Paper deep dive

A Robust Framework for Secure Cardiovascular Risk Prediction: An Architectural Case Study of Differentially Private Federated Learning

Rodrigo Tertulino, Laércio Alencar

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 103

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 7/20/2026, 5:16:02 AM

Summary

This paper presents FedCVR, a privacy-preserving Federated Learning framework for cardiovascular risk prediction that utilizes server-side adaptive optimization (Adam-based) to mitigate the effects of Differential Privacy (DP) noise and Non-IID data. The study demonstrates that FedCVR achieves a stable F1 score of 0.78 and AUC of 0.96 under a privacy budget of epsilon ≈ 13.4, outperforming stateless baselines like FedAvg and FedProx, as well as other adaptive optimizers like FedAdagrad and FedYogi. The work validates the hypothesis that server-side momentum acts as a temporal denoiser, enabling clinical utility recovery under realistic privacy constraints.

Entities (13)

FedCVR · framework · 98%Cardiovascular Risk Prediction · task · 95%Differential Privacy · technique · 95%Federated Learning · technique · 95%AUC · metric · 90%Cleveland · dataset · 90%epsilon · hyperparameter · 90%F1 Score · metric · 90%FedAdagrad · baseline-model · 90%FedAvg · baseline-model · 90%FedProx · baseline-model · 90%FedYogi · baseline-model · 90%Framingham · dataset · 90%

Relation Signals (10)

FedCVR → achievesmetric → AUC

confidence 95% · achieve a stable... Area Under the Curve (AUC) of 0.96

FedCVR → achievesmetric → F1 Score

confidence 95% · achieve a stable F1 score of 0.78

FedCVR → appliedto → Cardiovascular Risk Prediction

confidence 95% · FedCVR, a privacy-preserving Federated Learning framework applied to heterogeneous clinical networks... cardiovascular risk prediction

FedCVR → usestechnique → Differential Privacy

confidence 95% · FedCVR is a privacy-preserving Federated Learning framework... integrating Differential Privacy (DP) to provide formal, mathematical guarantees

FedCVR → usestechnique → Federated Learning

confidence 95% · FedCVR, a privacy-preserving Federated Learning framework

FedCVR → operateswith → epsilon

confidence 90% · under the operational privacy budget (epsilon approximately 13.4)

FedCVR → outperforms → FedYogi

confidence 90% · FedCVR statistically outperforms... other adaptive optimizers (FedAdagrad, FedYogi)

FedCVR → outperforms → FedAdagrad

confidence 90% · FedCVR statistically outperforms... other adaptive optimizers (FedAdagrad, FedYogi)

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Accurate cardiovascular risk prediction is crucial for preventive healthcare; however, the development of robust Artificial Intelligence (AI) models is hindered by the fragmentation of clinical data across institutions due to stringent privacy regulations. This paper presents a comprehensive architectural case study validating the engineering robustness of FedCVR, a privacy-preserving Federated Learning framework applied to heterogeneous clinical networks. Rather than proposing a new theoretical optimizer, this work focuses on a systems engineering analysis to quantify the operational trade-offs of server-side adaptive optimization under utility-prioritized Differential Privacy (DP). By conducting a rigorous stress test in a high-fidelity synthetic environment that reflects the feature space and clinical context of real-world datasets (Framingham, Cleveland), we systematically evaluate the system's resilience to statistical noise. The validation results demonstrate that integrating server-side momentum as a temporal denoiser enables the architecture to achieve a stable F1 score of 0.78 and an Area Under the Curve (AUC) of 0.96 under the operational privacy budget (epsilon approximately 13.4), compared to a non-private baseline with an F1 score of 0.84. FedCVR statistically outperforms standard stateless baselines (FedAvg, FedProx) and other adaptive optimizers (FedAdagrad, FedYogi) under identical privacy constraints. Our findings confirm that server-side adaptivity is a structural prerequisite for recovering clinical utility under realistic privacy budgets, providing a validated engineering blueprint for secure multi-institutional collaboration.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

101,532 characters extracted from source content.

