Pith. sign in

REVIEW 3 major objections 4 minor 41 references

FedCVR, an adaptive federated aggregation method with client-level differential privacy, outperforms FedAvg on five real cardiovascular datasets, achieving 79.2% F1 and 0.96 AUC under a practical privacy budget.

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 · deepseek-v4-flash

2026-08-02 08:12 UTC pith:QTJXJP5C

load-bearing objection Real-data validation of FedCVR is a genuine step forward, but the central adaptive-advantage-under-DP claim lacks a DP-controlled FedAvg arm and the privacy accounting is under-specified. the 3 major comments →

arxiv 2607.19403 v1 pith:QTJXJP5C submitted 2026-07-07 cs.LG cs.AIcs.CRcs.CY

Recovering Clinical Utility Under Differential Privacy: Empirical Validation of Adaptive Federated Aggregation on Heterogeneous Cardiovascular Datasets

Rodrigo Tertulino , Laercio Alencar , Ricardo Almeida This is my paper
classification cs.LG cs.AIcs.CRcs.CY
keywords Federated LearningDifferential PrivacyAdaptive AggregationCardiovascular Risk PredictionNon-IID HeterogeneityClinical UtilityFedCVRHeart Disease
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 validates FedCVR, a federated learning framework with server-side adaptive (Adam-style) aggregation, on five real, heterogeneous cardiovascular datasets rather than only synthetic data. The authors claim that adaptive aggregation acts as a temporal denoiser for differential privacy noise, canceling zero-mean Gaussian noise across rounds while preserving the true gradient signal. They report that FedCVR achieves an F1-Score of 79.2% and AUC of 0.96 at a privacy budget of about 4.2, statistically outperforming standard FedAvg on all metrics tested, and that the privacy-utility degradation is graceful. A sympathetic reader would care because this suggests privacy-preserving federated learning can be practically deployed across hospitals for cardiovascular risk prediction without centralizing patient data.

Core claim

The paper reports that FedCVR preserves its adaptive advantage on real data under client-level differential privacy: an F1-Score of 79.2% and AUC of 0.96 at a noise multiplier of 0.8 (epsilon approximately 4.2), with FedCVR statistically outperforming FedAvg on all evaluated metrics (paired t-tests, all p ≤ 0.003, significant under the Bonferroni-corrected threshold). Convergence to 95% of final performance takes about 45 rounds versus roughly 85 for FedAvg, and the sensitivity analysis shows F1 declining from 79.5% (no privacy) to 78.1% at the strictest regime tested (epsilon approximately 1.2), a relative decline of less than 2%.

What carries the argument

The central mechanism is server-side adaptive moment estimation (an Adam-style optimizer at the aggregation server), which maintains exponential moving averages of the first and second moments of the aggregated client updates across communication rounds. Because the Gaussian differential privacy noise is zero-mean and independent across rounds, the first moment progressively cancels the noise in expectation while accumulating the true gradient signal—a 'temporal denoiser' effect. This server-side adaptivity is combined with client-side gradient clipping (to L2 norm C = 1.0) and calibrated Gaussian noise injection, providing client-level differential privacy where the protected unit is the pa

Load-bearing premise

The load-bearing premise is that the binarized 'high-risk' label means the same clinical thing across all five datasets; the paper does not report the Framingham threshold, so if that mapping captures a different clinical quantity, the federated model is trained on an inconsistent target and the cross-institution generalization claim collapses.

What would settle it

Re-run the experiment with a documented, clinically standard Framingham risk threshold (for example, the standard 10-year risk cutoff) instead of the unreported mapping and check whether FedCVR's advantage over FedAvg persists; or switch the privacy mechanism to record-level DP-SGD and see whether the reported negligible privacy cost (0.3 points) still holds.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the adaptive advantage holds, hospitals with scarce local data can benefit from collaborative learning without centralizing patient records, with smaller institutions gaining the most.
  • Privacy budgets in the range of epsilon 2–4 maintain utility within roughly 0.5% of the non-private baseline, giving deployers a concrete calibration range for regulatory compliance.
  • Stateless aggregation methods like FedAvg appear to be a structural bottleneck under combined non-IID data and differential privacy; server-side adaptive optimization may become a default component in privacy-preserving federated systems.
  • The client-level leave-one-institution-out evaluation protocol offers a reusable template for measuring generalization to entirely unseen institutions, a more deployment-relevant criterion than random resampling.
  • The lightweight 13-feature neural network runs on CPU-only hardware, suggesting low infrastructure barriers for clinical settings.

