Federated Learning Aggregation (FLAg) Strategies in Healthcare: A Systematic Survey and Taxonomy Based on Technical and Data Challenges

1. Introduction

Deep Learning (DL) has brought significant advantages in medical image analysis, including automated feature extraction, effective handling of large-scale image data, and improved model utility in both classification [1] and segmentation [2] workloads. However, models trained on single institutional data collections often lack generalization, resulting in significant utility variations across multi-institutional data distributions.

Multi-institutional studies in the medical domain [3,4] are facilitated through Centralized Data Sharing (CDS) in which DL models are trained on data pooled in one center (Figure 1a). CDS is hindered by privacy concerns (among others) as medical data must comply with confidentiality and security regulations, i.e., Personal Health information under the General Data Protection Regulation (GDPR) [5] and the Health Insurance Portability and Accountability Act (HIPAA) [6]. Federated Learning (FL) [7,8] enables multi-institutional collaborations while circumventing data sharing (Figure 1b). A central server initializes a global model parameters and broadcasts them to all collaborating institutions. During a training cycle (i.e., a federated round) each institution performs local training on its own data, sends the model updates to the central server, the server collects the updated models from all institutions and aggregates them in a global consensus model. It then broadcasts the updated global model back to all the collaborators. This process is repeated iteratively until model convergence is reached.

FL was made popular in 2017 with the Federated Averaging (FedAvg) algorithm firstly used for auto-complete typing functionality of phones [9] and then facilitating the very first FL study in healthcare [10]. During FedAvg the server aggregates collaborator updates by averaging their model parameters. In their evaluation, both the convex and non-convex problems were shown to converge toward the global minimum as the number of local training epochs and communication rounds increased.

This strategy works remarkably well for independent and identical datasets (IID), as they have a similar objective for all the clients. In medical applications, however, data tends to be non-IID as they are collected from different healthcare institutions that exhibit variations in terms of disease frequency, patient ethnicity, geographic location, gender, and socioeconomic status. This leads to biased models that even further worsen healthcare disparities [11,12]. Thus, fair aggregation strategies are required to ensure the global model neither favors some institutions over others, nor a patient population group dominates the global consensus model.

FL comes with other technical challenges (such as heterogeneous computational resources across collaborators) that can create communication bottlenecks, increase synchronization delays, under constrained network conditions hinder convergence, and potentially lead to model divergence. The most straightforward solution is a wait-free FL framework that allows asynchronous training [13]. Since the primary focus of FL is to incorporate knowledge from institutions with varying levels of computational capability, FLAg strategies should account for the computational load of clients without compromising model fairness, as highlighted in [9,14,15].

Medical datasets are often highly imbalanced and noisy due to variations in annotation quality or protocol, and image resolution. DL systems in healthcare are expected to provide reliable and well-calibrated predictions, rather than solely achieving high predictive utility [11]. Data heterogeneity and communication constraints, can cause clients to complete local training at different rates. In synchronous FL, the server must wait for slower clients, commonly referred to as stragglers, before performing model aggregation, resulting in increased training latency and reduced system efficiency. Therefore, FLAg strategies designed for healthcare applications should not only ensure fair contributions from participating hospitals, but also effectively mitigate the impact of stragglers [9,13,14].

Existing literature [16] has divided FL methods into horizontal, vertical, and federated transfer learning, describing their key differences and features. Kairouz et al. [17] provided a comprehensive theory of effective and efficient implementations of FL, while Rauniyar et al. [18] created a taxonomy-based review of FL applications in medical fields, with a focus on cancer detection and future model development. Pati et al [19] and Raza et al [20] gave an overview of FL applications in healthcare, mainly focusing on identifying threats and their mitigation through varying privacy preserving mechanisms. Other studies focused on FL problems when targeting a particular medical application, such as COVID-19 [21,22,23], cancer diagnostics [24,25,26] or rare diseases [27].

Other recent surveys have discussed various challenges to overcome FL including non-IID, stragglers, privacy, security, data uncertainty, asynchronous training, and a very limited focus given to aggregation strategies available in the current practice. In this review, we provide a comprehensive and structured overview of FLAg strategies and how they are addressing the challenges faced in FL Healthcare ecosystem that relate to either to data or technical issues (Figure 2).