Expand or collapse full text

RESEARCH Peer-to-Peer Networking and Applications (2026) 19:105 https://doi.org/10.1007/s12083-026-02265-z Rodrigo Tertulino rodrigo.tertulino@ifrn.edu.br Laércio Alencar lapea@ifrn.edu.br 1 Federal Institute of Education, Science, and Technology of Rio Grande do Norte (IFRN), Natal 3030-290, RN, Brazil Abstract Accurate cardiovascular risk prediction is crucial for preventive healthcare; however, the development of robust Artificial Intelligence (AI) models is hindered by the fragmentation of clinical data across institutions due to stringent privacy regu- lations. This paper presents a comprehensive architectural case study validating the engineering robustness of FedCVR, a privacy-preserving Federated Learning framework applied to heterogeneous clinical networks. Rather than proposing a new theoretical optimizer, this work focuses on a systems engineering analysis to quantify the operational trade-offs of server-side adaptive optimization under utility-prioritized Differential Privacy (DP). By conducting a rigorous stress test in a high-fidelity synthetic environment that reflects the feature space and clinical context of real-world datasets (Framing- ham, Cleveland), we systematically evaluate the system’s resilience to statistical noise. The validation results demonstrate that integrating server-side momentum as a temporal denoiser allows the architecture to achieve a stable F1-score of 0.78 and an Area Under the Curve (AUC) of 0.96 under the operational privacy budget ( ε≈13.4 ), compared to a non-private baseline of F1-score 0.84. FedCVR statistically outperforms standard stateless baselines (FedAvg, FedProx) and other adaptive optimizers (FedAdagrad, FedYogi) under identical privacy constraints. Our findings confirm that server-side adaptivity is a structural prerequisite for recovering clinical utility under realistic privacy budgets, providing a validated engineering blueprint for secure multi-institutional collaboration. Keywords Federated learning · Differential privacy · Cardiovascular disease · Adaptive optimization · System architecture · Synthetic data Received: 28 February 2026 / Accepted: 13 June 2026 © The Author(s) 2026 A robust framework for secure cardiovascular risk prediction: An architectural case study of differentially private federated learning Rodrigo Tertulino 1 · Laércio Alencar 1 limits both the statistical power and the geographic repre- sentativeness of predictive models, with consequences that are particularly severe for low and middle-income coun- tries, which bear a disproportionate share of the global car- diovascular disease burden [3]. From a systems engineering perspective, the core obsta- cle is that the institutions most capable of improving model quality are precisely those most constrained from sharing their data. Regulations such as GDPR [4] in Europe and HIPAA [5] in the U.S. impose strict boundaries on data movement, effectively transforming each clinical site into an isolated data silo [6]. Any viable solution must therefore operate entirely within those boundaries, treating data pri- vacy not as a post-hoc constraint but as a first-class archi- tectural requirement. To address data privacy challenges, Federated Learning (FL) has been introduced as a distributed ML approach [7] in which raw data never leaves each institution’s premises. Our prior work [8] validated a baseline FL pipeline for 1 Introduction The growing adoption of Artificial Intelligence (AI) in clini- cal decision support has positioned cardiovascular risk pre- diction as one of the most impactful application domains in health informatics. Multi-institutional datasets are essential for training models that generalize across diverse patient populations; yet assembling such datasets remains a per- sistent challenge, as clinical records are dispersed across hospitals, subject to strict regulatory controls, and rarely shareable in raw form [1, 2]. The resulting fragmentation 1 3 105 Page 2 of 23Peer-to-Peer Networking and Applications (2026) 19:105 cardiovascular risk prediction, demonstrating that combin- ing FedProx with client-side SMOTETomek can achieve clinically useful Recall under Differential Privacy. How- ever, that work identified a critical open challenge: under the simultaneous pressure of Non-IID data distributions and DP noise, even robust aggregators experience significant utility degradation, and the question of whether server-side adaptive optimization can systematically recover this lost utility remained unanswered. The primary objective of this study is to address this gap by implementing and rigorously evaluating FedCVR, a framework that replaces stateless aggregation with an Adam-based adaptive optimizer at the server level, effec- tively functioning as a temporal denoiser across commu- nication rounds. By benchmarking against five distinct baselines (FedAvg, FedProx, FedCluster, FedAdagrad, and FedYogi) under identical privacy constraints, we isolate the specific contribution of server-side momentum to utility recovery under DP. Furthermore, this work provides a critical analysis of the privacy-utility trade-off by integrating Differential Privacy (DP) to provide formal, mathematical guarantees against inference attacks [9]. Local DP is applied by injecting cali- brated Gaussian noise into client gradients before transmis- sion, and we quantify how the adaptive optimizer modifies the effective privacy cost experienced at the global model level. The primary contributions of this paper are summarized as follows: ●Robust Framework for Clinical Non-IID Data: We develop FedCVR, a framework incorporating server- side adaptive optimization designed to mitigate gradient divergence caused by the statistical heterogeneity (Non- IID) inherent in multi-institutional medical data. ●Comprehensive Benchmarking & Resilience: We evaluate FedCVR against five distinct baselines: Fe- dAvg, FedProx, FedCluster, FedAdagrad, and FedYogi. We demonstrate that the specific momentum-based tun- ing in FedCVR acts as a superior temporal denoiser, sig- nificantly outperforming both stateless and other adap- tive optimizers in the presence of Differential Privacy (DP) noise. ●Privacy-Utility Trade-off Analysis: We provide a rig- orous empirical analysis of the impact of DP noise on model utility. Unlike previous works that treat DP as a black box, we quantify the specific privacy budget (ε) required to maintain clinical viability (Recall/Sensitiv- ity) without compromising patient anonymity. ●Statistical Validation: We report comprehensive per- formance metrics (F1-Score, AUC, Recall) validated through independent runs ( N=5 ) and statistical signif- icance testing ( p<0.05 ), ensuring reproducibility and confirming that the performance gains are not artifacts of random initialization. This paper is structured as follows: Section 2 reviews the relevant literature on FL, contextualizing the work within the current state of the art. Section 3 provides the technical back- ground on federated optimization under the dual challenge of Non-IID data and DP noise, establishing the motivation for the FedCVR design. Section 4 distinguishes the framework from standard approaches, followed by Section 5, which details the specific architecture of the FedCVR aggregation model. Section 6 discusses data availability challenges in healthcare and justifies the approach of using generated data. Section 7 details the proposed methodology, including the dataset characteristics and the FL-DP implementation. Sec- tion 8 presents the empirical results from the experiments, featuring a comparative analysis of the baseline and DP- enhanced models. Section 9 discusses the broader implica- tions of these findings, and finally, Section 10 concludes the paper and outlines potential directions for future research. 2 Related work Applying ML to large-scale EHRs holds immense potential for advancing Cardiovascular Disease (CVD) research [10]. However, progress is fundamentally constrained by the chal- lenge of “data silos,” where valuable clinical data remains fragmented across different institutions due to stringent pri- vacy regulations and security concerns [11]. Additionally, data fragmentation limits the development of robust, gen- eralizable models that perform well across diverse patient populations [12, 13]. Federated Learning (FL) addresses these challenges by enabling collaborative ML across multiple sites without requiring the sharing of raw data, thereby mitigating issues of inconsistent data quality, unbalanced datasets, and data privacy. Its privacy-preserving characteristics are crucial for maintaining compliance with stringent regulations, includ- ing GDPR and HIPAA [14, 15]. A significant challenge in real-world FL is statistical heterogeneity, where data distributions differ substantially between institutions (i.e., data is non-IID). The work by [16] provides a demonstration of this issue and FL’s advantage. Their study on detecting hypertrophic cardiomyopathy found that models trained at a single institution generalized poorly when tested on data from another country. However, a multi- national FL approach significantly improved the model’s gen- eralizability, underscoring FL’s ability to create more robust models by leveraging diverse, international data sources. A significant hurdle in medical research has tradition- ally been the acquisition of sufficiently large and diverse 1 3 Page 3 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 datasets for model training within individual institutions, often compounded by the complexities of privacy and data ownership associated with centralized patient data sharing [17]. FL offers a compelling solution by facilitating multi- institutional collaboration through distributed model train- ing, where data remains localized with its owners, and only aggregated model updates are shared. Studies have corroborated FL’s efficacy, demonstrating that it can achieve up to 99% model quality comparable to centralized approaches, even when training involves ten dis- tinct institutions [18]. In this work [19], a global FL initiative leveraged data from 6,134 glioblastoma patients across 71 geographically diverse sites, resulting in a robust and gen- eralizable model for glioblastoma sub-compartment detec- tion. This emphasizes the critical role of diverse datasets in developing clinically impactful models. Similarly, FL’s diagnostic utility was validated in hypertrophic cardiomy- opathy using M&M and ACDC datasets subsets, confirm- ing performance comparable to centralized systems while upholding patient privacy [20]. Beyond data heterogeneity, the operational deployment of FL introduces more complex challenges related to resource constraints and data lifecycle management [21]. A significant hurdle is the computational heterogeneity across clients, with some institutions lacking the powerful GPUs required for local training of large models. To address this, Wang et al. [22] proposes CAFE AU LAIT, a compute- aware framework in which resource-constrained clients contribute by generating data, which is then used by a strong central server to participate in federated training alongside more powerful clients. While this approach democratizes participation, it also highlights a potential performance gap between models trained on synthetic data versus those trained on real, local updates, presenting a clear trade-off between inclusivity and model utility. Synthetic data has also been explored to solve another critical issue: data incompleteness. The FLIGAN frame- work, proposed by [23], leverages federated GANs to generate high-quality synthetic data that augments client datasets suffering from severe class imbalances or insuffi- cient data volume. Their results show that this technique can significantly improve model accuracy, though it introduces additional computational overhead during training. Further- more, the data lifecycle within a trained model is a growing concern, driven by regulations such as the “right to be for- gotten.” Naively retraining a model from scratch to unlearn a user’s data is computationally prohibitive. In response, efficient Federated Unlearning (FU) methods, such as QUICKDROP, have been developed that use a small, dis- tilled synthetic dataset to approximate gradients, enabling rapid, computationally inexpensive removal of data from a global model [24]. A consistent theme throughout the literature is the advan- tage of FL in facilitating multi-institutional collaboration that would otherwise be impossible. Studies have shown that FL can achieve model quality comparable to that of centralized approaches, even across multiple participating sites, underscoring the critical role of diverse datasets in creating clinically impactful models. Beyond diagnostics, FL has been instrumental in advancing research on complex diseases, such as brain tumors, by enabling collaboration without disclosing private patient information [25]. Meanwhile, FL inherently provides a strong baseline for privacy by keeping data localized; the model updates are not immune to inference attacks [26]. Most of the cited works primarily focus on addressing the challenge of model per- formance under data heterogeneity. A critical, often unad- dressed, limitation or future work noted in these studies is the need for stronger mathematical privacy guarantees. However, the challenge of data fragmentation due to privacy constraints restricts the full potential of FL in clinical applications. While many studies propose FL as a solution for privacy-preserving healthcare AI, a significant gap remains: most existing work addresses model perfor- mance under data heterogeneity but stops short of integrat- ing formal, mathematically grounded privacy guarantees. Researchers continue to highlight DP-enhanced FL as a via- ble direction for digital health [27]. The present work fills this gap by implementing a federated framework for cardio- vascular risk prediction with Differential Privacy, conduct- ing a rigorous comparative analysis of the privacy-utility trade-off, and providing empirical evidence of the practical viability of a truly secure FL system. 