Where Pith is reading between the lines

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

  • The privacy guarantee claimed is client-level, not record-level; the paper's 'privacy-preserving' framing is therefore weaker than typical patient-level differential privacy. Applying record-level DP-SGD with per-batch clipping would likely incur a much larger utility cost than the reported 0.3-point drop, so the negligible-privacy-cost result likely does not transfer to record-level privacy.
  • The Framingham outcome is binarized into the same 'high-risk' label as the UCI datasets, but the threshold and mapping are not reported. If the Framingham 10-year risk score cutoff maps to a different clinical quantity, the federated model trains on an inconsistent target across sites, which would invalidate the cross-institution generalization claim regardless of aggregation improvements.
  • The statistical tests rest on only five folds (four degrees of freedom); while the FedCVR-versus-FedAvg comparisons pass the Bonferroni-corrected threshold, the non-significant high-privacy comparison (p = 0.045 uncorrected) should be read as a directional trend, not evidence that privacy loss is negligible.
  • A testable extension: rerun the same five-dataset network under record-level DP-SGD to quantify the true patient-level privacy-utility trade-off, and document the Framingham label binarization threshold to test the target-consistency premise.

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

3 major / 4 minor

Summary. This paper reports a real-data validation of FedCVR, a federated learning framework that combines server-side Adam-style adaptive aggregation with client-side gradient clipping and Gaussian noise. Five public cardiovascular datasets (Framingham, Cleveland, Hungarian, Switzerland, Long Beach VA) are harmonized to a 13-feature schema and arranged as a heterogeneous federation with leave-one-institution-out cross-validation. The authors report F1 79.2% and AUC 0.96 for the complete FedCVR model at noise multiplier σ=0.8, claim a privacy cost of only 0.3 percentage points relative to the non-private variant, and compare against FedAvg in a three-arm ablation. The paper asserts that FedCVR preserves its adaptive advantage under differential privacy on real data, with paired t-tests significant after Bonferroni correction, thereby extending prior synthetic results.

Significance. If the headline comparative claim were fully supported, this would be a valuable deployment-oriented validation: the use of five genuinely distinct public cardiovascular sources, a client-level leave-one-institution-out protocol, public code, and measured performance metrics are real strengths. The paper also honestly discusses several limitations, including the small number of folds and the instability of precision under extreme class skew. However, the main conclusion is currently not supported because the real-data ablation lacks a FedAvg arm run under the same differential privacy conditions, and the privacy accounting is not reported with enough detail for the stated ε≈4.2 to be verifiable. These are load-bearing gaps, not cosmetic issues.

