consistent synthetic initialization' circular with respect to the outcome, and any observed utility gain cannot be interpreted as evidence that synthetic data contribute independent information. The paper does not analyze the sensitivity of this label-transfer mechanism, nor does it compare against pseudo-labeling on real data or an alternative label-free pretraining objective.
  4. [Algorithm 2, lines 17–20] As written, the pseudocode places the weighted-averaging and Adam-update steps after the 'end for' of the round loop, implying that aggregation occurs only once after all R rounds, rather than per round. This contradicts the surrounding description in §5.5 and the standard FL protocol the paper claims to follow. If this is a formatting artifact, the pseudocode should be corrected; as it stands, it is not a faithful reference for the implemented method.
  5. [§9.1, §10.4] All experiments use a single dataset generated by the Synthea simulator, not real clinical EHR data. Section 9.1 acknowledges this restriction, and Section 10.4 correctly notes that results 'cannot be interpreted as multicenter clinical validation.' Given the abstract's practical clinical framing and the claim of robustness under realistic heterogeneity, the absence of any real-world EHR validation is a serious external-validity limitation. This alone would warrant a qualified conclusion, even if the internal results were consistent.
minor comments (5)
  1. [§1.1] The contribution list states 'Synthetic data quality is accessed using membership inference attacks' — 'accessed' should be 'assessed'.
  2. [§10.4] The limitations paragraph contains a typo: 'Several limitations can aslobe identified' should be 'can also be identified'.
  3. [Table 6] The caption reads 'Final performance (5 clients)' but the table contains blocks for 5, 10, and 15 clients. The caption should reflect all federation sizes.
  4. [§9.3, Figure 6] The text says 'Figure 6 are included for the per-round global trajectory and per-client variation,' but the figure shows only global AUC, with no per-client variation. Either add the per-client curves or correct the text.
  5. [§9.2, Table 5] The AUC column in Table 5 is labeled 'AUC' while the text uses 'AUROC' elsewhere. For consistency, unify the terminology.

Circularity Check

0 steps flagged

No significant circularity: SynPre-FL is an empirical combination of standard components; the task-aware label transfer in Eq. 6 is distillation, not a self-referential prediction.

full rationale

The paper's claims are empirical rather than derived from first principles. The core components—autoencoder+diffusion generation (Section 4), class-balanced FedProx local objective (Eqs. 7-8), FedAdam server update (Eqs. 9-11), Platt calibration (Eq. 12), and threshold selection (Eq. 13)—are standard methods with external citations; no load-bearing claim rests on the authors' prior work or on an imported uniqueness theorem. The only potentially circular-looking step is Section 4.6, where synthetic labels are assigned by thresholding h_lab trained on real data (Eq. 6). That is pseudo-labeling/distillation: the synthetic cohort inherits label information from the real labels, but the federated model is still trained and evaluated on real held-out labels, so the synthetic warm-start does not define the evaluation target. The paper's own Section 9.7 states that the difference between full SynPre-FL and the no-synthetic-pretraining variant is not statistically significant, and Section 10.4 concedes single-dataset, no-DP, and non-federated explainability limitations. Those concessions undercut the abstract's claim of consistent improvement empirically, but undercutting is not circularity. No fitted parameter is renamed as a prediction, no ansatz is smuggled in via citation, and no known result is repackaged under new coordinates. The framework is self-contained as an empirical study: every claimed contribution is either measured directly or explicitly qualified. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, forces, or formal objects. Its free parameters are mostly standard ML hyperparameters plus a validation-fitted decision threshold. The load-bearing assumptions are the representativeness of Synthea data, the validity of label transfer from h_lab, and the sufficiency of empirical privacy metrics.