3 Background: federated optimization under privacy constraints This section provides the technical foundations for contex- tualizing the FedCVR design. Readers already familiar with the standard FL formulation (as formalized in, e.g., [7]) are directed to our prior work [8] for a detailed treatment. Here, we focus on the two specific failure modes that motivate the FedCVR architecture: client drift under Non-IID data and gradient corruption under Differential Privacy. 3.1 The federated optimization objective In a Horizontal FL system, N clients each hold a local private dataset D i . The global objective is to minimize the weighted average of local loss functions without centralizing data: F(w)= N ∑ i=1 |D i | |D| F i (w),where|D|= N ∑ i=1 |D i | (1) 1 3 105 Page 4 of 23Peer-to-Peer Networking and Applications (2026) 19:105 In each communication round t, the server broadcasts the current global model w t to a subset S t of clients; each client minimizes F i locally via gradient descent ( w (t+1) i =w t −η∇F i (w t ) ) and returns its update; the server aggregates these updates into a new global model w t+1 . 3.2 Why standard aggregation fails under the dual challenge Standard FedAvg [7] operates a stateless aggregation: the server treats each round independently, computing w t+1 as a weighted average of client updates with no memory of the optimization trajectory. This design is adequate under IID data and without privacy noise, but breaks down along two orthogonal axes simultaneously present in real-world clini- cal deployments: 1. Statistical Heterogeneity (Non-IID): When clinical data distributions differ across institutions, each local gradient ∇F i (w t ) points toward a different local opti- mum. Their arithmetic mean diverges from the true global gradient, a phenomenon known as client drift. Proximal regularization (FedProx [28]) partially miti- gates this by penalizing excessive local deviation but leaves server-side aggregation stateless. 2. DP Noise Corruption: When Differential Privacy is enforced, each transmitted gradient g k t contains Gauss- ian noise: g k t =∇F k (w t )+N(0,σ 2 C 2 I) . Stateless aggregators directly incorporate this noise into the model update at each round, leading to erratic convergence. The noise variance does not decrease over rounds under a fixed σ, so the problem accumulates. The central hypothesis of this work is that a stateful server- side optimizer, one that maintains a temporal history of gradients, can act as a denoising filter across rounds. By averaging zero-mean Gaussian noise across multiple com- munications, momentum-based adaptivity transforms pri- vacy noise from a persistent adversary into a manageable stochastic perturbation. The FedCVR architecture opera- tionalizes this hypothesis, as detailed in Section 5. 4 Distinguishing FedCVR from state-of-the- art baselines To articulate the engineering advantages of the proposed framework, it is necessary to distinguish FedCVR not only from the canonical Federated Averaging (FedAvg) but also from regularization-based, clustering-based, and other adaptive baselines used in this study. The distinctions are rooted in how each method handles the server-side aggrega- tion state and responds to the noise introduced by Differen- tial Privacy (DP). 4.1 Versus stateless aggregators (FedAvg and FedProx) FedAvg and FedProx operate under a stateless server-level paradigm. ●FedAvg computes the global model simply as a weight- ed average of client updates. It treats every communica- tion round as an independent event, lacking any “memo- ry” of the optimization trajectory [7, 29]. In the presence of DP noise, this statelessness is detrimental: the random Gaussian noise added in round t directly impacts the model w t+1 without any historical smoothing, leading to erratic convergence. ●FedProx improves upon FedAvg by adding a proximal term ( μ 2 ||w−w t || 2 ) to the client’s local loss function to limit local drift [28, 30]. However, the server aggrega- tion remains a simple average. While FedProx handles statistical heterogeneity better than FedAvg, it offers no specific mechanism to filter out the aggregation noise caused by privacy mechanisms. On scalability with increasing client counts: as the num- ber of participating clients N grows, the aggregated gra- dient ̄g t benefits from statistical variance reduction (the mean’s noise variance scales as σ 2 /|S t | ). For FedCVR, this implies that the server learning rate η s could be mod- erately increased in larger networks without destabilizing the momentum update, as each round provides a less noisy gradient estimate. This is an inherent architectural advan- tage over stateless baselines, where larger N under Non-IID settings amplifies client drift. Empirical validation of this scaling behavior for N≫5 is an open research direction that we identify as a priority for future work. 4.2 Versus clustering strategies (FedCluster) FedCluster employs a structural approach by grouping clients with similar data distributions to reduce Non-IID effects [31]. While effective for heterogeneity, it essentially runs parallel instances of FedAvg within clusters. It does not fundamentally alter the gradient optimization steps. Consequently, FedClus- ter still suffers from the same sensitivity to DP noise within each cluster as standard FedAvg. Furthermore, FedCVR offers a more streamlined deployment by eliminating the preprocess- ing step of similarity analysis and cluster formation. 1 3 Page 5 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 4.3 Versus other adaptive optimizers (FedAdagrad and FedYogi) The closest relatives to FedCVR are other server-side adap- tive optimizers proposed in the FedOpt framework [32]. Importantly, the FedCVR aggregation rule is functionally equivalent to FedAdam as defined in Reddi et al. [32]; it maintains first and second moment estimates ( m t ,v t ) and applies bias-corrected adaptive updates. The contribution of this work does not lie in proposing a new optimizer, but in providing a systematic empirical validation of the Adam- based adaptive update as a DP noise denoiser in the specific domain of cardiovascular risk prediction under clinically realistic Non-IID conditions, a question that prior theoreti- cal work left as an open engineering problem. Within the FedOpt family, distinct mechanical differ- ences explain the performance ordering observed in our experiments: ●FedAdagrad adapts learning rates based on the ac- cumulation of squared gradients (second moment) but lacks a momentum component (first moment) [33]. Without momentum, FedAdagrad cannot effectively av- erage out the oscillatory noise introduced by DP across rounds, as the temporal filtering effect requires both mo- ment estimates. ●FedYogi utilizes both first and second moments, mak- ing it structurally closest to FedCVR [34]. However, FedYogi employs an additive update rule for the second moment: v t ←v t−1 +(1−β 2 ) ̄g 2 t . Under DP-SGD, the gradient clipping operation bounds each gradient to a norm of C, systematically reducing gradient magni- tudes. The additive rule in FedYogi tends to accumulate an increasing second moment across rounds, despite this suppression, effectively inflating the denominator and causing the adaptive step size to shrink progressively. Under high-variance DP noise, this accumulation is ex- acerbated, causing FedYogi to underestimate the true learning signal and converge more slowly than the mul- tiplicative update used in FedCVR. The multiplicative rule ( v t ←β 2 v t−1 +(1−β 2 ) ̄g 2 t ) in FedCVR is more responsive to the current gradient estimate, preserving adaptivity even when clipped gradients are small. 4.4 The FedCVR advantage: stateful denoising In contrast to the methods above, FedCVR’s server-side momentum introduces a temporal filtering effect. By main- taining the first moment vector m t , the server computes a moving average of successive gradient estimates: m t =β 1 m t−1 +(1−β 1 )(∇L+g noise,t ) (2) Since g noise,t ∼N(0,σ 2 C 2 I) is i.i.d. zero-mean across rounds, its contribution to m t decays as β k 1 with round lag k, causing E[m t ]→∇L as t→∞ . This temporal denois- ing is the primary architectural mechanism that enables FedCVR to recover clinical utility under DP budgets where stateless aggregators fail to converge reliably. 5 FedCVR aggregation model The core of the FedCVR framework is its server-side adap- tive optimization logic, which stabilizes convergence in the presence of differential privacy noise. Unlike standard FedAvg, which is stateless and updates the global model w by simply averaging positional weights, FedCVR maintains internal state vectors, specifically, the first ( m t ) and second ( v t ) moment estimates of the gradients. These states allow the server to act as a centralized optimizer, adaptively scal- ing the learning rate for each parameter based on the histori- cal trajectory of updates. In this framework, clients function as distributed gra- dient calculators. In each communication round t, a set of clients S t compute gradients g k t on their local private data. These gradients are transmitted to the server (potentially with added noise for privacy, as detailed in Section 5.2). The server then executes the adaptive aggregation step as follows: 1. Gradient Aggregation: The server computes the true average pseudo-gradient ̄g t from the participating clients: ̄g t ← 1 |S t | ∑ k∈S t g k t (3) 2. Moment Estimation: The server updates the exponen- tial moving averages of the gradient ( m t , representing momentum) and the squared gradient ( v t , representing uncentered variance): m t ←β 1 m t−1 +(1−β 1 ) ̄g t (4) v t ←β 2 v t−1 +(1−β 2 ) ̄g 2 t (5) 3. Bias Correction: To counteract the initialization bias towards zero (especially in early rounds), bias-corrected estimates ˆm t and ˆv t are computed: ˆm t ← m t 1−β t 1 (6) ˆv t ← v t 1−β t 2 (7) 1 3 105 Page 6 of 23Peer-to-Peer Networking and Applications (2026) 19:105 4. Adaptive Update: Finally, the global model parameters w t are updated using the adaptive learning rate derived from the moments: w t+1 ←w t −η ˆm t √ ˆv t +τ (8) Here, η represents the server-side learning rate, while β 1 and β 2 control the exponential decay rates for the moment esti- mates (typically set to 0.9 and 0.999, respectively), and τ is a small scalar for numerical stability. This mechanism effec- tively filters out high-frequency noise introduced by both data heterogeneity (Non-IID) and the Differential Privacy mechanism, ensuring a smoother convergence trajectory. 5.1 System orchestration and privacy preservation The complete operational workflow of FedCVR integrates the server-side adaptive aggregation with a privacy-preserv- ing client-side execution. This architecture, illustrated in Fig. 1, orchestrates the interplay between the central server and the distributed clients to ensure both convergence sta- bility and data privacy. The server-side coordination is formally described in Algorithm 1. The server acts as the central synchronizer, responsible for initializing parameters, selecting the subset of clients S r for each round r, and applying the previously defined adaptive FedCVR update rule. This centralization removes the computational burden from clients, leaving them solely responsible for computing the gradient. Algorithm 1 FedCVR server-side orchestration. Complementing the server, the client-side procedure (Algorithm 2) enforces the privacy constraints. Instead of Fig. 1 System architecture of the FedCVR framework. The diagram illustrates the cyclic federated process: The central server broadcasts the global model ( w t ) to participating hospitals (Clients). Each client trains locally on private data and applies a Differential Privacy (DP) noise mechanism before uploading the protected gradients ( ̃g t ) for adaptive aggregation 1 3 Page 7 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 transmitting raw gradients, each client executes Differen- tially Private Stochastic Gradient Descent (DP-SGD). This involves two critical steps before transmission: (1) Clipping the per-sample gradients to a maximum norm C to bound the sensitivity of the update, and (2) Noise Injection, where Gaussian noise is added to the aggregated batch gradient. This ensures that the update ̃g sent to the server satisfies Dif- ferential Privacy, preventing the reconstruction of sensitive patient records from the gradient information. Algorithm 2 ClientUpdate with DP-SGD. 5.2 Privacy preservation in federated learning Although FL’s decentralized design prevents direct data sharing, transmitted model updates remain vulnerable to inference attacks that can reconstruct individual records [35]. Differential Privacy (DP) provides a formal, mathe- matical shield against such attacks. The (ε,δ) -DP guarantee, for any two adjacent datasets D 1 and D 2 and any output set S, is: Pr[M(D 1 )∈S]≤e ε ·Pr[M(D 2 )∈S]+δ (9) where ε (privacy budget) bounds the maximum information leakage from any single record, and δ is the probability that this bound fails. A smaller ε yields a stronger privacy guar- antee. In our multi-institutional FL context, DP is imple- mented at the client level using DP-SGD [36]: per-sample gradients are clipped to a maximum L 2 norm C and Gauss- ian noise proportional to σC is added before transmission. The formal mechanism and its interaction with the FedCVR adaptive optimizer are detailed in Section 5.3. The key engineering insight exploited by FedCVR is that the noise injected by DP-SGD is zero-mean Gaussian, i.e., E[g noise ]=0 . A stateful server that maintains a tem- poral history of gradients via exponential moving averages can progressively cancel this noise across rounds, whereas a stateless aggregator (FedAvg) absorbs the full noise variance at each step. Algorithm 3 formalizes the compara- tive procedure used to isolate this effect. Algorithm 3 Comparative experiment: FedCVR vs. DP-FedCVR. The two-phase design ensures that any performance dif- ference is attributable solely to the privacy mechanism. This allows a direct measurement of the “cost of privacy” under the FedCVR adaptive optimizer, which is then compared against the same cost measured for stateless baselines in Section 8. In healthcare, DP has gained traction due to the grow- ing need for robust data protection amid increasingly strin- gent regulatory environments [37]. The significance of DP extends beyond compliance, fostering trust among patients regarding data handling [38]. However, the central chal- lenge remains that injected noise must not overly degrade diagnostic accuracy, which is vital for clinical decision- making [39]. As healthcare data is digitized, integrating DP becomes increasingly critical to safeguard patient identities [40]. 