major comments (3)
  1. [§4.4, Tables 3–4] The central comparison is not privacy-controlled. Table 3 defines FedAvg (Baseline) as “standard weighted averaging” and FedCVR (Complete) as adaptive + DP, but no FedAvg row is reported with noise multiplier σ=0.8 (or any DP setting). Table 4 then reports paired t-tests of FedCVR vs. FedAvg on F1, AUC, convergence, and stability. If those tests use the Table 3 baseline, they compare a DP-protected method against a non-DP method, not against FedAvg under the same client-level Gaussian mechanism. This cannot support the abstract’s claim that FedCVR “statistically outperforms standard FedAvg” under the operational privacy budget. The sentence in §4.4 claiming “identical experimental conditions” is internally inconsistent if DP status differs across arms. The synthetic benchmark result (FedAvg ~72% F1 under DP) is not real-data evidence. Please add FedAvg with the same DP noise (at least σ=
  2. [§3.1.3, Table 5] The privacy budget is not reproducibly defined. The abstract and Table 5 state ε≈4.2 for σ=0.8, but no δ is reported, and the manuscript does not give the composition formula, the number of rounds, the sampling rate (all clients participate every round?), or the privacy-accounting method used to obtain ϵ. Equations (8)–(9) define clipping and noise, but not how the cumulative ε is computed. For Gaussian mechanisms, ε depends on q, T, σ, and δ; without these, “ε≈4.2” is an assertion rather than a demonstrated privacy guarantee. Please provide the full accounting (e.g., Rényi DP composition used in the code), including the δ value and the exact protocol. Also, since §3.1.3 explicitly states this is client-level DP protecting whole-institution participation rather than record-level DP, the manuscript should avoid implying patient-level guarantees in the abstract and framing.
  3. [§3.3.2, Table 2] The outcome binarization is underspecified and may affect label comparability across institutions. The text says “The outcome variable is binarized into high-risk … and low-risk classes,” but no threshold or mapping is given. The Framingham dataset is a 10-year CVD risk score, while the UCI-derived datasets are documented heart-disease diagnoses; these are clinically different quantities. If the Framingham “high-risk” label is defined by a different risk score threshold than the diagnosis labels in the other four sites, then the federated model is trained on an inconsistent target across sites, which would undermine the cross-institution generalization claim. Please report the exact per-dataset binarization rule and justify that the resulting binary labels represent the same clinical outcome across all five clients.
minor comments (4)
  1. [§4.5 vs. Table 4] The text gives p=0.542 for Global vs. Local (Framingham), while Table 4 reports p=0.540. Please reconcile. Also clarify how “Convergence Speed” and “Training Stability” are measured per fold for the paired t-test; as written, the units for these metrics are unclear.
  2. [§3.3.3] The protocol first describes client-level leave-one-institution-out cross-validation and then a final global model trained on all five clients and evaluated on a composite test set. It is not fully clear whether Table 1 reports the final composite model or fold-aggregated results. Please state which model produced Table 1 and how the confidence intervals relate to the cross-validation folds.
  3. [References] Several references are self-citations to works marked “submitted” or to a 2026 companion paper; the present manuscript depends on that companion for the theoretical temporal-denoising result. This is acceptable, but the self-citation reliance should be made more explicit. Also, the GDPR reference appears twice in the bibliography.
  4. [§4.4, “Important Methodological Note”] The note distinguishes cross-fold SD from cross-institution SD, which is helpful. However, the same paragraph also uses “stable” to describe both fold-level consistency and clinical heterogeneity; consider using different terms to avoid confusion.

Circularity Check

0 steps flagged

No circular reduction: measured real-data validation with only minor background self-citation.

full rationale

The paper's central claim is empirical, not definitional. FedCVR is taken from prior self-cited work, and the new contribution is a measured evaluation on five real datasets: F1=79.2%, AUC=0.96, and paired t-tests against FedAvg. These are experiment outputs, not fitted parameters disguised as predictions; no equation in Sections 3-4 reduces to its own input. The self-citations to [Tertulino and Alencar, 2026] and [Tertulino, 2025] provide background motivation and a prior synthetic benchmark, but the current real-data results are independently generated and externally compared against standard FedAvg and local models, so the self-citation is not load-bearing in a circular way. The closest concern is a comparison-asymmetry issue: Table 3 defines the FedAvg baseline as standard weighted averaging without listing DP noise, while FedCVR Complete is evaluated at sigma=0.8, so the abstract's claim of outperforming 'standard FedAvg under the operational privacy budget' is not backed by a DP-perturbed FedAvg arm. That is a methodological validity gap, not a circularity: it does not make the prediction equivalent to the input by construction. The paper also acknowledges the limited statistical power of five institutional folds in Section 5.4. Overall, no circular step is established; score 2 reflects only minor background self-citation and does not indicate circular reasoning.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper's quantitative claims rest on several tuned hyperparameters (server LR, σ, C) and on unstated assumptions about label harmonization across the five datasets and about the client-level DP accounting. No new physical or architectural entities are introduced beyond the FedCVR design carried over from prior work.