2. Aggregation Strategies

FL emerged as one of the key methods for decentralized model training in privacy sensitive domains such as healthcare. The effectiveness of FL is highly influenced by the aggregation approach, i.e., how the central server combines the client parameters. A wide range of FLAg strategies have been proposed in various domains since 2016. Each of these FLAg strategies was developed to address various challenges such as the presence of system heterogeneity between different institutions, the non-IID nature of the data, and the variation in the client reliability due to resource constraints. For those new to the field, these strategies can be challenging to grasp because they are scattered and inconsistently presented across the literature. There is a need for a clear and structured overview that organizes these approaches into a coherent road map, helping readers navigate and interpret the existing body of work more effectively. This section aims to categorize FLAg strategies according to their underlying strategy (Figure 2), followed by a comparative discussion of their strengths and limitations in real-world scenarios. A summarized version of all families can be found in (Suppl.Table S1).

2.1. Parameter Aggregation

This is one of the earliest strategies, where a global model is constructed by aggregating parameters from each local client model after every few iterations of local training (Figure 2).

The historically first aggregation strategy (FedAvg [9]) falls under this category. Clients train locally before sending updates to the server, requiring fewer communication rounds while being relatively easy to implement. A shortcoming of this method are large update sizes, which is prone to model divergence under non-IID data. FedProx [14] sought to fix this by adding a proximal term to weight drift, forcing the global consensus model to maintain the impact of variable local updates. HeteroFL [28] performs partial weight aggregation that enables clients with heterogeneous resources to train models with different capacities while still contributing to a single global consensus model. Similarly, FlexiFed [29] addresses model architecture heterogeneity by aligning parameters across different client architectures with a reference map before aggregation, instead of blindly averaging the parameters.

Parameter aggregation comes with its own issues, including the risk of information leakage through model inversion [30], computational cost, and slower convergence when it comes large DL models [9,14], and the potential to wash out local patterns, especially in the case of non-IID data, leading to utility issues [12].

2.2. Gradient Aggregation

To reduce communication overhead and have fine-grained control over global optimization, an alternative is for the server to aggregate local gradients from clients rather than model weights (Figure 2).

FedSGD [9] and DPSGD-FL [31] were the earlier established gradient based aggregation strategies for FL which demonstrates that decentralized training can achieve convergence behavior comparable to centralized stochastic gradient descent (CSGD). FedGS [32] proposes a balanced framework that sends reduced information through gradient sparsification with a higher degree of freedom for additional control over communication and computation trade-off. FedNova [33] addresses the issue of objective inconsistency caused by the unfairly scaled gradients resulting from different numbers of local update steps across clients. It normalizes the local gradient updates to ensure a more balanced contribution from clients during aggregation.

As clients share the direction of optimization, gradient aggregation reduces communication overhead across distributed clinical sites. However, this approach may lead to oscillatory optimization with instability in updates. Under heterogeneous medical environments, conflicting gradient directions may degrade global model utility compared to conventional weight aggregation.

2.3. Cluster-Based Aggregation

Cluster-based aggregation was proposed to overcome the client drift inherent in heterogeneous datasets. In this approach, clients are organized into clusters based on criteria such as data distribution, computational capacity, or client availability, and aggregation is performed per cluster (Figure 2). In most cases, each cluster maintains its own global model, resulting in multiple cluster-level models rather than a single global model representing all the clients.

Along these lines, ClusterFL [34] groups clients based on the similarity of their updates and performs aggregation within each cluster. By dropping slow-converging or weakly correlated clients within each cluster, this method reduces communication overhead and accelerates convergence. However, this method also introduces cluster-wise dropout for slow clients and correlation-based client selection, which can exclude clinically valuable but resource-constrained participants, partially undermining the purpose of FL. FedCluster [35] extended this idea to grouping clients based on time zones and availability, employing cluster cycling to alternate training between clients. This reduces client drift and improves convergence under non-IID data settings.

The aforementioned strategies are based on hard association assumptions to form clusters of IID clients, but comes with the shortcoming that FL cannot effectively exploit similarities between different clusters. To overcome this, a soft clustering way is proposed. Fedsoft [36] allows every local dataset to follow a mixture of multiple source distributions. However, this approach does not explicitly consider similarities between clusters, potentially limiting its ability to leverage shared structure across clients. FedSim [37] further encourages collaboration among similar clients, while minimizing the variance from dissimilar clients through clustering.

Methods such as IFCA [38] which use dynamic clustering can cause clients to shift between clusters, potentially destabilizing training. Moreover, the generation of multiple global models challenges the traditional notion of generalization, as most clustering techniques prioritize cluster-level utility over a single unified global consensus model.

By leveraging clients with similar data distributions, cluster-based aggregation handles heterogeneous datasets, including differences between populations or institutions. Additionally, it provides a form of partial personalization, allowing separate cluster-level models rather than a single global model, which better accommodates local variability.