5.3 Differential privacy implementation and budgeting We implement Local Differential Privacy using the Gauss- ian Mechanism via Opacus [36, 41]. For every batch, gradi- ents are clipped to a maximum norm C and noise is added: ̃g= g clipped +N ( 0,σ 2 C 2 I ) (10) The clipping threshold C=1.0 is kept constant and uniform across all clients throughout all experiments. A site-specific sensitivity threshold could, in principle, better accommodate local data variance; however, it would reveal 1 3 105 Page 8 of 23Peer-to-Peer Networking and Applications (2026) 19:105 per-client gradient magnitudes to the server, introducing an additional privacy surface. Adaptive per-client clipping is therefore left as future work. All client training configura- tions use this shared C, ensuring that the privacy account- ing is consistent and reproducible via the Opacus RDP accountant. 5.3.1 Privacy budget accounting (RDP) To rigorously track privacy loss across T=100 communi- cation rounds, we employ Rényi Differential Privacy (RDP) accounting rather than the standard (ε,δ) -accounting. RDP provides tighter composition bounds for iterative mecha- nisms, such as SGD. We convert the RDP budget back to standard (ε,δ) for interpretability at the end of training. The privacy budget is computed using the actual mini-batch subsampling parameters of our implementa- tion. With a batch size L=32 and a local dataset per client of |D k |≈6,000 samples (30,000 total samples distributed over 5 clients), the effective per-step sam- pling rate is q=L/|D k |≈0.0053 . Each client performs E=5 local epochs per round, yielding approximately |D k |/L·E=937.5 optimization steps per round and T≈93,750 total DP-SGD steps over 100 rounds. Given noise multiplier σ=1.0 and δ=10 −5 (satisfying the stan- dard requirement δ≪1/|D k | ), the RDP accountant yields a cumulative budget of ε≈13.4 . All accounting was per- formed using the Opacus RDP accountant to ensure cor- rectness. The experiments assume full client participation across all 100 rounds (no dropout). In scenarios with cli- ent dropout, the RDP accountant should be re-run using the actual number of participants per round, since the cumu- lative ε depends on the product q×T ; the Opacus library supports this adjustment via its subsampling_scheme parameter. 5.3.2 Justification of the utility-prioritized regime While theoretical cryptography often strives for ε<1 , engineering deployments in complex healthcare tasks typi- cally require a “Utility-Prioritized” regime ( ε∈[10,15] ). In Federated Learning, where the threat model focuses on preventing the server from reconstructing exact inputs rather than protecting against adversaries with infinite computational power, this range is widely accepted as the operational “sweet spot” [7, 42]. Comparable ε values have been reported in large-scale deployments: Apple’s iOS key- board uses ε≈8 for on-device learning, and the U.S. Cen- sus Bureau deployed differential privacy with ε≈17.14 for the 2020 Decennial Census [42]. In FL healthcare settings, several peer-reviewed frameworks operate in the ε∈[5,20] range while maintaining clinical viability [36]. As detailed in Table 1, forcing a “Strict Academic” bud- get ( ε<1 ) in high-dimensional risk prediction typically results in random-guess performance (AUC ≈ 0.5), render- ing the system clinically useless. Our architectural choice of ε≈13.4 reflects a deliberate engineering trade-off that balances diagnostic precision with the prevention of trivial data leakage. 5.4 Client-side optimization: local DP-SGD To strictly adhere to differential privacy guarantees while maintaining computational efficiency, we implement Local Differentially Private Stochastic Gradient Descent (DP- SGD). Contrary to full-batch gradient descent approaches, often simplified in the FL literature, which compute gradi- ents over the entire local dataset D k in a single step, our framework employs mini-batch optimization, consistent with the Opacus library implementation. The local training process on client k proceeds as follows for each training round: 1. The local dataset D k is partitioned into mini-batches B, each of size L (Sampling Rate q=L/|D k | ). 2. For each mini-batch, per-sample gradients g i =∇ℓ(w;x i ,y i ) are computed individually for each sample (x i ,y i )∈B . This step is crucial for DP as it allows regulating the influence of individual records. 3. To bound the sensitivity of the update, each per-sample gradient is clipped to a maximum L 2 norm C: ̄g i =g i /max ( 1, ∥g i ∥ 2 C ) (11) Table 1 Operational privacy regimes in federated healthcare systems. The selected regime for this case study is highlighted Regime ε rangeOperational context & implications Strict / Academic <1.0 High Privacy, Low Utility. Theo- retical guarantee, but typically destroys model convergence in deep learning tasks. Moderate 1.0−10.0 Standard Trade-off. Used in simple convex problems; often requires massive datasets ( N>10 6 ) to maintain utility. Utility-Prioritized 10.0 - 15.0 Balanced (Current Study). Prevents reconstruction attacks while maintaining clinical validity (AUC >0.9 ). Typical for trusted consortia. Weak >20.0 Low Privacy. Vulnerable to mem- bership inference attacks; offers little advantage over obfuscation. 1 3 Page 9 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 4. Gaussian noise is added to the sum of clipped gradients before the optimization step: ̃g B = ∑ i∈B ̄g i +N(0,σ 2 C 2 I) (12) 5. The local model parameters are updated using the sto- chastic noisy gradient: w←w−η 1 L ̃g B . This mechanism ensures that the privacy budget ( ε ) is accounted for using the subsampled Gaussian mechanism rather than assuming a full-batch composition, which would yield looser privacy guarantees. 5.5 Performance metrics Model performance is evaluated using the standard binary classification metrics summarized in Table 2. Following established practice in clinical ML [43], we place particu- lar emphasis on Recall (Sensitivity) and F1-Score, as the cost of a false negative (missed high-risk patient) greatly exceeds that of a false positive in cardiovascular screen- ing. The AUC of the ROC curve provides a threshold-inde- pendent measure of discriminative power. Communication Rounds N R assesses convergence efficiency specific to the federated context [44]. 6 Dataset and federated learning methods 6.1 Types of federated learning Different types of FL in healthcare address unique data distribution challenges across medical institutions and devices. These variations in FL models ensure that sensitive healthcare data can be utilized for advanced model training while maintaining privacy and compliance with regulatory standards [45]. The primary types of FL used in healthcare include Horizontal Federated Learning (HFL), Vertical Federated Learning (VFL), and Federated Transfer Learning (FTL), each of which caters to specific scenarios depending on how the data is distributed among the participants, mak- ing them vital in different healthcare settings [46]. The FedCVR framework presented in this study is an application of Horizontal Federated Learning (HFL). There- fore, this approach was chosen because it directly models the common real-world scenario in which collaborating institutions, such as the simulated hospitals in the experi- ments, collect data using a common schema but for differ- ent patient cohorts. Specifically, all clients in the framework share an identical feature space (e.g., age, systolic blood pressure, cholesterol) and predictive target (risk), while each client’s dataset contains a unique, non-overlapping set of samples. Horizontal data partitioning closely resembles typical multi-institutional clinical research, making HFL the appropriate and natural strategy for the cardiovascular risk prediction task. Figure 2 illustrates the HFL approach used in this study. The conceptual global dataset is horizontally partitioned, with each participating hospital holding a dis- tinct set of patient samples while sharing an identical feature space. The sustainability and advancement of clinical AI rely on access to diverse datasets; however, the sensitive nature of medical records presents serious privacy risks, necessi- tating safeguards. While strategies such as data consortia and cross-institutional aggregation platforms are emerging, they often involve complex legal agreements and lengthy approval processes [47]. 6.2 Model architecture To balance predictive power with communication effi- ciency, we employed a lightweight Deep Neural Network (DNN) designed for tabular clinical data. The architecture consists of: ●Input Layer: 6 neurons corresponding to the selected clinical features (e.g., Age, Hypertension, Glucose Level). ●Hidden Layers: Two fully connected layers with 64 and 32 neurons, respectively, utilizing ReLU activation functions to capture non-linear interactions between risk factors. ●Output Layer: A single neuron with a Sigmoid activa- tion function to output the probability of cardiovascular risk P(y=1|x)∈[0,1] . The model was implemented in PyTorch and optimized using Binary Cross-Entropy (BCE) loss. Table 2 Performance metrics used in this study. TP=True Positive, TN=True Negative, FP=False Positive, FN=False Negative MetricFormula / Definition Accuracy (TP+TN)/(TP+TN+FP+FN) Precision TP/(TP+FP) Recall (Sensitivity) TP/(TP+FN) F1-Score 2· Prec·Rec Prec+Rec Global Loss L global = ∑ C i=1 N i N L (i) local Comm. Rounds N R Total server ↔ client cycles to convergence 1 3 105 Page 10 of 23Peer-to-Peer Networking and Applications (2026) 19:105 6.3 Datasets To establish a clinically relevant foundation for the FL framework, we analyzed the feature space and demographic characteristics of several prominent multicenter datasets, including the Framingham Heart Study (FHS) [48], IEEE Comprehensive Heart Disease Dataset (IEEECHD) [49], Cleveland [50], Hungarian [51], and Faisalabad Institute of Cardiology, Pakistan (FIC) [52]. These sources collectively represent a broad spectrum of demographic and clinical profiles across Massachusetts, Pakistan, Europe, and multiple locations within the United States. Understanding this heterogeneity is instrumental in mitigating potential model biases by ensuring the inclusion of diverse age groups, ethnicities, medical backgrounds, and lifestyle characteristics [53]. Building on insights from these real-world sources, this study uses a synthetically generated dataset to con- duct federated experiments. This approach was selected to ensure a strictly controlled, reproducible, and privacy- centric environment that avoids the legal complexities of sharing raw patient data. The synthetic dataset, comprising 30,000 samples, is designed to reflect the feature space and clinical context of the referenced public datasets, using the same set of clinically established cardiovascular risk factors. It includes six key features reflecting the defined clinical context: demographic data (age), clinical param- eters (systolic blood pressure, diastolic blood pressure, cholesterol), and binary lifestyle factors (smoking status, diabetes). The predictive target is a binary label indicating whether ‘Low Risk’ or ‘High Risk’ was assigned to each record based on rule-based logic derived from established clinical guidelines (Table 3). These labels are determinis- tic, generated by applying fixed clinical thresholds (e.g., systolic BP >140 mmHg, cholesterol >240 mg/dL, or presence of diabetes and smoking), and do not incorporate the probabilistic uncertainty inherent in real clinical diag- noses, where risk stratification typically results from cali- brated scoring models that output continuous probability estimates. This deterministic labeling ensures a controlled benchmarking environment; future work should incorpo- rate probabilistic labels to better assess generalizability to real-world risk assessments. All experiments were conducted under the Non-IID data distribution scenario, which is the primary focus of this study. Non-IID (Non-Independent and Identically Distrib- uted) Setting: A Non-IID environment was created by parti- tioning the data based on specific clinical and demographic characteristics to model a realistic deployment scenario. The process created "specialized" clients (e.g., a geriatric center, Fig. 2 Schematic of the Horizontal Federated Learning (HFL) parti- tioning scheme. The global clinical dataset is distributed across dis- tinct clients (e.g., hospitals), with each client sharing the same feature space (columns) but holding a disjoint subset of patient records (rows). This partition strategy simulates a realistic Non-IID multi-institutional environment in which the data structure is consistent but local distribu- tions vary 1 3 Page 11 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 a cardiology referral center), forcing the global model to learn from heterogeneous data sources, thereby testing the true robustness of the aggregation algorithm. Before train- ing, a standard preprocessing step was applied, in which all numerical features were normalized using StandardScaler [54]. To ensure consistency across the federated network, the scaler was fit on the entire conceptual training dataset before the data was partitioned and distributed to the clients. To ensure consistency across the FL model training process, only common features shared across datasets were selected, thereby preserving each dataset’s unique contextual relevance while enabling joint model optimi- zation. Each dataset posed unique challenges, including class imbalance, sparsity, missing values, and outliers. To navigate these challenges, particularly in the foundational stages of research, the study explicitly adopted synthetic data. Instead of using protected health information, we pro- grammatically generated a dataset that mimics the statisti- cal properties and key features of real-world patient records for cardiovascular risk assessment [55]. This approach was deliberately chosen as a robust, privacy-centric first step in model development. The advantages of utilizing synthetic data in this initial phase are manifold: 1. Inherent Privacy Preservation: By design, synthetic data contains no real patient information, thereby com- pletely circumventing privacy risks and eliminating the need for complex and time-consuming ethical board (e.g., CEP/IRB) approvals and data sharing agreements. Moreover, allows for rapid and agile development and testing of the FL framework. 