free parameters (6)
  • server learning rate η_server = 1.0
    Chosen by grid search on a validation set (Section 3.5); not independently justified and affects the convergence behavior that underlies the F1=79.2 result.
  • noise multiplier σ = 0.8 (operational)
    Selected from the sweep {0.0, 0.8, 1.1, 1.5} as the 'operational privacy budget' (Table 5); the headline results are reported at this chosen value.
  • gradient clipping norm C = 1.0
    Set by hand (Section 3.5); interacts with σ to determine the effective privacy budget and the noise magnitude.
  • local epochs E = 5
    Set by hand (Section 3.3.4); influences client drift and update noise statistics.
  • batch size = 32
    Set by hand (Section 3.5).
  • server momentum β1, β2 = 0.9, 0.999
    Standard Adam values 'following standard practice' (Section 3.1.2); the temporal-denoising behavior and the reported stability depend on them.
axioms (4)
  • domain assumption The five public datasets, after harmonization to the 13-feature UCI schema with site-specific median imputation, form a realistic non-IID multi-institution network whose learned target ('high-risk vs low-risk') is consistent across sites.
    Section 3.3.2 binarizes outcomes across sources without reporting the threshold used for Framingham's 10-year risk score; if labels are not clinically equivalent across sites, the federated objective is ill-defined.
  • domain assumption Client-level DP (protecting whether an entire hospital participates) is a meaningful privacy guarantee for the claimed clinical deployment and is compatible with GDPR/LGPD.
    Explicitly stated in the 'Important Note' (Section 3.1.3); the abstract and conclusion nonetheless frame ε≈4.2 as a strong privacy guarantee, and the regulatory compatibility is asserted, not established.
  • domain assumption The Rényi/moments accountant composition over 100 rounds with noise multiplier 0.8 yields ε≈4.2 for the described client-level update perturbation.
    Table 5 reports ε≈4.2/2.2/1.2 for σ=0.8/1.1/1.5 with no δ, no sampling ratio, no composition formula, and no code in the text; the value appears assumed or computed off-screen.
  • standard math Adam's exponential moving average cancels zero-mean Gaussian DP noise in expectation without incurring extra privacy cost.
    Invoked in Section 2.1/4.4 from Tertulino and Alencar (2026); the present paper does not re-derive it.

pith-pipeline@v1.3.0-alltime-deepseek · 19720 in / 17894 out tokens · 157739 ms · 2026-08-02T08:12:32.060102+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Recovering Clinical Utility Under Differential Privacy: Empirical Validation of Adaptive Federated Aggregation on Heterogeneous Cardiovascular Datasets." pith.science (2026). https://pith.science/paper/QTJXJP5C

@misc{pith2026260719403,
  author       = {Pith},
  title        = {Pith review of: Recovering Clinical Utility Under Differential Privacy: Empirical Validation of Adaptive Federated Aggregation on Heterogeneous Cardiovascular Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QTJXJP5C}},
  note         = {Machine review of arXiv:2607.19403}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Validating federated learning frameworks on real clinical data is an essential step between proof-of-concept demonstrations in controlled synthetic environments and deployment in real multicenter healthcare settings. A prior architectural study by the same authors (Tertulino and Alencar, 2026) demonstrated, on a synthetic six-feature benchmark, that server-side adaptive optimization acts as a temporal denoiser for Differential Privacy noise, answering an open challenge identified in the original pipeline work (Tertulino, 2025). That study used synthetically generated data and explicitly identified real-world validation as a priority future direction. The present work addresses this gap by validating the FedCVR framework on five publicly available real cardiovascular datasets (Framingham, Cleveland, Hungarian, Switzerland, and Long Beach VA), harmonized to the 13-attribute UCI Heart Disease schema and configured as a heterogeneous federated scenario with leave-one-institution-out cross-validation. Results demonstrate that FedCVR preserves its adaptive advantage on real data, achieving an F1-Score of 79.2% and AUC of 0.96 under the operational privacy budget (noise multiplier = 0.8, privacy budget epsilon approximately 4.2), while statistically outperforming standard FedAvg on all evaluated metrics (paired t-tests, all p <= 0.003, significant under the Bonferroni-corrected threshold). The measured privacy cost on real data confirms the graceful degradation pattern observed in the synthetic experiments, providing empirical evidence of the framework's clinical viability in genuine multicenter contexts.

Figures

Figures reproduced from arXiv: 2607.19403 by Laercio Alencar, Ricardo Almeida, Rodrigo Tertulino.