2.4. Knowledge Distillation Aggregation

Knowledge distillation aggregation allow clients to maintain different local models while contributing to an FL consortium. In this framework, clients send their predictions, logits, or a smaller model to the server instead of the local model parameters (Figure 2). Unlike conventional FL, where the global model is obtained by directly averaging client model parameters,the server uses the clients predictions to train a global student model. This allows clients who have different computational capacities and data constraints to train heterogeneous models and make them adaptable to the shared FL global ecosystem.

Knowledge distillation has evolved through several research directions. FedMD [39] was among the first methods to demonstrate that clients with heterogeneous model architectures can collaborate by exchanging prediction logits instead of model parameters. Building on this idea, FedDF [40] introduced server-side ensemble distillation, where the global model is trained by aggregating the predictions of local models rather than averaging their parameters. FedBE [41] further enhanced this framework by incorporating Bayesian inference to generate multiple candidate global models before compressing them into a single distilled model, accounting for uncertainty during aggregation. Subsequent methods extended these ideas to address specific challenges, including data-free distillation (FedFTG [42]), improved initialization and privacy with auxiliary data (FedAUX [43]), and ensemble-style model fusion (FedFusion [44]), among others (Suppl.Table S1).

Overall, knowledge distillation offers greater flexibility than parameter aggregation by allowing clients to train heterogeneous models tailored to their local data modalities (e.g., medical imaging, electronic health records, or biosignals) and computational resources, making it well suited to heterogeneous healthcare environments. However, because aggregation relies on model outputs rather than model parameters or raw data, the distilled global model may lose information contained in local representations, particularly for rare or complex patterns. Additionally, although sharing logits reduces the exposure of model parameters, prediction outputs can still reveal sensitive information through inference attacks, and biases present in local models may propagate into the distilled global model.

2.5. Representation Aggregation

Representation-based aggregation addresses non-IID data by aggregating feature-level representations learned by client models instead of their complete model parameters (Figure 2). By aligning latent feature representations across clients, these strategies improve collaboration under heterogeneous data distributions, such as class imbalance or varying patient populations.

Representation-based aggregation has evolved to improve feature alignment under heterogeneous data distributions. FedPer [45] was among the first methods to separate shared feature extraction from task-specific prediction by aggregating only the feature extractor while keeping classifiers local, enabling personalization under non-IID settings. Building on representation learning, MOON [46] introduced contrastive learning to align local and global feature representations, reducing client drift and improving training stability. FedProto [47] further generalized this idea by exchanging class prototypes instead of model parameters, allowing clients to communicate compact feature-level summaries while remaining architecture-independent. Subsequent methods extended these ideas to improve representation learning and personalization, including explicit decoupling of shared representations and local heads (FedRep [48]) and adaptive feature-aware aggregation for highly heterogeneous data (FedMix [49]), among others (Suppl.Table S1).

However, the quality of the aggregated model depends on the learned feature representations, which may become biased toward dominant client populations, potentially skewing the global model toward particular distributions [50]. Careful design of weighting and representation alignment strategies is critical to ensure equitable and effective aggregation.

2.6. Ensemble-Based Aggregation

Ensemble-based aggregation addresses client heterogeneity by combining the predictions of multiple local models rather than directly averaging their parameters (Figure 2). By operating in the prediction space, these strategies mitigate utility degradation caused by direct parameter averaging under strong statistical heterogeneity, such as that encountered by FedAvg [9].

Ensemble-based aggregation has evolved through several research directions. FedBE [41] introduced Bayesian model averaging in function space, treating client models as posterior samples to explicitly account for uncertainty during aggregation. FedAUX [43] further improved ensemble quality by assigning confidence-aware weights to client predictions using auxiliary data, allowing more reliable models to contribute more strongly to the aggregated output. More recently, ensemble techniques have been extended to graph-structured data, where ensemble Graph Neural Networks (GNNs) [51] combine node embeddings or predictions from multiple GNNs to improve robustness and generalization. Other ensemble-based strategies, including distillation-driven approaches such as FedDF, FedFusion, and FedSEAL, further compress ensemble knowledge into a single deployable global model (Suppl.Table S1).

Despite providing robustness and flexibility, ensemble-based aggregation necessitates the maintenance of multiple models, which increases computational complexity and resource requirements.

2.7. Attention-Based Aggregation

Another way to handle non-IID data is through an attention-based aggregation strategy. This approach introduces an adaptive mechanism for combining client updates by assigning importance to model parameters, representations or feature embeddings for each client during aggregation in the server (Figure 2). Instead of averaging all the parameters of the model, the server uses attention mechanisms to dynamically weigh client contributions based on their relevance and informativeness. This approach is particularly effective in addressing challenges posed by the non-IID nature of data distributions or multi-modality modeling, where each client trained data from different modalities.