2. Controlled and Reproducible Environment: The data generation process is fully controllable. We can precisely define the data distributions, feature relationships, and the degree of class imbalance. Furthermore, by using a fixed seed for the random number generator, we ensure the dataset is reproducible. This cornerstone of scientific validation allows other researchers to replicate the exper- iment under identical conditions. 3. Flexibility for Model Testing: A synthetic dataset can be tailored to test specific model capabilities. In case we created a known class imbalance to evaluate how the federated model performs under conditions common in real clinical data, but which may be inconsistent across different real-world datasets. A model trained on synthetic data must ultimately be fine- tuned and validated on real-world clinical data before any potential deployment [56]. To evaluate the robustness and real-world applicability of the proposed FedCVR frame- work, we designed a more challenging benchmark using a Non-Independently and Identically Distributed (non-IID) data partitioning strategy (See Table 4). Unlike standard datasets, this approach moves beyond the assumption that each client holds a perfectly representative sample of the data. Therefore, the work serves as a crucial proof-of-concept, demonstrating that the proposed FL architecture is functional and effective, thus paving the way for future studies involv- ing real-world patient data. A comprehensive overview of the dataset’s attributes and corresponding descriptions is provided in Table 5, offering a detailed understanding of the data utilized in the analysis. To simulate, we programmati- cally distributed the synthetic patient records among the five clients based on specific clinical and demographic charac- teristics. Each client, therefore, received a biased dataset representing a “specialized” hospital with a unique patient profile. Hence, this non-IID setup creates a more realistic and demanding test environment, as the server-side aggre- gation algorithm must effectively reconcile learning across diverse, potentially conflicting data distributions to produce a single, well-generalized global model. The model’s abil- ity to converge under these conditions strongly indicates its potential for practical deployment. 7 Implementation and experimental methodology To validate the theoretical advantages of the FedCVR frame- work, we developed a robust experimental environment in Python. The orchestration of server-client communica- tion rounds, data distribution, and secure model aggrega- tion was implemented using the Flower ecosystem [57]. Table 3 Distribution of patients by risk category in the synthetic data- set Risk category N ◦ of patients Percentage 0 (Low Risk)24,87082.9% 1 (High Risk)51317.1% Total30,000100.0% Table 4 Non-IID data partitioning strategy simulating specialized hos- pitals Client IDSimulated specialtyPrimary data characteristic N ◦ of patients Hospital 1General clinic (Young population) Patients with age < 50 years 13,130 Hospital 2Cardiology referral center Systolic BP > 135 and Cholesterol > 220 4,730 Hospital 3Diabetes and smok- ing control clinic Patients who are dia- betic or smokers 9,250 Hospital 4Community hospitalGeneral population not fitting other criteria 1,620 Hospital 5Geriatric centerPatients with age > 65 years 1,270 Total30,000 1 3 105 Page 12 of 23Peer-to-Peer Networking and Applications (2026) 19:105 The underlying Deep Neural Network (DNN) architectures were constructed using PyTorch [58], while Differential Pri- vacy (DP) guarantees, specifically the gradient clipping and noise injection mechanisms detailed in Algorithm 2, were enforced via the Opacus library [36]. Data preprocessing and performance evaluation relied on Scikit-learn [59] and NumPy [60]. The computational experiments operate on patient records characterized by key clinical features, including age, systolic and diastolic blood pressure, cholesterol levels, smoking, and diabetes status. This setup enables a detailed comparative analysis between the proposed FedCVR and state-of-the-art baselines. 7.1 Evaluation strategy and metrics Model performance was evaluated using standard classi- fication metrics: Accuracy, Precision, Recall (Sensitivity), Loss, F1-score, and Area Under the ROC Curve (AUC). A particular emphasis was placed on the performance metrics of the minority class (high-risk patients), as this group is frequently underrepresented in clinical datasets but holds significant importance for intervention strategies. The experimental design follows a rigorous compara- tive methodology, where the FedCVR framework is bench- marked against distinct categories of aggregation strategies under identical privacy constraints ( ε≈13.4 ): ●Stateless Baselines: Standard FedAvg and FedProx, which do not maintain server-side state. ●Adaptive Baselines: FedAdagrad and FedYogi, which, like FedCVR, use adaptive optimization but differ in their moment estimation logic. ●Clustering Approaches: FedCluster, used to assess performance when clients are grouped by data similarity. To ensure a fair comparison, each strategy is evaluated over N=5 independent runs with a fixed random_state seed for data partitioning and model initialization, such that all methods are evaluated on identical data splits. Algo- rithm 3 formalizes the comparative procedure used to iso- late the contribution of DP on model utility. 7.2 Hyperparameter selection and experimental justification The experimental design was structured to conduct rigor- ous comparative benchmarking of the proposed FedCVR framework against state-of-the-art aggregation strategies (FedAvg, FedProx, FedAdagrad, FedYogi, and FedCluster) under varying privacy constraints. To ensure a fair assess- ment of the server-side optimization effects, a core set of hyperparameters was held constant across all analysis runs. The key hyperparameters for the federated environment were selected as follows: ●Network Topology: The number of participating cli- ents was fixed at num_clients = 5. This simulates a consortium of major clinical institutions where data is siloed but collaboratively trained. ●Training Duration: The analysis was run for commu- nication_rounds = 100. This duration provides sufficient iterations for the adaptive moment estimation vectors ( m t ,v t ) in FedCVR and baseline adaptive opti- mizers to stabilize and demonstrate long-term conver- gence behavior. ●Local Computation: Each client performed local_ epochs = 5 with a learning rate of η c =0.01 . These values were chosen to balance meaningful local feature extraction with the frequency of global aggregation, mitigating the risk of excessive client drift in the Non- IID setting. ●Reproducibility: A random_state = 42 was used throughout all stochastic processes, including synthetic data generation, partitioning, and model initialization, en- suring that performance differences are attributable solely to the aggregation strategy and not initialization variance. The experiment manipulates two primary independent vari- ables: (1) the Aggregation Strategy and (2) the Privacy Budget. For the latter, the DP-enabled simulations utilized the Opacus library with noise multipliers σ∈0.5,1.1,1.5 , representing a spectrum from moderate to strict privacy protection. By holding all other hyperparameters constant (as detailed in Table 6), this setup isolates the server-side adaptive optimization’s specific capability to recover utility from noisy updates. Table 5 Properties of the synthetic dataset used for the analysis PropertyDescription Dataset nameSynthetic health risk dataset Data typesInteger (int), Floating-Point (float) Dataset sourceProgrammatically generated via Python script using NumPy. The data is artificial and does not contain any real patient information. ContextDesigned to simulate a binary classification task for cardiovascular risk prediction within both IID and non-IID FL environments. Total rows and columns 30,000 rows × 7 columns. The number of records is required to ensure robust data availability for each specialized client in the non-IID scenario. Attributes (Features) age, systolic_bp, diastolic_bp, cho- lesterol, smoker (0/1), diabetic (0/1) Target labelA binary variable indicating risk (0: Low Risk, 1: High Risk). 1 3 Page 13 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 7.3 Sensitivity of τ under High-Privacy Noise Regimes The stability constant τ plays a particularly important role under high-privacy conditions. When σ is large (e.g., σ=1.5 , ε≈6.6 ), the injected Gaussian noise dominates the gradient signal, causing the estimated second moment ˆv t to approach zero in parameter dimensions where the true gradient is small. In this regime, the adaptive update denominator ( √ ˆv t +τ) is governed almost entirely by τ. A value too small ( τ<10 −4 ) would produce numerically explosive updates, while a value too large ( τ>10 −2 ) would suppress adaptivity, effectively reducing Fed- CVR to standard gradient descent. The selected value τ=10 −3 follows the Adam optimizer convention [?] and was empirically confirmed to yield stable convergence across all tested privacy regimes in our experiments, with AUC variation below 0.01 when τ was varied in the range [10 −4 ,10 −2 ] . 8 Experimentation and results This section presents a comprehensive evaluation of the FedCVR framework. We begin by characterizing the com- putational environment and the statistical properties of the clinical dataset, highlighting the heterogeneity and outliers that pose challenges to FL. Subsequently, we detail the con- vergence behavior of the global model, analyze the impact of differential privacy on utility, and conclude with a rigor- ous statistical benchmark against state-of-the-art baselines. 8.1 Experimental environment specification To ensure reproducibility and transparency, we first detail the computational infrastructure used for this study. The experiments were conducted on a standardized Linux-based server environment, specifically selected to demonstrate the framework’s efficiency without relying on specialized hard- ware acceleration, such as high-end GPUs. Table 7 outlines the specific hardware and software com- ponents. The use of standard CPU resources underscores the lightweight nature of the FedCVR aggregation algorithm, making it suitable for deployment in resource-constrained clinical networks. 8.2 Data characterization and complexity analysis Before evaluating the predictive models, it is essential to understand the statistical characteristics of the underlying clinical data. The dataset comprises features with distinct distributional patterns that directly influence the learning difficulty. Figure 3 presents the frequency distributions of the key numerical features used in the study. As observed, the Age feature follows an approximately uniform distribution, ensuring balanced representation across different demographics. In contrast, the clinical indicators, Systolic BP, Diastolic BP, and Cholesterol, exhibit Bell- shaped (Gaussian) distributions. This statistical heterogeneity confirms that the dataset accurately reflects the physiological variations commonly found in the general population. Furthermore, the presence of outliers and data dispersion is a critical factor for robust model training. Figure 4 visual- izes the central tendency and variance of these features. Table 6 Hyperparameter settings for the federated learning experi- ments HyperparameterDescriptionValue General federated learning parameters num_clientsNumber of participating clients (hospitals) in the analysis. 5 communication_roundsTotal number of learn- ing rounds for the main experiments. 100 random_stateSeed used for all random processes to ensure reproducibility. 42 Client-side training parameters local_epochsNumber of training epochs each client per- forms locally per round. 5 client_learning_rate ( η c ) Learning rate for the cli- ent-side SGD optimizer. 0.01 batch_sizeBatch size for local train- ing (optimized for Opacus DP accounting). 32 Server-side adaptive optimizer parameters (FedCVR, FedYogi, FedAdagrad) server_learning_rate ( η s ) Global learning rate applied to the aggregated update. 0.1 beta_1 ( β 1 ) Exponential decay rate for the first moment estimates ( m t ). 0.9 beta_2 ( β 2 ) Exponential decay rate for the second moment estimates ( v t ). 0.999 tau ( τ )Numerical stability con- stant (adaptivity epsilon). 10 −3 Table 7 Analysis environment specifications ComponentSpecification CPUIntel Xeon (4 vCores) RAM16 GiB OSUbuntu 22.04 LTS LanguagePython 3.10 LibrariesFlower 1.5, PyTorch 2.0, Opacus 1.3, Scikit-learn 1 3 105 Page 14 of 23Peer-to-Peer Networking and Applications (2026) 19:105 The boxplots reveal significant variance, particularly in Cholesterol levels (centered around 225 mg/dL) and Sys- tolic Blood Pressure. The outliers beyond the whiskers rep- resent high-risk patients, which are crucial for the model to identify correctly. Including these extreme values ensures that the trained FedCVR model is robust not only to typi- cal cases but also to critical anomalies often associated with cardiovascular events. 