Figure 1
Figure 1. Figure 1: FedCVR Framework Architecture. The system operates under a federated client-server paradigm, in which five heterogeneous healthcare institutions (Framingham, Cleveland, Hungarian, Switzerland, and Long Beach VA) perform local training on their private patient data. Every client computes the same type of output: a model-parameter update vector (∆θk) perturbed with calibrated Gaussian noise N (0, σ2C2 I) bef… view at source ↗
Figure 2
Figure 2. Figure 2: Receiver Operating Characteristic curve demonstrating the global model’s discriminative capacity. The Area Under the Curve of 0.96 indicates excellent separation between high-risk and low-risk patient classifications across all possible decision thresholds [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Convergence analysis comparing FedCVR (blue solid line) against FedAvg (red dashed line) over 100 communication rounds. FedCVR demon￾strates faster convergence and superior stability, effectively mitigating client drift phenomena caused by statistical heterogeneity. is high, preventing overshooting in response to outlier up￾dates. Quantitatively, FedCVR achieves 95% of its final perfor￾mance by round 42 (4… 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

41 extracted references · 4 canonical work pages

  1. [1]

    B., Mironov, I., Talwar, K., and Zhang, L

    Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., and Zhang, L. (2016). Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (CCS) , pages 308--318. DOI: 10.1145/2976749.2978318

    arXiv 2016

  2. [2]

    V., and Singh, M

    Abhishek, Bhagat, H. V., and Singh, M. (2023). A machine learning model for the early prediction of cardiovascular disease in patients. In 2023 Second International Conference on Advances in Computational Intelligence and Communication (ICACIC) , pages 1--5. DOI: 10.1109/ICACIC59454.2023.10435210

    arXiv 2023

  3. [3]

    Estudo técnico sobre anonimização de dados na lgpd: Uma visão de processo baseado em risco e técnicas computacionais

    Autoridade Nacional de Proteção de Dados (ANPD) (2021). Estudo técnico sobre anonimização de dados na lgpd: Uma visão de processo baseado em risco e técnicas computacionais. Technical report, ANPD -- Autoridade Nacional de Proteção de Dados, Brasília, Brazil

    2021

  4. [4]

    Banerjee, S., Barik, S., Das, D., and Ghosh, U. (2023). Ehr security and privacy aspects: A systematic review. In IFIP International Internet of Things Conference , pages 243--260. Springer

    2023

  5. [5]

    J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K

    Beutel, D. J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K. H., Parcollet, T., Porto Buarque de Gusm \ a o, P., and Lane, N. D. (2020). Flower: A friendly federated learning research framework

    2020

  6. [6]

    Heart disease data set

    Cleveland Heart Disease Dataset (1988). Heart disease data set. UCI Machine Learning Repository (Donated by R. Detrano). Available at: https://archive.ics.uci.edu/ml/datasets/Heart+Disease. Accessed: Jul. 26, 2024

    1988

  7. [7]

    Dubey, M., Tembhurne, J., and Makhijani, R. (2025). Enhancing federated learning through differential privacy: Introducing FedHybrid for multicenter diverse heart disease datasets. IEEE Trans. Emerg. Top. Comput. Intell. , pages 1--14

    2025

  8. [8]

    and Roth, A

    Dwork, C. and Roth, A. (2014). The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science , 9(3--4):211--407. DOI: 10.1561/0400000042

    doi:10.1561/0400000042 2014

  9. [9]

    Cardiovascular study data

    Framingham Heart Study Dataset (2022). Cardiovascular study data. Kaggle Repository (Uploaded by Aasheesh200). Available at: https://www.kaggle.com/datasets/aasheesh200/framingham-heart-study-dataset. Accessed: Apr. 24, 2023

    2022

  10. [10]

    GDPR (2016a). Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/ EC (General Data Protection Regulation)

    2016

  11. [11]

    GDPR (2016b). Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/ EC (General Data Protection Regulation)

    2016

  12. [12]

    E., Yagi, R., Homilius, M., Ichihara, G., Katsumata, Y., Gaggin, H

    Goto, S., Solanki, D., John, J. E., Yagi, R., Homilius, M., Ichihara, G., Katsumata, Y., Gaggin, H. K., Itabashi, Y., MacRae, C. A., and Deo, R. C. (2022). Multinational federated learning approach to train ecg and echocardiogram models for hypertrophic cardiomyopathy detection. Circulation , 146(10):755--769. DOI: 10.1161/CIRCULATIONAHA.121.058696

    work page doi:10.1161/circulationaha.121.058696 2022

  13. [13]

    Governance, I. (2018). Preparing for the eu gdpr in research settings / guidance. Available em: https://www.itgovernance.eu/en-ie/eu-general-data-protection-regulation-gdpr-ie. accessed: 13.10.2022

    2018

  14. [14]