Attention-based aggregation has evolved through several research directions. The Attentive Aggregation method [52] was among the first to introduce attention mechanisms for federated aggregation by iteratively assigning adaptive weights to client updates based on their similarity to the global model. Building on this idea, FedAMP [53] enabled personalized aggregation by allowing clients to receive greater contributions from similar peers, thereby reducing negative transfer under heterogeneous data distributions. More recently, HEALNET [54] extended attention mechanisms to multimodal healthcare by learning modality-specific parameter spaces within a unified global model. Subsequent methods further adapted attention for specialized tasks, including hierarchical client clustering for traffic prediction (FedDA [55]) and attention-guided client selection (FedABC [56]), among others (Suppl.Table S1).

The main advantage of attention-based aggregation is that it prioritizes more relevant and informative clients, allowing adaptive personalization [57]. However, this type of bias amplification directly conflicts with fairness requirements in healthcare AI. The attention weights are not interpretable even though they are mathematically defined. These weights do not always provide clear or clinically meaningful explanations for decision-making. This lack of traceability can hinder trust and adoption in real-world applications.

2.8. Meta-Learning Aggregation

The meta-learning global model is built based on client model learning. Instead of learning one static global model from local parameters or local SGD from clients, the server will learn the initialization rules and adaptation to each client’s data. Meta-learning focuses on learning algorithm utility for arbitrary tasks across devices (Figure 2).

FedMeta [58] demonstrates this by facilitating collaborative training across distributed clients with a meta-learner, which can quickly adapt to a client’s local data using only a few gradient updates. Building on this idea, pFedMe [59] formulates FL as a bi-level optimization problem, learning a global initialization that can rapidly adapt to individual client needs. Similarly, Per-FedAvg [60] extends the traditional FedAvg framework by incorporating meta-learning, allowing clients to achieve personalization through a few local gradient updates. Moving beyond shared initialization, a meta-learning inspired personalized FL approach, pFedHN [61] introduces a model generator approach, enabling stronger and more flexible personalization with relatively small communication overhead.

Meta-learning introduces a shift from simply learning a single global model to learning how to learn across diverse clients [7]. This becomes especially important in healthcare settings, where data are inherently heterogeneous. By capturing shared knowledge while allowing rapid adaptation to local distributions, meta-learning enables models to perform better on rare diseases and personalization [62], which are often under-represented in global training. However, meta-learning also increases computational complexity to the system as it requires additional optimization steps to learn adaptable model parameters. These strategies are sensitive to data distribution shifts, which can negatively impact utility if local datasets are non-IID [63]. A critical concern in clinical settings is explainability, and interpretable meta-learning models tend to function as complex black-box systems, making it difficult for clinicians to understand or trust their predictions, thereby limiting their practical adoption in healthcare environments.

2.9. Optimization-Based Aggregation

In optimized-based aggregation, the server treats the aggregation process as a constrained optimization problem instead of naive parameter averaging. These strategies actively adjust how updates are combined to improve convergence, stability, and fairness, especially under non-IID data (Figure 2).

Optimization-based aggregation has evolved through several research directions. Oort [64] improves training efficiency by prioritizing clients that provide high-quality and timely updates, accelerating convergence compared with uniform client sampling. FedDyn [65] further stabilizes optimization by introducing dynamic regularization to reduce divergence caused by heterogeneous client objectives. In parallel, FedAdam and FedYogi [66] adapt server-side optimization by incorporating momentum and adaptive learning rates, enabling more robust convergence under heterogeneous data distributions. Subsequent methods further refined optimization through drift correction and update alignment, including FedDC [67] and correlation-aware optimization in FedCorr [68], among others (Suppl.Table S1).

Optimization-based aggregation combines local updates into a coherent and high-performing global model with stability and faster convergence. However, rapid convergence makes the system sensitive to hyper parameters, with even minor misconfiguration in learning rates or momentum terms significantly affecting model utility. In healthcare, particularly in the presence of rare disease cases, adaptive aggregation mechanisms may carelessly control under-represented updates if they contribute less to the global objective.

2.10. Sparsification Aggregation

In sparsification-based aggregation clients send masked parameters or a subset of model parameters to the server, with the goal of easing the communication bottleneck. Masking strategies of this nature include top-k gradient selection, magnitude pruning, random sparsification, or structured pruning masking, enforcing only the important parameter subsets to reach the server (Figure 2).