8.3 Convergence and global model performance The core of our evaluation focuses on the learning dynam- ics of the FedCVR aggregator under Non-IID data distri- butions. By monitoring the global model’s metrics at each communication round, we can assess the efficiency of the adaptive optimization strategy. Figure 5 illustrates the evolution of key performance indicators over 100 communication rounds for the FedCVR non-private baseline (no DP, ε=∞ ). The results demon- strate a highly efficient optimization profile. The Average Loss (dotted gray line) drops sharply from an initial 0.50 to a stable plateau of approximately 20% within the first 60 rounds. Concurrently, the F1-Score improves dramati- cally, stabilizing at 84%, the utility ceiling achievable with- out privacy constraints. This result is particularly significant given the class imbalance in the dataset; the high F1-score confirms that the model maintains a robust balance between Precision (79%) and Recall (78%), rather than merely pre- dicting the majority class. Under the operational DP budget ( ε≈13.4 , σ=1.0 ), the F1-Score settles at 78%, a cost- of-privacy of approximately 6 percentage points, which the FedCVR adaptive aggregator recovers relative to the 72% achieved by FedAvg under the same constraints (Table 8). To further validate the benefits of the federated approach, Fig. 6 compares the performance of the aggregated global model against individual client models trained on their local data partitions. The contrast is striking. Individual clients (Clients 1-5) exhibit volatile learning curves due to local data bias (Non-IID). Some clients converge quickly but to subopti- mal minima, while others struggle with instability. In com- parison, the aggregated Global Model (bottom-right panel) demonstrates a smooth, monotonic convergence trajec- tory. This confirms that the FedCVR strategy effectively synthesizes diverse local knowledge into a generalizable global model, mitigating the bias inherent in any single institution’s data. 8.4 Comparative benchmarking: FedCVR vs. state- of-the-art To quantify the specific engineering advantages of our framework, we conducted a rigorous benchmark against standard baselines (FedAvg, FedProx) and advanced adap- tive optimizers (FedAdagrad, FedYogi). All algorithms were evaluated under identical Differential Privacy constraints ( ε≈13.4 ) to simulate a realistic, privacy-preserving clini- cal environment. Figure 7 presents the Receiver Operating Characteristic (ROC) curves for all evaluated strategies. This metric is cru- cial for clinical decision support systems as it measures the trade-off between sensitivity and specificity. The FedCVR framework (solid green line) achieves the dominant performance with an Area Under the Curve (AUC) of 0.96. The sharper "elbow" of the curve indicates superior sensitivity at low false-positive rates, a critical requirement Fig. 3 Frequency distribution of numerical features. Histograms with Kernel Density Estimation (KDE) overlays showing the statistical distribution of Age, Systolic BP, Diastolic BP, and Cholesterol in the patient cohort. The distributions indicate a representative sample suit- able for unbiased model training 1 3 Page 15 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 for clinical screening tools where minimizing false alarms is as important as detecting positive cases. For a granular statistical validation, Table 8 details the mean performance metrics across N=5 independent runs. The results show that FedCVR outperforms the stan- dard FedAvg baseline with high statistical significance ( p<0.001 ), confirming that stateless aggregation is insuf- ficient for noisy, private FL. Furthermore, FedCVR dem- onstrates a statistically significant improvement ( p<0.05 ) over FedYogi. These findings suggest that the specific momentum tun- ing in FedCVR acts as a superior “temporal denoiser,” effec- tively averaging out the random Gaussian noise added by the DP mechanism while preserving the true gradient direction. 8.5 Privacy-utility trade-off analysis Finally, we analyze the resilience of the FedCVR frame- work to varying levels of privacy noise. A robust FL system must maintain clinical utility even when strict privacy guar- antees are enforced. Figure 8 illustrates the learning dynamics under privacy constraints. The comparative analysis reveals two distinct Fig. 5 Evolution of global model performance metrics using Fed- CVR. The chart tracks key classification metrics (Accuracy, Preci- sion, Recall, F1-Score) and Training Loss over 100 communication rounds. The rapid convergence of Loss (dotted gray) and stabilization of F1-Score (solid purple) demonstrate the efficiency of the adaptive optimization strategy Fig. 4 Statistical distribution of numerical features. Box plots sum- marizing the central tendency and dispersion of Age, Blood Pressure (Systolic/Diastolic), and Cholesterol. The independent scales reveal the specific variance and outlier profile of each clinical marker in the synthetic dataset 1 3 105 Page 16 of 23Peer-to-Peer Networking and Applications (2026) 19:105 phases. Initially (Rounds 0-30), the DP-enabled model (dashed line) experiences instability and higher loss due to the injected noise. However, in the second phase, the adap- tive optimizer effectively corrects this trajectory, enabling the model to converge to a utility level nearly identical to that of the non-private baseline. Table 9 quantifies this trade-off across three privacy regimes. Even under a strict privacy regime (“High Privacy”, σ=1.5 , ε≈6.6 ), the model maintains a functional F1-Score of 0.65. While there is an expected drop in per- formance compared to the non-private baseline (F1-Score = 0.84, ε=∞) and the operational regime (F1-Score = 0.78, ε≈13.4 ), this "graceful degradation" confirms that Fed- CVR remains a viable candidate for real-world deployment where regulatory compliance (such as GDPR or HIPAA) Table 8 Comprehensive benchmark under differential privacy ( ε≈13.4 ). Mean ± std. dev. across N=5 independent runs. Statistical significance was assessed via a two-tailed t-test comparing each method against FedCVR MethodTypeAccuracyRecallF1-ScoreAUC p-value (vs. FedCVR) FedAvgBaseline0.85 ± 0.020.65 ± 0.040.72 ± 0.030.88 ± 0.02 <0.001 ∗ FedProxRegularization0.87 ± 0.010.68 ± 0.020.74 ± 0.020.90 ± 0.01 <0.001 ∗ FedClusterClustering0.88 ± 0.020.70 ± 0.030.75 ± 0.030.91 ± 0.02 <0.001 ∗ FedAdagradAdaptive0.89 ± 0.020.72 ± 0.030.76 ± 0.020.92 ± 0.01 <0.001 ∗ FedYogiAdaptive0.91 ± 0.010.76 ± 0.020.77 ± 0.020.94 ± 0.01 0.014 ∗ FedCVR (Ours)Adaptive0.92 ± 0.010.78 ± 0.020.78 ± 0.020.96 ± 0.01— Significance levels: ∗ p<0.05 , ∗ p<0.001 . Note: with N=5 runs, statistical power is limited; the reported significance levels should be interpreted alongside the effect sizes and confidence intervals provided above Fig. 6 Performance analysis of individual client models vs. the feder- ated average. The grid displays the learning trajectories for each client on their local Non-IID data partitions. The bottom-right panel (Fed- erated Average) illustrates the aggregated global model, which con- verges more smoothly and is more stable than the volatile individual client updates 1 3 Page 17 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 is non-negotiable. The framework enables healthcare net- works to adjust the σ parameter to achieve the optimal bal- ance between legal compliance and diagnostic accuracy. 8.6 Privacy budget analysis and limitations Regarding the differential privacy parameters, we achieved a cumulative privacy budget of ε≈13.4 with a fixed δ=10 −5 . This value satisfies the standard requirement δ≪1/|D k | with respect to each client’s local dataset ( |D k |≈6,000 , giving 1/|D k |≈1.67×10 −4 ), which is the relevant privacy unit for the DP-SGD mechanism applied locally at each client. While theoretical DP literature often advocates for ε<1 for strong privacy guarantees [42], practical deployments of FL in healthcare face a challenging trade-off between pri- vacy and diagnostic utility. An extremely low ε would intro- duce excessive noise, potentially degrading the model’s sensitivity to critical cardiovascular risks. Therefore, the reported ε≈13.4 represents a utility-prioritized configura- tion, ensuring the model remains clinically relevant while offering protection against reconstruction attacks, compa- rable to industrial deployments of DP-FL. Future iterations of this work will explore techniques to tighten the privacy budget without sacrificing utility, includ- ing adaptive noise scheduling, personalized DP (where different clients receive different ε allocations based on their data sensitivity), and the integration of Secure Aggregation (SecAgg) protocols to reduce the required noise magnitude. 8.7 Ablation study: isolating the contribution of momentum To isolate the specific contribution of the first moment (momentum) component in FedCVR’s adaptive aggrega- tion, we conducted an ablation study comparing FedCVR (with β 1 =0.9 , full Adam-based update) against a momen- tum-free variant (FedCVR-NoMom, with β 1 =0 , equiva- lent to a server-side RMSProp-like update), both evaluated under the operational privacy budget ( ε≈13.4 ). All other hyperparameters were held constant. Table 10 presents the results. Removing the momentum component causes a statis- tically significant drop in Recall ( −7 p, p<0.01 ) and AUC ( −0.03 ), while FedCVR-NoMom remains competi- tive with FedAdagrad (Table 8). This confirms that the temporal denoising mechanism is primarily driven by the first-moment exponential moving average, not merely by adaptive scaling of the second moment. Without the his- torical gradient averaging provided by β 1 , the server-side update cannot effectively cancel the round-to-round Gauss- ian noise introduced by DP-SGD. Fig. 7 Comparative Receiver Oper- ating Characteristic (ROC) analy- sis. FedCVR (Green) demonstrates superior discriminative capability compared to baselines, achieving the highest Area Under the Curve (AUC = 0.96) and maintaining bet- ter Sensitivity at low False Positive Rates 1 3 105 Page 18 of 23Peer-to-Peer Networking and Applications (2026) 19:105 8.8 Minority class recall under differential privacy With a class imbalance of approximately 17.1% (high-risk patients) in the dataset, preserving Recall for the minority class under DP is a critical clinical requirement. Missed high-risk patients represent false negatives with poten- tially severe consequences for patient outcomes. Table 9 demonstrates that FedCVR maintains meaningful Recall across all tested privacy regimes (0.78 at ε≈13.4 ; 0.65 at ε≈6.6 ), outperforming FedAvg’s Recall of 0.65 even under a less strict privacy budget. This robustness can be attributed to the temporal denoising property of the first moment: the gradient direction associated with the minority class is consistent across rounds (high-risk patients share distinctive fea- ture patterns across the Non-IID partitions), whereas the injected Gaussian noise is zero-mean and round-indepen- dent. As a result, the momentum vector m t accumulates the minority-class gradient signal while the noise com- ponent decays to zero in expectation. This mechanism is particularly effective under class imbalance, where the minority signal, although smaller in magnitude, is more consistent than the majority-class gradient in Non-IID settings. Notably, even under the strictest tested regime ( ε≈6.6 ), the framework identifies 65% of high-risk patients, a level that compares favorably to FedAvg’s performance under significantly more relaxed privacy constraints. Table 9 Impact of DP noise on FedCVR utility (Mean F1-Score) Noise multiplier ( σ )Privacy budget ( ε )F1-score None (Baseline) ∞ 0.84 ± 0.01 0.5 (Low)80.50.81 ± 0.02 1.0 (Moderate)13.40.78 ± 0.03 1.5 (High)6.60.65 ± 0.05 Table 10 Ablation study: effect of server-side momentum on FedCVR under DP ( ε≈13.4 ). Mean ± std. dev. across N=5 runs VariantRecallF1-scoreAUC FedCVR-NoMom ( β 1 =0 ) 0.71 ± 0.030.74 ± 0.020.93 ± 0.01 FedCVR ( β 1 =0.9 ) 0.78 ± 0.020.78 ± 0.020.96 ± 0.01 Fig. 8 Impact of differential pri- vacy on FedCVR model perfor- mance. The comparative analysis reveals two distinct phases: (1) an initial instability period (Rounds 0-30) caused by DP noise injec- tion, followed by (2) a robust con- vergence phase where the FedCVR adaptive optimizer effectively filters the noise 1 3 Page 19 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 9 Discussion This study presents the design and validation of a robust framework for Federated Learning (FL) in cardiovascu- lar risk prediction, operating under stringent privacy con- straints. While initial baselines using the canonical FedAvg and FedProx algorithms provided functional aggregation under standard conditions, they exhibited significant perfor- mance degradation when subjected to the dual challenges of Non-IID data and Differential Privacy (DP) noise. The primary contribution of this work lies in the architectural validation of FedCVR, which not only outperforms state- less aggregators but also demonstrates statistical superior- ity over state-of-the-art adaptive strategies, such as FedYogi and FedAdagrad. 9.1 Architectural analysis: momentum as a privacy denoiser The superior performance of FedCVR over stateless base- lines (FedAvg) and even standard adaptive configurations (FedYogi) is not merely a result of hyperparameter tun- ing, but rather a structural consequence of the interaction between server-side momentum and the Differential Privacy mechanism. In a DP-FL system, every client update ∆w k contains two components: the true gradient signal ∇L and the injected Gaussian noise N(0,σ 2 ) . Stateless aggregators, such as FedAvg, compute the arithmetic mean of these updates at each round t. Consequently, the global model w t+1 directly inherits the variance of the noise injected at round t, leading to the erratic convergence observed in our baselines. In contrast, the adaptive aggregation logic employed in FedCVR introduces a temporal filtering effect. By main- taining a history of updates via the first moment vector m t : m t =β 1 m t−1 +(1−β 1 )(g signal +g noise ) (13) The system effectively averages out the zero-mean Gauss- ian noise over multiple rounds ( E[g noise ]→0 ), acting as a temporal denoiser that allows the persistent clinical signal to accumulate while damping stochastic noise. This mechanism also explains the performance advantage over other adaptive methods. While FedCluster attempts to mitigate heterogeneity by grouping clients, it cannot filter against temporal gradient noise. Furthermore, compared to FedYogi, which employs a more aggressive additive update rule for the second moment, FedCVR’s conserva- tive moment estimation proved more robust to the "gradient explosion" often caused by the clipping operations inherent to DP-SGD. 9.2 Revisiting the privacy-utility trade-off A significant finding of this research is the re-evaluation of the “cost of privacy.” Common literature suggests that imple- menting strong privacy guarantees ( ε<10 ) incurs a prohibi- tive performance cost. Our results provide a more nuanced picture. While the analysis revealed predictable instability in the initial training rounds (0–30), the FedCVR aggregator demonstrated a remarkable capacity to recover utility as its moment vectors accumulated sufficient gradient history. Under the operational regime ( σ=1.0 , ε≈13.4 ), Fed- CVR converges to an F1-Score of 0.78, approximately 6 percentage points below the non-private ceiling of 0.84. Crit- ically, this is a differential cost: FedAvg under the same bud- get achieves only F1 = 0.72 ± 0.03. The adaptive optimizer therefore recovers approximately two-thirds of the privacy- induced degradation relative to stateless aggregation. Even under high-privacy regimes ( σ=1.5 , ε≈6.6 ), the degra- dation is graceful (F1 = 0.65) rather than catastrophic. This implies that the trade-off is not a fixed barrier, but rather an engineering problem that can be partially resolved through stateful aggregation. By employing an optimizer that inher- ently dampens noise over time, we demonstrate that formal privacy guarantees do not require abandoning clinical utility. 