    Hudaib, A., Obeid, N., Albashayreh, A., Mosleh, H., Tashtoush, Y., and Hristov, G. (2025). Exploring the implementation of federated learning in healthcare: a comprehensive review. Cluster Comput. , 28(5)

    2025

  15. [15]

    Heart disease data set

    Hungarian Heart Disease Dataset (1988). Heart disease data set. UCI Machine Learning Repository. Available at: https://archive.ics.uci.edu/ml/datasets/Heart+Disease. Accessed: Jul. 26, 2024

    1988

  16. [16]

    Heart disease dataset (comprehensive)

    IEEE Comprehensive Heart Disease Dataset (2020). Heart disease dataset (comprehensive). IEEE Dataport (Created by Manu Siddhartha). Available at: https://dx.doi.org/10.21227/dz4t-cm36. Accessed: Apr. 24, 2023

    work page doi:10.21227/dz4t-cm36 2020

  17. [17]

    Khimani, A., Hornback, A., Jain, N., Avula, P., Jaishankar, A., and Wang, M. D. (2024). Predicting cardiovascular disease risk in tobacco users using machine learning algorithms. In 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) , pages 1--5. DOI: 10.1109/EMBC53108.2024.10782885

    arXiv 2024

  18. [18]

    Kingma, D. P. and Ba, J. (2015). Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR) . arXiv:1412.6980

    Pith/arXiv arXiv 2015

  19. [19]

    Koskela, A., J \"a lk \"o , J., and Honkela, A. (2020). Computing tight differential privacy guarantees using fft. In International Conference on Artificial Intelligence and Statistics , pages 2560--2569. PMLR

    2020

  20. [20]

    J., Alagona, P., Clark, B

    Laslett, L. J., Alagona, P., Clark, B. A., Drozda, J. P., Saldivar, F., Wilson, S. R., Poe, C., and Hart, M. (2012). The worldwide environment of cardiovascular disease: prevalence, diagnosis, therapy, and policy issues: a report from the american college of cardiology. Journal of the American College of Cardiology , 60(25S):S1--S49

    2012

  21. [21]

    K., Zaheer, M., Sanjabi, M., Talwalkar, A., and Smith, V

    Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A., and Smith, V. (2020). Federated optimization in heterogeneous networks. In Proceedings of Machine Learning and Systems (MLSys) , volume 2, pages 429--450

    2020

  22. [22]

    Liu, J., Jia, J., Zhang, H., Yun, Y., Wang, L., Zhou, Y., Dai, H., and Dou, D. (2024). Efficient federated learning using dynamic update and adaptive pruning with momentum on shared server data. ACM Transactions on Intelligent Systems and Technology , 15(6):1--28

    2024

  23. [23]

    Liu, Z., Chen, Y., Zhao, Y., Yu, H., Liu, Y., Bao, R., Jiang, J., Nie, Z., Xu, Q., and Yang, Q. (2023). Carefl: Enhancing smart healthcare with contribution-aware federated learning. AI Magazine , 44(1):4--15

    2023

  24. [24]

    O., Ballard, B

    Lopez, E. O., Ballard, B. D., and Jan, A. (2023). Cardiovascular disease. In StatPearls [Internet] . StatPearls Publishing

    2023

  25. [25]

    Mammen, P. M. (2021). Federated learning: Opportunities and challenges

    2021

  26. [26]

    B., Moore, E., Ramage, D., and y Arcas, B

    McMahan, H. B., Moore, E., Ramage, D., and y Arcas, B. A. (2016). Federated learning of deep networks using model averaging. CoRR , abs/1602.05629

    Pith/arXiv arXiv 2016

  27. [27]

    Mironov, I. (2017). R \'e nyi differential privacy. In 2017 IEEE 30th Computer Security Foundations Symposium (CSF) , pages 263--275. DOI: 10.1109/CSF.2017.11

    doi:10.1109/csf.2017.11 2017

  28. [28]

    and Silva, E

    Nascimento, B. and Silva, E. (2023). Lei geral de prote c \ a o de dados ( LGPD ) e reposit \'o rios institucionais: reflex \ o es e adequa c \ o es. Em Quest. , 29(e-127314). DOI: https://doi.org/10.1590/1808-5245.29.127314

    work page doi:10.1590/1808-5245.29.127314 2023

  29. [29]

    of Standards, N. I. and (NIST), T. (2020). NIST Privacy Framework: A Tool for Improving Privacy through Enterprise Risk Management (Privacy Framework)