free parameters (6)
  • Synthetic pretraining epochs = 10
    Hand-picked duration; ablation (Table 9) shows removing pretraining has no significant effect, so this choice is not validated by sensitivity analysis (§9.1, §9.2).
  • Synthetic cohort size M = 5000
    Set in §9.1 without sensitivity analysis; the number of synthetic samples may affect pretraining quality.
  • FedProx regularization coefficient μ = not reported
    Introduced in Eq. 8, but no value is given anywhere in the paper, making the local objective underspecified.
  • Latent dimension L and diffusion steps T = L=32, T=300
    Design choices in §4.3-4.4; no ablation or analysis of sensitivity to these values.
  • Synthetic label assignment threshold = 0.5
    Eq. 6 thresholds h_lab at 0.5 to create synthetic labels; no validation of this decision boundary.
  • Decision threshold T* = 0.38 (0.3803 ± 0.0581)
    Fitted on validation to maximize F1 (§6.3, §9.5); legitimate but a fitted operating point rather than an a priori constant.
axioms (4)
  • domain assumption Synthea-simulated EHR data are an adequate proxy for real clinical tabular EHR data when evaluating federated learning methods.
    §9.1: 'All experiments ... conducted using a publicly available EHR dataset generated with the Synthea patient simulator'; the paper's external validity rests on this single dataset.
  • ad hoc to paper Pseudo-labels generated by thresholding a real-trained gradient boosting model define a task-consistent synthetic label distribution.
    §4.6 Eq. 6: y~_j = I(h_lab(x~_j) >= 0.5). If the label-transfer model is biased, the synthetic pretraining signal is biased.
  • standard math Standard diffusion/DDPM training in the learned latent space converges to the real latent distribution.
    Assumed by the use of the DDPM forward/reverse process in §4.4-4.5; no formal convergence or distributional guarantee is established.
  • domain assumption Privacy can be adequately assessed by membership-inference AUROC and nearest-neighbor distance rather than formal differential privacy.
    §8.3 reports empirical privacy metrics and §10.4 acknowledges 'formal privacy preserving mechanisms such as differential privacy is not applied.'

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SynPre-FL: Synthetic data-driven pretraining integrated Federated Learning training framework." pith.science (2026). https://pith.science/paper/UZOKQRZV

@misc{pith2026260719524,
  author       = {Pith},
  title        = {Pith review of: SynPre-FL: Synthetic data-driven pretraining integrated Federated Learning training framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZOKQRZV}},
  note         = {Machine review of arXiv:2607.19524}
}
Share X LinkedIn Reddit HN
read the original abstract

Federated learning (FL) offers a promising approach to privacy-preserving clinical risk prediction, but its deployment remains limited by restricted data sharing, client heterogeneity, class imbalance, and the lack of realistic tabular electronic health record (EHR) benchmarks. Synthetic data generation may alleviate data scarcity, yet its integration with federated optimisation has received limited systematic study. We propose SynPre-FL, a unified framework combining high-fidelity synthetic EHR generation with synthetic-pretrained FL for robust prediction under non-IID conditions. A latent autoencoder-diffusion model generates privacy-preserving synthetic cohorts, which are used to warm-start federated training. This pretraining is followed by heterogeneity-aware optimisation using class-balanced local objectives, proximal regularisation, and adaptive server aggregation. Post-hoc calibration and federated-safe explainability support reliable and interpretable risk estimates. Experiments show that the synthetic generator preserves univariate, bivariate, and multivariate structure while protecting against membership-inference and reconstruction attacks. The generated data achieve strong downstream utility under TSTR, TRTS, and model-based evaluations. Across federated settings with 5, 10, and 15 heterogeneous clients, SynPre-FL consistently improves robustness and scalability over baseline methods, especially under severe non-IID fragmentation. Calibration improves probability reliability, while SHAP analysis produces stable and clinically coherent feature attributions across federation sizes. SynPre-FL therefore provides a practical and reproducible framework for combining synthetic data with FL to enable privacy-aware, interpretable, and robust clinical prediction from distributed tabular EHR data.

Figures

Figures reproduced from arXiv: 2607.19524 by Akarsh K Nair, Amna Burzic, Andy Burton, David Baldwin, David J. Brown, Emma O'Dowd, Jun He, Mufti Mahmud, Muhammad Arifur Rahman, Nicholas Shopland, Yuan Shen.