9.3 Computational and communication overhead From an engineering deployment perspective, the overhead introduced by FedCVR is minimal. While the server per- forms additional operations to calculate moments m t and v t , this computational cost is negligible compared to the net- work latency of transmitting model updates over the internet. Crucially, the communication cost remains identical to FedAvg, as the vector sizes transmitted between the cli- ent and the server remain unchanged; the moment vectors are stored and updated solely on the server. The memory requirement on the server increases linearly by a factor of 3 (storing w t ,m t ,v t ), which is trivial for modern server infrastructure, given the relatively small parameter space of tabular data models used in clinical risk scores. 9.4 Limitations While this study demonstrates the engineering feasibility of the FedCVR framework, we acknowledge specific limita- tions inherent to its current validation scope: ●Synthetic Data: We utilized synthesized data to pre- cisely control the degree of statistical heterogeneity (Non-IID) and isolate the optimizer’s behavior. While this allows for rigorous ablation studies, it does not fully capture the missing values, heterogeneous formats, and 1 3 105 Page 20 of 23Peer-to-Peer Networking and Applications (2026) 19:105 unstructured noise typical of raw Electronic Health Re- cords (EHRs). Alternative strategies, such as applying reweighted focal loss directly on real imbalanced da- tasets [61], could complement this approach by avoid- ing bias from synthetic label generation entirely. Future validation on real-world multi-center datasets (e.g., eICU or MIMIC-IV) is required to confirm clinical generalizability. ●Threat Model Scope: The current framework assumes a passive, honest-but-curious threat model: the server faithfully executes the aggregation protocol but may at- tempt to infer information from individual (noisy) client updates. Active adversarial scenarios, in which a mali- cious server or compromised clients attempt to manipu- late training via model poisoning, gradient inversion, or Byzantine attacks, are outside the scope of this architec- tural validation. Furthermore, the framework does not implement Secure Multi-Party Computation (SMPC) or Secure Aggregation (SecAgg) [62], which would cryp- tographically hide individual updates from the server. Integrating these defense layers is a critical next step toward deployment in fully adversarial environments. ●Architectural Scalability: The experiments employ a lightweight DNN on six structured clinical features with five participating clients. As the number of clients N increases, the gradient aggregation benefits from re- duced variance (averaging over more estimates), which suggests that η s could potentially be increased with- out compromising stability. However, the framework’s behavior under N≫5 , with high-dimensional inputs (e.g., 12-lead ECG waveforms or echocardiographic images requiring convolutional or transformer-based local models), and with the associated increase in per- parameter clipping sensitivity has not been empirically validated and constitutes a prioritized future direction. ●Edge Deployment Overhead: FedCVR was designed and validated for institutional-grade hospital servers with sufficient resources for Opacus-based per-sample gradient computation (which scales linearly with batch size). Deployment on resource-constrained edge devices (e.g., wearable sensors or embedded medical devices) would require architectural adaptations, including mod- el compression, gradient sparsification, and communi- cation-efficient protocols, as the DP-SGD overhead and 100-round communication frequency could introduce unacceptable latency on constrained network links. ●Interoperability Assumptions: The experiments as- sume a unified feature schema across all clients. A real- world deployment would require a semantic interop- erability layer (e.g., HL7 FHIR) as a prerequisite for federated training. 10 Conclusion and future directions In this work, we addressed the critical challenge of deploying robust Machine Learning models for cardiovascular risk pre- diction within a privacy-constrained, multi-institutional envi- ronment. We introduced and validated FedCVR, a framework that mitigates the adverse effects of Non-IID data distribu- tions and Differential Privacy noise via server-side adaptive optimization. The following sections summarize our primary findings and outline strategic avenues for future research. 10.1 Conclusion This study systematically evaluated FedCVR, a framework that applies an Adam-based adaptive update rule at the server level and is thus functionally equivalent to FedAdam [32], against a comprehensive suite of baselines including stateless aggre- gators (FedAvg and FedProx) and other adaptive optimizers (FedYogi and FedAdagrad). Building on the baseline FL pipe- line validated in our prior work [8], this study demonstrates that server-side adaptivity is a structural prerequisite for recov- ering clinical utility under DP constraints, a question that the prior work explicitly identified as the primary open challenge. Our extensive empirical analysis on heterogeneous syn- thetic clinical data shows that FedCVR achieves a final AUC of 0.96 and an F1-Score of 0.78 under the operational privacy budget ( ε≈13.4 ), compared to a non-private ceil- ing of F1 = 0.84 (cost-of-privacy: −6 p) and FedAvg’s F1 ≈ 0.72 under the same DP constraints (recovered margin: +6 p). These results confirm that server-side momentum acts as an effective temporal denoiser, progressively cancel- ing zero-mean Gaussian privacy noise across communica- tion rounds while preserving the clinical signal. For medical applications where data is statistically heterogeneous, and privacy is non-negotiable, simple aggregation is insuffi- cient; server-side adaptivity is a prerequisite for stabilizing convergence in privacy-preserving federated AI. 10.2 Future directions Building upon these findings, future research will expand the scope of this framework in five key directions: ●Validation on Real-World Multi-Modal Data: While this study validated the framework on structured clinical features, future work will transition to large-scale, real- world datasets (e.g., MIMIC-IV or eICU). Furthermore, we aim to incorporate multimodal data, combining tabu- lar EHR data with unstructured clinical notes (NLP) or time-series signals (such as ECG) to enhance predictive accuracy. 1 3 Page 21 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 ●Enhanced Security Layers: The current privacy mod- el focuses on Differential Privacy to prevent inference attacks. Future iterations will integrate cryptographic techniques, such as Secure Multi-Party Computation (SMPC) or Homomorphic Encryption, to protect the model updates themselves during transmission. Addi- tionally, we will explore defense mechanisms against Byzantine failures and model poisoning attacks. ●Personalized Federated Learning: While FedCVR optimizes a single global model, clinical heterogeneity often requires localization. We plan to investigate Per- sonalized Federated Learning (pFL) techniques, such as fine-tuning the global model on local client data or learning client-specific headers, to further improve per- formance for specific patient demographics. ●Probabilistic Label Generation and Real-Data Vali- dation: Future experiments will incorporate probabilis- tic risk labeling (e.g., using the Framingham Risk Score as a continuous generation mechanism) and validate the framework on publicly available real-world multi-center datasets, comparing performance against the synthetic benchmarks established here. ●Asynchronous Client Updates: The current FedCVR architecture assumes synchronous client updates, where the server waits for all selected clients to re- turn gradients before performing the aggregation step. In real hospital networks, institutional availability is inconsistent. Extending FedCVR to asynchronous settings would require buffered aggregation mecha- nisms in which the server triggers aggregation upon receiving a threshold fraction of client updates. The temporal denoising property of the momentum vec- tor would need to be re-evaluated in this context, as stale gradients from delayed clients could corrupt the first-moment estimate, requiring staleness-weighted aggregation strategies. Acknowledgements The author would like to acknowledge the sup- port of the Software Engineering and Automation Research Labora- tory, where the research was developed and conducted. The infrastruc- ture and resources provided were crucial for completing the work. Author Contributions R.T. conceptualized the study, developed the FedCVR framework, conducted the experiments, and wrote the main manuscript text. L.A. supervised the research, contributed to the ar- chitectural design, and critically revised the manuscript for important intellectual content. All authors reviewed the manuscript. Funding The Article Processing Charge (APC) for the publication of this research was funded by the Coordenação de Aperfeiçoamen- to de Pessoal de Nível Superior - Brasil (CAPES) (ROR identifier: 00x0ma614). This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data Availability No datasets were generated or analysed during the current study. Declarations Competing interests The authors declare no competing interests. Conflicts of Interest The authors declare that they have no known competing financial interests or personal relationships that could have influenced the work reported in this paper. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adap- tation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit h t t p : / / c r e a t i v e c o m m o n s . o r g / l i c e n s e s / b y / 4 . 0 / . References 1. Gupta A, Misra S, Pathak N, Das D (2023) FedCare: Federated learning for resource-constrained healthcare devices in IoMT sys- tem. IEEE Trans Comput Social Syst 10(4):1587–1596 2. Khimani A, Hornback A, Jain N, Avula P, Jaishankar A, Wang MD (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), p 1–5. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / E M B C 5 3 1 0 8 . 2 0 2 4 . 1 0 7 8 2 8 8 5 3. Lopez EO, Ballard BD, Jan A (2023) Cardiovascular disease. In: StatPearls [Internet], StatPearls Publishing 4. Mittal S, Gupta S, Bansal K, Aggarwal G (2024) Democratizing GDPR compliance: AI-driven privacy policy interpretation. In: Proceedings of the 2024 16th International Conference on Con- temporary Computing, IC3-2024, Association for Computing Machinery, New York, NY, USA, p 735–743. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 7 5 8 8 8 . 3 6 7 6 1 4 2 5. Jensen BK, Cline M, Guynes CS (2007) Hippa, privacy and orga- nizational change: a challenge for management. SIGCAS Com- put Soc 37(1):12–17. https:/ /doi.or g/10.11 45/12 73353.1273354 6. Shah SM, Khan RA (2020) Secondary use of electronic health record: Opportunities and challenges. IEEE Access 8:136947– 136965. https:/ /doi.or g/10.11 09/AC CESS.2020.3011099 7. McMahan HB, Moore E, Ramage D, y Arcas BA (2016) Fed- erated learning of deep networks using model averaging. arXiv:1602.05629 8. Tertulino R (2025) A robust pipeline for differentially private fed- erated learning on imbalanced clinical data using SMOTETomek and FedProx, submitted to the Journal of the Brazilian Computer Society. arXiv:2508.10017 9. Thumula K, Holla H, Gutti C, Sasikumar A, Gogineni H (2025) PrivFED: Protecting user privacy in federated learning systems through differential privacy. In: 2025 8th International Confer- ence on Electronics, Materials Engineering & Nano-Technology (IEMENTech), p 1–6. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / I E M E N T e c h 6 5 1 1 5 . 2 0 2 5 . 1 0 9 5 9 6 6 5 10. Ahmed S, Kaiser MS, Chaki S, Aloteibi S, Moni MA (2025) Federated learning model with dynamic scoring-based client selection for diabetes diagnosis. Knowl-Based Syst 320:113662. https:/ /doi.or g/10.10 16/j. knosys.2025.113662 1 3 105 Page 22 of 23Peer-to-Peer Networking and Applications (2026) 19:105 11. Silverio A, Cavallo P, De Rosa R, Galasso G (2019) Big health data and cardiovascular diseases: A challenge for research, an opportunity for clinical care. Frontiers in Medicine Volume 6 - 2019. https:/ /doi.or g/10.33 89/fm ed.2019.00036 12. Qu Z, Zhao X, Sun L, Muhammad G (2025) DAQFL: Dynamic aggregation quantum federated learning algorithm for intelligent diagnosis in internet of medical things. IEEE Internet Things J 1–1. https:/ /doi.or g/10.11 09/JI OT.2025.3537614 13. Qiu L, Cheng J, Gao H, Xiong W, Ren H (2023) Federated semi- supervised learning for medical image segmentation via pseudo- label denoising. IEEE J Biomed Health Inform 27(10):4672–4683 14. Ullah F, Srivastava G, Xiao H, Ullah S, Lin JC-W, Zhao Y (2023) A scalable federated learning approach for collaborative smart healthcare systems with intermittent clients using medical imag- ing. IEEE J Biomed Health Inf 15. Benitez K, Malin B (2010) Evaluating re-identification risks with respect to the HIPAA privacy rule. J Am Med Inform Assoc 17(2):169–177 16. Goto S, Solanki D, John JE, Yagi R, Homilius M, Ichihara G, Kat- sumata Y, Gaggin HK, Itabashi Y, MacRae CA, Deo RC (2022) Multinational federated learning approach to train ECG and echo- cardiogram models for hypertrophic cardiomyopathy detection. Circulation 146(10):755–769. h t t p s : / / d o i . o r g / 1 0 . 