SparseFed [69] was among the first methods to reduce communication costs by applying global top-k update sparsification together with device-level gradient clipping, while also improving robustness against model poisoning attacks. Building on this idea, FedSPA [70] incorporated sparsification directly into the aggregation process by enforcing a shared sparse parameter structure across clients. Lossless Gradient Sparsification [71] further improved communication efficiency by representing gradients in an alternative space that increases their compressibility without sacrificing model utility. Subsequent strategies extended these ideas through adaptive parameter selection and sparsification strategies, including FedSparse [72] and Top-k/Rand-k gradient selection methods [73], among others (Suppl.Table S1).

Sparsification is compatible for large-scale applications, allowing efficient network bandwidth usage in distributed learning [72]. However, as this method depends on the selection strategy, fairness in aggregation of under-represented data distribution could be missed. It also complicates debugging, making it difficult to trace individual contributions for clinical validation because of irreversible sparsification.

2.11. Peer-to-Peer Aggregation

Peer-to-peer aggregation eliminates the need for a central server, allowing clients to converge through consensus rules without strict central averaging. Generally, peer-to-peer aggregator and client have a common predefined communication protocol that they abide (Figure 2).

The segmented gossip approach [74] was among the first decentralized aggregation strategies, combining model partitioning with gossip-based communication to improve bandwidth utilization and convergence without a central server. Building on FL optimization, SecureD-FL [75] employs the Alternating Direction Method of Multipliers (ADMM) to enable peers to collaboratively compute the global model through distributed consensus. BAFFLE [76] further extended this paradigm by leveraging blockchain and smart contracts to coordinate aggregation, while improving computational efficiency through parameter-space decomposition. Subsequent strategies expanded peer-to-peer aggregation to broader FL environments, including neighborhood-based serverless aggregation [77] and other blockchain-enabled frameworks such as Block-chain-based Federated Learning framework with Committee consensus BFLC [78], among others (Suppl.Table S1).

Peer-to-peer aggregation improves robustness by eliminating the single point of failure. However, it comes with notable challenges. For example, gossip or ring-based topologies can become inefficient as the number of clients grows [74]. The lack of a central authority makes it harder to detect and mitigate malicious updates, which increases security threats. The absence of centralized tracking complicates the contribution, fairness, and critical auditing concerns in sensitive healthcare data [79].

2.12. Confidential Aggregation

Confidential aggregation addresses privacy risks posed by the aggregator itself. The standard threat model assumes an honest-but-curious adversary: the aggregator executes the aggregation protocol faithfully, but may attempt to extract sensitive information from the individual model updates submitted by clients. Confidential aggregation strategies prevent such leakage by ensuring that the server can perform the aggregation without ever seeing individual client updates in plain text. (Figure 2)

Homomorphic encryption (HE) [80] allows the aggregator to perform computations directly on encrypted model updates, ensuring that the server never gains access to raw gradient or weight values, while still producing correct aggregated results. This provides strong cryptographic privacy guaranties at the cost of significant computational and communication overhead, driven by the explosion in ciphertext size and computational complexity, which can be prohibitive in large-scale healthcare deployments with numerous participants.

Secure aggregation protocols, most notably SAFELearn [81], enable the server to compute the sum of client updates without learning individual contributions. Through pairwise masking and secret sharing techniques, each client’s update is encrypted in such a way that the masks cancel out only after all updates are combined, ensuring that only the final aggregate is revealed to the server. While computationally more efficient than HE, secure aggregation remains vulnerable to inference attacks on the final aggregate itself and requires careful handling of client dropouts.

Trusted Execution Environments (TEEs) provide a hardware-based approach to confidential aggregation by processing client updates within an isolated, attested enclave, such as Intel SGX or AMD SEV. The aggregator’s own operating system, and its system administrators, cannot inspect computations within the enclave providing a practical alternative to pure cryptographic strategies with lower computational overhead. Recent work has demonstrated the feasibility of TEE-based FL in healthcare settings [82], though limitations in secure memory capacity and vulnerabilities to side-channel attacks remain active concerns.

Secure Multi-Party Computation (SMPC) extends the principle of confidential aggregation by distributing the aggregation computation across multiple non-colluding servers, such that no single party can access individual client updates in plain text. Secure Multi-Party Computation such SMPC protocols[83] and Arithmetic, Boolean, and Yao (ABY) [84,85] enable parties to jointly compute aggregate model parameters while maintaining privacy guarantees even if some servers are compromised. SMPC-based FL mechanisms [86] have been increasingly adopted for applications requiring strong privacy guarantees across institutional boundaries, at the cost of increased communication rounds and protocol complexity.

Confidential aggregation strategies form the backbone of privacy-preserving FL in healthcare by providing cryptographic security guarantees that client updates remain confidential during the aggregation process. However, the stronger the cryptographic guarantee, the higher the computational and communication overhead, which raise a critical tension in resource-constrained healthcare deployments. Furthermore, confidentiality during aggregation does not eliminate the risk of information leakage from the aggregated model itself, underscoring the need for complementary privacy-preserving techniques.