Figure 1
Figure 1. Figure 1: Stage-wise overview of the SynPre-FL framework illustrating the integration of synthetic EHR generation, synthetic pretraining, and heterogeneity-aware FL optimisation. The pipeline depicts the separation between offline synthetic data preparation and privacy-preserving FL training, followed by calibration and explainability modules for reliable risk prediction. adaptively scale aggregated client updates. … view at source ↗
Figure 2
Figure 2. Figure 2: Training loss of the autoencoder used for latent representation learning within the synthetic EHR generation pipeline. The reduction in reconstruction loss across 40 epochs indicates stable convergence 0 10 20 30 40 0.6 0.8 Epoch Diffusion Training Loss [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training loss of the latent diffusion model over 40 epochs. The gradual decrease demonstrates successful learning of the latent feature distribution and stable optimisation behaviour during synthetic data generation. 8. Results This section presents a comprehensive evaluation of the proposed synthetic EHR generation pipeline and its downstream application to federated learning. We report three main categor… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of real vs. synthetic correlation matrices. The synthetic data preserves major correlation structures, including clusters of respiratory symptoms and diabetic complications [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: UMAP projection of real vs. synthetic samples in latent feature space. The strong overlap across clusters indicates preservation of global multivariate structure. that the diffusion model captures the global structure of patient phenotypes. 8.3. Privacy Evaluation We evaluated privacy using two complementary ap￾proaches: (i) Membership Inference Attack (MIA) and (ii) nearest-neighbour (NN) distance analysi… view at source ↗
Figure 6
Figure 6. Figure 6: Global AUC plot over 10 FL rounds under varying client numbers in SynPre-FL configurations. The trajectory illustrates stable convergence behaviour and consistent per￾formance improvement. rounds start. Running it substantially longer would increase total computation without proportionate gain in downstream performance, especially since the pur￾pose of pretraining is not to maximise synthetic accu￾racy but… view at source ↗
Figure 8
Figure 8. Figure 8: Mean F1-score trend across decision thresholds. The optimal operating point is consistently below the default threshold of 0.5, with the calibrated tuned threshold centred around 0.38. reflected by lower ECE and Brier scores. AUROC remains unchanged across settings, as expected. This behaviour is desirable in clinical risk prediction, where the calibrated score should reflect a meaningful event probability… view at source ↗
Figure 9
Figure 9. Figure 9: Global SHAP summary visualisation for 15-client federated configuration. The distribution highlights consistent attribution patterns and preservation of clinically relevant predictors even with heterogeneous data partitions. Across all client counts, the global SHAP profiles re￾vealed a highly consistent ranking of clinically relevant predictors [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative local SHAP explanations illustrating patient-level prediction rationale. Individual feature contribu￾tions demonstrate coherent risk attribution patterns across federated client configurations. setups. This indicates that the decision patterns learned from the synthetic-pretrained global initialisation remain stable, even as the number of clients increases. 10. Discussion 10.1. Key Insights… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 1 linked inside Pith

  1. [1]

    S. Pati, S. Kumar, A. Varma, B. Edwards, C. Lu, L. Qu, J. J. Wang, A.Lakshminarayanan,S.hanWang,M.J.Sheller,K.Chang,P.Singh, D. L. Rubin, J. Kalpathy-Cramer, S. Bakas, Privacy preservation for federated learning in health care, Patterns 5 (2024) 100974. A.K.Nair et al.:Preprint submitted to ElsevierPage 17 of 18 SynPre-FL: Synthetic data-driven pretrainin...

    2024

  2. [2]

    Z. L. Teo, L. Jin, N. Liu, S. Li, D. Miao, X. Zhang, W. Y. Ng, T. F. Tan, D. M. Lee, K. J. Chua, J. Heng, Y. Liu, R. S. M. Goh, D. S. W. Ting, Federatedmachinelearninginhealthcare:Asystematicreview on clinical applications and technical architecture, Cell Reports Medicine 5 (2024) 101419

    2024

  3. [3]

    Giuffrè, D

    M. Giuffrè, D. L. Shung, Harnessing the power of synthetic data in healthcare:innovation,application,andprivacy, NPJdigitalmedicine 6 (2023) 186

    2023

  4. [4]

    S. M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions, in: I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, R. Garnett (Eds.), Advances in Neural InformationProcessingSystems,volume30,CurranAssociates,Inc., 2017

    2017

  5. [5]

    C. Sun, J. van Soest, M. Dumontier, Generating synthetic personal health data using conditional generative adversarial networks com- bining with differential privacy, Journal of Biomedical Informatics 143 (2023) 104404

    2023

  6. [6]

    J. Li, B. J. Cairns, J. Li, T. Zhu, Generating synthetic mixed- type longitudinal electronic health records for artificial intelligent applications, NPJ digital medicine 6 (2023) 98

    2023

  7. [7]

    El Emam, L

    K. El Emam, L. Mosquera, X. Fang, A. El-Hussuna, An evaluation of the replicability of analyses using synthetic health data, Scientific Reports 14 (2024) 6978

    2024

  8. [8]

    V. C. Pezoulas, D. I. Zaridis, E. Mylona, C. Androutsos, K. Aposto- lidis, N. S. Tachos, D. I. Fotiadis, Synthetic data generation methods in healthcare: A review on open-source tools and methods, Compu- tational and Structural Biotechnology Journal 23 (2024) 2892–2910

    2024

  9. [9]

    Ganzinger, M

    M. Ganzinger, M. Blumenstock, A. Fürstberger, L. Greulich, H. A. Kestler, M. Marschollek, C. Niklas, T. Schneider, C. Spreckelsen, E. Tute, J. Varghese, M. Dugas, Federated electronic data capture (fedc):Architectureandprototype, JournalofBiomedicalInformatics 138 (2023) 104280

    2023

  10. [10]

    X. Wang, H. G. Zhang, X. Xiong, C. Hong, G. M. Weber, G. A. Brat, C.-L. Bonzel, Y. Luo, R. Duan, N. P. Palmer, M. R. Hutch, A.Gutiérrez-Sacristán,R.Bellazzi,L.Chiovato,K.Cho,A.Dagliati, H. Estiri, N. García-Barrio, R. Griffier, D. A. Hanauer, Y.-L. Ho, J. H. Holmes, M. S. Keller, J. G. Klann MEng, S. L’Yi, S. Lozano- Zahonero,S.E.Maidlow,A.Makoudjou,A.Mal...