1 1 6 1 / C I R C U L A T I O N A H A . 1 2 1 . 0 5 8 6 9 6 17. Madathil NT, Dankar FK, Gergely M, Belkacem AN, Alrabaee S (2025) Revolutionizing healthcare data analytics with federated learning: A comprehensive survey of applications, systems, and future directions. Comput Struct Biotechnol J 28:217–238 18. Sheller MJ, Edwards B, Reina GA, Martin J, Pati S, Kotrotsou A, Milchenko M, Xu W, Marcus D, Colen R et al (2020) Federated learning in medicine: facilitating multi-institutional collabora- tions without sharing patient data. Sci Rep 10(1):12598 19. Pati S, Baid U, Edwards B, Sheller M, Wang S-H, Reina GA, Foley P, Gruzdev A, Karkada D, Davatzikos C et al (2022) Feder- ated learning enables big data for rare cancer boundary detection. Nat Commun 13(1):7346 20. Linardos A, Kushibar K, Walsh S, Gkontra P, Lekadir K (2022) Federated learning for multi-center imaging diagnostics: a simu- lation study in cardiovascular disease. Sci Rep 12(1):3551 21. Dhavamani M, Niranjan K (2022) A federated learning based approach for heart disease prediction. In: 2022 6th International Conference on Computing Methodologies and Communication (ICCMC), p 1117–1121. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / I C C M C 5 3 4 7 0 . 2 0 2 2 . 9 7 5 4 1 1 9 22. Wang J, Gounley J, Hanson H (2025) CAFE AU LAIT: Compute- aware federated augmented low-rank AI training. In: Proceedings of the Platform for Advanced Scientific Computing Conference, PASC ’25, Association for Computing Machinery, New York, NY, USA, p 1–12. https:/ /doi.or g/10.11 45/37 32775.3733580 23. Maliakel PJ, Ilager S, Brandic I (2024) FLIGAN: Enhancing feder- ated learning with incomplete data using GAN. In: Proceedings of the 7th International Workshop on Edge Systems, Analytics and Net- working, EdgeSys ’24, Association for Computing Machinery, New York, NY, USA, p 1–6. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 4 2 9 6 8 . 3 6 5 4 8 1 3 24. Dhasade A, Ding Y, Guo S, Kermarrec A-M, de Vos M, Wu L (2024) QuickDrop: Efficient federated unlearning via synthetic data generation. In: Proceedings of the 25th International Middle- ware Conference, Middleware ’24, Association for Computing Machinery, New York, NY, USA, p 266–278. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 5 2 8 9 2 . 3 7 0 0 7 6 4 25. Maddock S, Cormode G, Maple C (2024) FLAIM: AIM-based synthetic data generation in the federated setting. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, Association for Computing Machin- ery, New York, NY, USA, p 2165–2176. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 3 7 5 2 8 . 3 6 7 1 9 9 0 26. Ghavamipour AR, Turkmen F, Wang R, Liang K (2023) Feder- ated synthetic data generation with stronger security guarantees. In: Proceedings of the 28th ACM Symposium on Access Control Models and Technologies, SACMAT ’23, Association for Com- puting Machinery, New York, NY, USA, p 31–42. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 5 8 9 6 0 8 . 3 5 9 3 8 3 5 27. Rieke N, Hancox J, Li W, Milletari F, Roth HR, Albarqouni S, Bakas S, Galtier MN, Landman BA, Maier-Hein K et al (2020) The future of digital health with federated learning. NPJ digital medicine 3(1):119 28. An T, Ma L, Wang W, Yang Y, Wang J, Chen Y (2023) Consider- ation of FedProx in privacy protection. Electronics 12(20). https:/ /doi.or g/10.33 90/el ectronics12204364, h t t p s : / / w w w . m d p i . c o m / 2 0 7 9 - 9 2 9 2 / 1 2 / 2 0 / 4 3 6 4 29. Yurdem B, Kuzlu M, Gullu MK, Catak FO, Tabassum M (2024) Federated learning: Overview, strategies, applications, tools and future directions. Heliyon 10(19):e38137. h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . h e l i y o n . 2 0 2 4 . e 3 8 1 3 7 30. Li T, Sahu AK, Zaheer M, Sanjabi M, Talwalkar A, Smith V (2020) Federated optimization in heterogeneous networks. In: Dhillon I, Papailiopoulos D, Sze V (eds), Proceedings of Machine Learning and Systems, vol 2, p 429–450. h t t p s : / / p r o c e e d i n g s . m l s y s . o r g / p a p e r _ fi l e s / p a p e r / 2 0 2 0 / fi l e / 1 f 5 f e 8 3 9 9 8 a 0 9 3 9 6 e b e 6 4 7 7 d 9 4 7 5 b a 0 c - P a p e r . p d f 31. C. You, Z. Lu, J. Wang, C. Yan, FedDDB: Clustered federated learning based on data distribution difference. In: Proceedings of the 2022 5th International Conference on Algorithms, Computing and Artificial Intelligence, ACAI ’22, Association for Computing Machinery, New York, NY, USA, 2023. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 5 7 9 6 5 4 . 3 5 7 9 7 3 2 32. Reddi S, Charles Z, Zaheer M, Garrett Z, Rush K, Konečný J, Kumar S, McMahan HB (2021) Adaptive federated optimization. arXiv:2003.00295 33. Luo Y, Pan Z, Fu Q, Qin S (2024) FAdagrad: Adaptive federated learn- ing with differential privacy. In: 2024 IEEE International Conference on High Performance Computing and Communications (HPCC), p 508–515. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / H P C C 6 4 2 7 4 . 2 0 2 4 . 0 0 0 7 4 34. Xing X, Liu H (2024) Adaptive federated learning with high- efficiency communication compression. In: 2024 43rd Chinese Control Conference (C), p 1903–1909. h t t p s : / / d o i . o r g / 1 0 . 2 3 9 1 9 / C C C 6 3 1 7 6 . 2 0 2 4 . 1 0 6 6 1 2 0 7 35. Beguier C, du Terrail JO, Meah I, Andreux M, Tramel EW (2021) Differentially private federated learning for cancer prediction. arXiv:2101.02997 36. Yousefpour A, Shilov I, Sablayrolles A, Testuggine D, Prasad K, Malek M, Nguyen J, Ghosh S, Bharadwaj A, Zhao J et al (2022) Opacus: User-friendly differential privacy training in PyTorch, arXiv:2109.12298, https:/ /doi.or g/10.48 550/a rXiv.2109.12298 37. Fu Y, Wang T (2024) Benchmarking secure sampling protocols for differential privacy. In: Proceedings of the 2024 on ACM SIG- SAC Conference on Computer and Communications Security, CCS ’24, Association for Computing Machinery, New York, NY, USA, p 318–332. https:/ /doi.or g/10.11 45/36 58644.3690257 38. Zhang Y, Ye Q, Hu H (2025) Federated heavy hitter analytics with local differential privacy. Proc ACM Manag Data 3(1). h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 7 0 9 7 3 9 39. Di Fazio A (2024) Enhancing privacy in recommender systems through differential privacy techniques. In: Proceedings of the 18th ACM Conference on Recommender Systems, RecSys ’24, Association for Computing Machinery, New York, NY, USA, p 1348–1352. https:/ /doi.or g/10.11 45/36 40457.3688019 40. Koskela A, Jälkö J, Honkela A (2020) Computing tight differen- tial privacy guarantees using FFT. In: International Conference on Artificial Intelligence and Statistics, PMLR, p 2560–2569 41. Abadi M, Chu A, Goodfellow I, McMahan HB, Mironov I, Talwar K, Zhang L (2016) Deep learning with differential pri- vacy. In: Proceedings of the 2016 ACM SIGSAC Conference on 1 3 Page 23 of 23 105 Peer-to-Peer Networking and Applications (2026) 19:105 Computer and Communications Security, CCS ’16, Association for Computing Machinery, New York, NY, USA, p 308–318. https:/ /doi.or g/10.11 45/29 76749.2978318 42. Dwork C, Roth A (2014) The algorithmic foundations of differen- tial privacy, Foundations and Trends® in Theoretical Computer Science 9(3–4):211–407 43. Shivaprasad K, Wadhawan A (2023) Deep learning-based plant leaf disease detection. In: 2023 7th International Conference on Intelligent Computing and Control Systems (ICICCS), p 360– 365. https:/ /doi.or g/10.11 09/IC ICCS56967.2023.10142857 44. Savazzi S, Nicoli M, Rampa V, Kianoush S (2020) Feder- ated learning with mutually cooperating devices: A consensus approach towards server-less model optimization. In: ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p 3937–3941. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / I C A S S P 4 0 7 7 6 . 2 0 2 0 . 9 0 5 4 0 5 5 45. Nezhadsistani N, Moayedian NS, Stiller NS (2025) Blockchain- enabled federated learning in healthcare: Survey and state-of-the- art. IEEE Access 46. Chen M, Hao Y, Hwang K, Wang L, Wang L (2017) Disease pre- diction by machine learning over big data from healthcare com- munities. IEEE access 5:8869–8879 47. Ruan Y, Zhang X, Joe-Wong C (2024) How valuable is your data? optimizing client recruitment in federated learning. IEEE/ACM Trans Netw 32(5):4207–4221. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / T N E T . 2 0 2 4 . 3 4 2 2 2 6 4 48. Bhardwaj A (2022) Framingham heart study dataset, h t t p s : / / w w w . k a g g l e . c o m / d a t a s e t s / a a s h e e s h 2 0 0 / f r a m i n g h a m - h e a r t - s t u d y - d a t a s e t . Accessed 24 April 2023 49. Siddhartha M (2020) Heart disease dataset (comprehensive). https:/ /doi.or g/10.21 227/d z4t-cm36 50. Detrano R (1988) Cleveland heart disease dataset. uCI Machine Learning Repository. h t t p s : / / a r c h i v e . i c s . u c i . e d u / m l / d a t a s e t s / H e a r t + D i s e a s e . Accessed 26 July 2024 51. Janosi A (1988) Hungarian heart disease dataset. uCI Machine Learning Repository. h t t p s : / / a r c h i v e . i c s . u c i . e d u / m l / d a t a s e t s / H e a r t + D i s e a s e . Accessed 26 July 2024 52. Khan A (2022) The heart failure prediction dataset: FIC Paki- stan. h t t p s : / / w w w . k a g g l e . c o m / d a t a s e t s / a s g h a r a l i k h a n / m o r t a l i t y - r a t e - h e a r t - p a t i e n t - p a k i s t a n - h o s p i t a l . Accessed 18 Sept 2024 53. Dubey M, Tembhurne J, Makhijani R (2025) Enhancing federated learning through differential privacy: Introducing FedHybrid for multicenter diverse heart disease datasets. IEEE Trans Emerg Top Comput Intell 1–14 54. Zhang W (2025) Diabetes detection: Predicting Type I diabetes with machine learning algorithm. In: Proceedings of the 2025 International Conference on Health Big Data, HBD ’25, Associa- tion for Computing Machinery, New York, NY, USA, p 16–22. https:/ /doi.or g/10.11 45/37 33006.3733009 55. Johnson AEW, Pollard TJ, Shen L, Lehman L-WH, Feng M, Ghas- semi M, Moody B, Szolovits P, Celi LA, Mark RG (2016) MIMIC- I, a freely accessible critical care database. Sci Data 3(1):160035 56. Dankar FK, Madathil N (2023) Using synthetic data to reduce model convergence time in federated learning. In: Proceedings of the 2022 IEEE/ACM International Conference on Advances in Social Net- works Analysis and Mining, ASONAM ’22, IEEE Press, p 293– 297. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / A S O N A M 5 5 6 7 3 . 2 0 2 2 . 1 0 0 6 8 6 1 5 57. Naseri M, Fernandez-Marques J, Gao Y, Pan H (2024) Privacy- preserving federated learning using flower framework. In: Pro- ceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, Association for Comput- ing Machinery, New York, NY, USA, p 6422–6423. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 3 7 5 2 8 . 3 6 7 1 4 4 7 58. Stevens E, Antiga LPG, Viehmann T (2020) Vol. Electronic ISBN:9781617295263. h t t p : / / i e e e x p l o r e . i e e e . o r g / d o c u m e n t / 1 0 2 8 0 2 7 2 59. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Van- derplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duch- esnay E (2011) Scikit-learn: Machine learning in Python J Mach Learn Res 12(null):2825–2830 60. Harris CR, Millman KJ, van der Walt SJ, Gommers R, Virtanen P, Cournapeau D, Wieser E, Taylor J, Berg S, Smith NJ, Kern R, Picus M, Hoyer S, van Kerkwijk MH, Brett M, Haldane A, Del Río JF, Wiebe M, Peterson P, Gérard-Marchant P, Sheppard K, Reddy T, Weckesser W, Abbasi H, Gohlke C, Oliphant TE (2020) Array programming with NumPy. Nature 585(7825):357–362. https:/ /doi.or g/10.10 38/s4 1586-020-2649-2 61. Tan J, Kwan B-H, Ng DW-K, Hum Y-C (2025) Adaptive focal loss with personality stratification for stably mitigating hard class imbalance in multi-dimensional personality recognition. Sci Rep 15(1):39241. https:/ /doi.or g/10.10 38/s4 1598-025-22853-y 62. Bonawitz K, Ivanov V, Kreuter B, Marcedone A, McMahan HB, Patel S, Ramage D, Segal A, Seth K (2017) Practical secure aggregation for privacy-preserving machine learning. In: Pro- ceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, CCS ’17, Association for Com- puting Machinery, New York, NY, USA, p 1175–1191. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 1 3 3 9 5 6 . 3 1 3 3 9 8 2 Publisher's Note Springer Nature remains neutral with regard to juris- dictional claims in published maps and institutional affiliations. Rodrigo Tertulino holds a Ph.D. degree from the Department of Informatics and Engineering, Uni- versity of Coimbra (UC). He is part of the Security Software Group (SSE) at the Centre for Informatics and Systems of the University of Coimbra (CISUC). He is also an associate professor at the Federal Institute of Education, Science, and Technology of Rio Grande do Norte (IFRN), Brazil. His research interests include pri- vacy, security, federated learning, blockchain applications in health- care (EHR) systems, and machine learning, with a focus on education and healthcare. Laércio Alencar is an Informa- tion Technology student at the Fed- eral Institute of Rio Grande do Norte (IFRN) and serves as a research fellow at the Software Engineering and Automation Research Laboratory (LaPEA), the leading research laboratory at the Mossoró Campus. His research focuses on leveraging excellence in software development and infor- mation technologies to explore the intersection between Computer Science and Education. He special- izes in the application of Machine Learning and Educational Technologies to design and optimize digital solutions that personalize learning and enhance teaching processes. His contributions reflect a strong emphasis on Software Engineering innova- tions, applied to the fields of education and healthcare.

Instant research discovery