2.13. Byzantine-Robust Aggregation

Byzantine robustness refers to a system’s ability to maintain correct operation and reach consensus even when some nodes are compromised, malfunctioning, or acting maliciously. In FL aggregation, it tackles a general security threat within the FL network, where the adversaries could be any of the protocol participants. For example, a malicious client (or a coalition of clients) may submit intentionally corrupted or crafted updates to degrade the global model’s accuracy or to inject targeted backdoors. Unlike the honest-but-curious adversary in confidential aggregation, byzantine adversaries do not follow the protocol and may send arbitrary values. The goal of Byzantine-robust aggregation is to ensure that the final global model remains accurate and uncorrupted. This distinction is critical in healthcare, where a poisoned model could lead to misdiagnosis or incorrect clinical decisions. (Figure 2)

Statistically robust aggregation strategies defend against Byzantine attacks by replacing or modifying the standard averaging operation with aggregation functions that are inherently resistant to extreme values. Solutions like Multi-Krum [87] select a single client update that is most similar to its neighbors, effectively discarding outliers, while its generalized variant selects and averages a trusted subset of updates. Beyond statistical filtering, trust-based aggregation strategies maintain a reputation or trust score for each client and weight their contributions accordingly. FLTrust [88] introduces a trusted root model that validates client updates before aggregation, ensuring that only updates consistent with trustworthy behavior influence the global consensus model. Similarly, FORTA [89] enhances the Krum algorithm by incorporating discrete Fourier transform guidance to better distinguish between benign and malicious updates under heterogeneous data conditions.

An emerging direction in Byzantine-robust aggregation is the utilization of computational verifiability, which enables clients or auditors to cryptographically confirm that aggregation was performed correctly on legitimate updates. Techniques leveraging zero-knowledge proofs [90], TEEs [91], and authenticated data structures [92] allow participants to verify that the server has not tampered with the aggregation process and that all included updates originated from registered clients [93]. Verifiable aggregation addresses a critical trust gap in cross-institutional healthcare FL where no single authority is universally trusted, though these strategies currently incur significant computational and communication overhead that limits practical deployment.

Collectively, Byzantine-robust aggregation strategies provide a multi-layered defense against adversarial manipulation of the global model, with methodologies from statistical outlier rejection to trust-based client weighting to cryptographically verifiable computation. A key concern in healthcare is that aggressive filtering of anomalous updates may inadvertently suppress legitimate contributions from institutions with underrepresented patient populations, creating a tension between security and fairness. Verifiable aggregation addresses the trust gap in cross-institutional collaboration, where no single authority is universally trusted, though practical adoption remains constrained by computational overhead.

3. Challenges in Healthcare FL

Deploying FL in real-world healthcare settings comes with significant challenges, span across technical (Figure 3) and data-related challenges (Figure 3).

FL practitioners must carefully select the right FLAg strategy, as it directly affects the utility, fairness, and reliability of the model. Simple FLAg strategies can fail under complex conditions, leading to biased predictions or unstable training. In healthcare, where decisions can impact patient outcomes, careful consideration of both system constraints and data characteristics is essential to ensure robust, trustworthy, and clinically-relevant models.

This section discusses key technical and data-related challenges encountered when implementing FL in healthcare and highlights the FLAg strategies designed to address them.

3.1. Technical Challenges

3.1.1. Stragglers

Stragglers are clients whose local training or communication takes significantly longer than other clients per FL round. Their delayed model updates can slow down or stall the aggregation process, particularly in synchronous FL algorithms, where the server typically waits for updates from a predefined set of clients before proceeding to the next round.

Although a natural response for stragglers is to ease the synchronization requirement through asynchronous strategies, this strategy do not universally solve the problem and generally more effective in large-scale FL settings with hundreds of participating clients (e.g., at scale of 100 or more) to achieve utility and training efficiency comparable to centralized learning [94]. Straggler mitigation in cases like healthcare requires approaches that explicitly accommodates heterogeneous computational capabilities rather than relying solely on asynchronous participation. Existing aggregation strategies address this challenge through several complementary mechanisms. Sparsification based aggregation strategies such as [95] allow flexibility to clients to train and contribute sub-models suited for their computational capabilities and eliminate the likelihood that resource limited clients become persistent stragglers during aggregation. In contrast, rather than preventing the occurrence of stragglers, optimization-based aggregation strategies accommodate differences in local training progress and correct the resulting objective inconsistency during aggregation. For example, FedNova [33] uses normalized averaging allows fast clients perform more local updates per communication round and stragglers to train minimal rounds and attempt to remove objective inconsistencies and ensures fair contribution from all clients during aggregation. Clustering-based aggregation addresses the problem by grouping clients based on their computational capabilities and aggregate them without requiring every participant to progress at the pace of the slowest client. Peer-to-peer aggregation further relaxes dependence on a centralized synchronization barrier. Instead of eliminating or grouping the clients, clients exchange and update models with peers according to their availability. Faster participants can therefore continue exchanging updates without stalling for stragglers, effectively distributing the synchronization burden across the network.