    2022

  11. [11]

    S. Li, Y. Ning, M. E. H. Ong, B. Chakraborty, C. Hong, F. Xie, H. Yuan, M. Liu, D. M. Buckland, Y. Chen, N. Liu, Fedscore: A privacy-preserving framework for federated scoring system develop- ment, Journal of Biomedical Informatics 146 (2023) 104485

    2023

  12. [12]

    Thakur, S

    A. Thakur, S. Molaei, P. C. Nganjimi, F. Liu, A. Soltan, P. Schwab, K.Branson,D.A.Clifton, Knowledgeabstractionandfilteringbased federated learning over heterogeneous data views in healthcare, npj Digital Medicine 7 (2024) 283

    2024

  13. [13]

    P.Chen,X.Du,Z.Lu,J.Wu,P.C.Hung,Evfl:Anexplainablevertical federated learning for data-oriented artificial intelligence systems, Journal of Systems Architecture 126 (2022) 102474

    2022

  14. [14]

    J. L. Corcuera Bárcena, P. Ducange, F. Marcelloni, G. Nardini, A. Noferi, A. Renda, F. Ruffini, A. Schiavo, G. Stea, A. Virdis, Enabling federated learning of explainable ai models within beyond- 5g/6g networks, Computer Communications 210 (2023) 356–375

    2023

  15. [15]

    E.Briola,C.C.Nikolaidis,V.Perifanis,N.Pavlidis,P.Efraimidis, A federated explainable ai model for breast cancer classification, in: Proceedings of the 2024 European Interdisciplinary Cybersecurity Conference, EICC ’24, Association for Computing Machinery, New York, NY, USA, 2024, p. 194–201. URL:https://doi.org/10.1145/ 3655693.3660255. doi:10.1145/365569...

    arXiv 2024

  16. [16]

    Ducange, F

    P. Ducange, F. Marcelloni, A. Renda, F. Ruffini, Federated learning of xai models in healthcare: a case study on parkinson’s disease, Cognitive Computation 16 (2024) 3051–3076

    2024

  17. [17]

    L. Zhao, H. Xie, L. Zhong, Y. Wang, Explainable federated learning scheme for secure healthcare data sharing, Health Information Science and Systems 12 (2024) 49

    2024

  18. [18]

    Abdulkadir, Explainable, domain-adaptive, and federated artificial intelligence in medicine, IEEE/CAA Journal of Automatica Sinica 10 (2023) 859–876

    A.Chaddad,Q.Lu,J.Li,Y.Katib,R.Kateb,C.Tanougast,A.Bouri- dane, A. Abdulkadir, Explainable, domain-adaptive, and federated artificial intelligence in medicine, IEEE/CAA Journal of Automatica Sinica 10 (2023) 859–876

    2023

  19. [19]

    Reddi, Z

    S. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Konečný, S. Kumar, H. B. McMahan, Adaptive federated optimization, 2021. URL:https://arxiv.org/abs/2003.00295.arXiv:2003.00295

    Pith/arXiv arXiv 2021

  20. [20]

    A. Chen, D. O. Chen, Simulation of a machine learning enabled learninghealthsystemforriskpredictionusingsyntheticpatientdata, Scientific Reports 12 (2022) 17917

    2022

  21. [21]

    Van Calster, D

    B. Van Calster, D. J. McLernon, M. Van Smeden, L. Wynants, E. W. Steyerberg,Calibration:theachillesheelofpredictiveanalytics,BMC medicine 17 (2019) 230

    2019

  22. [22]

    1321–1330

    C.Guo,G.Pleiss,Y.Sun,K.Q.Weinberger,Oncalibrationofmodern neural networks, in: International conference on machine learning, PMLR, 2017, pp. 1321–1330

    2017

  23. [23]

    Dorent, P

    R. Dorent, P. Golland, W. W. III, Connecting jensen-shannon and kullback-leibler divergences: A new bound for representation learning, 2025. URL:https://arxiv.org/abs/2510.20644. arXiv:2510.20644. A.K.Nair et al.:Preprint submitted to ElsevierPage 18 of 18

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.