    2020

  30. [30]

    A., Foley, P., Gruzdev, A., Karkada, D., Davatzikos, C., et al

    Pati, S., Baid, U., Edwards, B., Sheller, M., Wang, S.-H., Reina, G. A., Foley, P., Gruzdev, A., Karkada, D., Davatzikos, C., et al . (2022). Federated learning enables big data for rare cancer boundary detection. Nature communications , 13(1):7346

    2022

  31. [31]

    Qiu, L., Cheng, J., Gao, H., Xiong, W., and Ren, H. (2023). Federated semi-supervised learning for medical image segmentation via pseudo-label denoising. IEEE journal of biomedical and health informatics , 27(10):4672--4683

    2023

  32. [32]

    J., Charles, Z., Zaheer, M., Garrett, Z., Rush, K., Kone c n \'y , J., Kumar, S., and McMahan, H

    Reddi, S. J., Charles, Z., Zaheer, M., Garrett, Z., Rush, K., Kone c n \'y , J., Kumar, S., and McMahan, H. B. (2021). Adaptive federated optimization. In International Conference on Learning Representations (ICLR) . arXiv:2003.00295

    Pith/arXiv arXiv 2021

  33. [33]

    Roh, Y., Heo, G., and Whang, S. E. (2021). A survey on data collection for machine learning: A big data - ai integration perspective. IEEE Transactions on Knowledge and Data Engineering , 33(4):1328--1347. DOI: 10.1109/TKDE.2019.2946162

    arXiv 2021

  34. [34]

    Shah , S. M. and Khan , R. A. (2020). Secondary use of electronic health record: Opportunities and challenges. IEEE Access , 8:136947--136965. DOI: 10.1109/ACCESS.2020.3011099

    arXiv 2020

  35. [35]

    J., Edwards, B., Reina, G

    Sheller, M. J., Edwards, B., Reina, G. A., Martin, J., Pati, S., Kotrotsou, A., Milchenko, M., Xu, W., Marcus, D., Colen, R. R., et al . (2020). Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Scientific reports , 10(1):12598

    2020

  36. [36]

    Silverio, A., Cavallo, P., De Rosa, R., and Galasso, G. (2019). Big health data and cardiovascular diseases: A challenge for research, an opportunity for clinical care. Frontiers in Medicine , Volume 6 - 2019. DOI: 10.3389/fmed.2019.00036

    arXiv 2019

  37. [37]

    Tertulino, R. (2025). A robust pipeline for differentially private federated learning on imbalanced clinical data using SMOTETomek and FedProx . Submitted to the Journal of the Brazilian Computer Society

    2025

  38. [38]

    and Alencar, L

    Tertulino, R. and Alencar, L. (2026). A robust framework for secure cardiovascular risk prediction: An architectural case study of differentially private federated learning. Peer-to-Peer Networking and Applications , 19(105). DOI: 10.1007/s12083-026-02265-z

    work page doi:10.1007/s12083-026-02265-z 2026

  39. [39]

    A., and Gogineni, H

    Thumula, K., Holla, H., Gutti, C., Sasikumar, A. A., and Gogineni, H. (2025). Privfed: Protecting user privacy in federated learning systems through differential privacy. In 2025 8th International Conference on Electronics, Materials Engineering & Nano-Technology (IEMENTech) , pages 1--6. DOI: 10.1109/IEMENTech65115.2025.10959665

    arXiv 2025

  40. [40]

    Yan, Y., Zhu, L., Li, Y., Xu, X., Goh, R. S. M., Liu, Y., Khan, S., and Feng, C.-M. (2024). A new perspective to boost performance fairness for medical federated learning. In Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024 , pages 13--23. Springer. DOI: 10.1007/978-3-031-72117-5\_2

    doi:10.1007/978-3-031-72117-5 2024

  41. [41]

    Yousefpour, A., Shilov, I., Sablayrolles, A., Testuggine, D., Prasad, K., Malek, M., Nguyen, J., Ghosh, S., Bharadwaj, A., Zhao, J., Cormode, G., and Mironov, I. (2021). Opacus: User-friendly differential privacy library in PyTorch

    2021