Collectively, these strategies illustrate a shift from strict synchronization toward adaptive and selective aggregation strategies to mitigate the negative effects of stragglers and the heterogeneity caused by them in FL systems.

3.1.2. Asynchronous Training

Asynchronous training allows the global model to update continuously as client updates arrive, rather than waiting for all participating sites to complete local training.

While this improves system responsiveness and avoids delays caused by stragglers, it introduces challenges such as slower or unstable convergence and scalability issues due to inconsistent update timing and heterogeneous data distributions.

Traditional asynchronous aggregation suffers from communication bottleneck when the consortium is large, resulting in suboptimal convergence and while the parameter traffic to the server is congested. Theoretical analysis of Gradient-based Asynchronous Decentralized-PSGD [94] shows that the clients to use stale weights to compute gradients shows consistent convergence rate similar to D-PSGD and SGD, suggesting that the issue is not with asynchronism itself but the way it is managed. Methods like FedBuff [96] attempt a buffered asynchronous aggregation where the server aggregates clients updates in a secure buffer before performing the update. This provides a good guaranty of convergence and achieves better scalability, especially in large healthcare collaborative settings with many participating nodes. Similarly, hierarchical FL Async-HFL[97] utilizes hierarchical asynchronous aggregation to aggregate in multi-layer manner (tier 1 -cloud layer, tier 2 - gateway layer , tier 3 - edge layer), which collaboratively optimize model convergence and stability. Despite these advances, fully asynchronous strategies still struggle with the dual challenges of staleness and data heterogeneity.

Strategies such as knowledge distillation based or meta-learning based can focus on issue caused by staleness and data heterogeneity. By solely relying on sharing knowledge predictions or initializations, these approaches indirectly makes sure to reduce the impact of negative effects of data heterogeneity and staleness.

Rather than eliminating the challenges of asynchronism, these approaches progressively constrain and manage it. The main focus of these strategies is turning a source of instability into a tunable design.

3.1.3. Varying Resource Environments

Large healthcare institutions may be wealthy in terms of resources, such as powerful GPUs, stable network connectivity, and dedicated AI teams. On the other hand, smaller clinics or regional hospitals might operate on limited CPU resources, intermittent connectivity, and heavy clinical workloads. Apart from that, data heterogeneity may further be amplified due to differences in imaging devices and acquisition protocols. This resource variation immediately breaks the assumption that all clients can train the same model under the same conditions, bringing model heterogeneity. Heterogeneous models vary in model size and types across clients, which creates stale updates and inconsistency in training with each client.

Many aggregation strategies have attemped to handle model heterogeneity. For instance, clustering-based strategies such as ClusterFL [34] and FedCluster [35] take the first step by preserving model diversity, allowing multiple client-specific models to coexist and contribute collectively, rather than collapsing them into a single averaged solution. Knowledge distillation approaches like FedMD [39], FedDF [44], and FedKD [98] shift the focus from parameter sharing to knowledge sharing, enabling clients with different model architectures and capacities to exchange soft predictions, thereby decoupling learning from strict structural uniformity. These approaches alone do not resolve the instability introduced by inconsistent local training, which is where optimization-based strategies such as FedProx [14], FedNova [33], and SCAFFOLD [99] play a critical role by correcting objective mismatches, mitigating client drift, and stabilizing convergence despite uneven update frequencies.

Attention-based aggregation strategies like FedAtt [57] and FedAMP [53] introduce a more selective mechanism, dynamically weighting client contributions based on their relevance or reliability, ensuring that faster or higher-quality updates are emphasized without entirely discarding slower participants. Together, these strategies illustrate a shift from rigid averaging toward adaptive, multi-faceted aggregation, where diversity is not suppressed but structured—balancing flexibility, stability, and fairness in the face of real-world healthcare heterogeneity. Peer-to-peer aggregation like Fedlesscan [100] includes a clustering-based FL strategy designed specifically for peer-to-peer environments to make it staleness-aware, mitigate slow model updates and avoid wasted contribution of clients.

These approaches shift aggregation from a rigid averaging process to a more flexible, heterogeneity-aware mechanism that sustain learning in healthcare environments.

3.1.4. Quantization Aware

The transmission of full-precision model updates becomes substantial communication overhead in FL, particularly when large models are trained across client with resource heterogeneity. Quantization-aware strategies address this by allowing clients to send compressed updates while aiming to retain sufficient information for effective global aggregation.

One of the most commonly discussed problems in compression methods, as discussed in [101], [102] and [103] is Quantization noise. This is the noise generated when high precision numerical value is represented in low precision numerical value. In FL, the quantization noise is not uniform across clients, where differences in hardware and implementation can introduce variability in how updates are quantized, making aggregation noisier and potentially less reliable. This issue becomes more in non-IID healthcare data settings, where client updates are already misaligned. Adding quantization noise can further distort their contribution, amplifying heterogeneity rather than mitigating it.

Ultimately, the effect of quantization depends on type of information being communication. Strategies such as Gradient-based and parameter based approaches provides a effective framework for applying quantization to the updates exchanged between clients and the server. However, quantization may introduce approximation errors into the transmitted updates, leading to the development of methods that adjust the quantization process according to the characteristics of the local updates. For instance, FedWSQ [102] employs distribution-aware non-uniform quantization to reduce quantization error by considering the statistical distribution of local model updates, while weight standardization is incorporated to improve learning stability under heterogeneous data. Adaptive quantization is considered in AdaGQ [104], where the quantization resolution is dynamically adjusted according to gradient characteristics across training rounds and heterogeneous communication conditions across clients.

Collectively, these approaches demonstrate that the effects of quantization can be controlled through distribution-aware, and adaptive quantization mechanisms rather than by applying a uniform fixed-precision representation to all client updates. This balance is more important in heterogeneous healthcare environments, where variations in local data distributions and system resources may result in different update and communication characteristics across participating clients.

3.1.5. Security

In FL for healthcare, security in aggregation refers to ensuring both that the model training process and the output are robust against adversarial manipulation by adversaries. Unlike privacy threats that target data confidentiality, security threats aim to compromise model integrity. The standard threat model assumes a Byzantine adversary: a malicious client (or a coalition of clients) that submits arbitrarily corrupted updates to degrade model accuracy or inject a targeted backdoor. Or a malicious server that manipulates the aggregation process. In healthcare, a poisoned model could lead to systematic misdiagnosis with direct clinical consequences.

Model poisoning attacks allow malicious clients to send manipulated updates that degrade model utility or introduce backdoor, particularly dangerous in healthcare, where incorrect predictions can have clinical consequences. Inference attacks such as gradient inversion or membership inference can exploit shared updates to reconstruct sensitive patient data, even without direct access to it. Unreliable or heterogeneous clients may behave unpredictably, either due to faults or adversarial intent, making it difficult to distinguish between benign noise and malicious contribution. Secure communication overhead can become significant, especially when dealing with bandwidth-constrained hospital networks, creating a tension between security and efficiency.

FL must ensure model integrity in the presence of malicious or faulty clients, which is addressed through robust aggregation strategies. Algorithms such as Multi-Krum [105] select updates that are closest to the majority, filtering out anomalous or adversarial contributions, while trimmed mean and coordinate-wise median reduce the influence of extreme gradient values by aggregating only central tendencies. Bulyan [106], a hybrid method, further strengthens the robustness by combining outlier filtering with robust averaging to defend against strong Byzantine attacks.

To reduce reliance on centralized servers and improve communication resilience, FL aggregation approaches such as gossip learning and decentralized SGD (D-PSGD) [94] allow clients to exchange model updates directly in a peer-to-peer network, eliminating single points of failure while still enabling global learning. Finally, meta-learning-based approaches such as FedMeta [58], Per-FedAvg [60], and pFedMe [59] shift the objective from learning a single global model to learning a shared initialization that can rapidly adapt to each client’s local distribution, enabling strong personalization in heterogeneous healthcare environments.

Together, these aggregation strategies form a multi-layered framework where privacy, robustness, efficiency, and personalization are jointly addressed rather than treated in isolation.

3.2. Data Related Challenges

3.2.1. Uncertainty

Uncertainty in healthcare FL can broadly arise from aleatoric and epistemic sources. Aleatoric uncertainty reflects the noise present in clinical observations, such as measurement noise and ambiguity in medical data. In contrast, epistemic uncertainty arises from limitations in the model’s knowledge, which may result from insufficient, sparse, or unrepresentative training data across participating clients

The main challenge with uncertainty in healthcare is that it is very difficult to distinguish outlier data or data anomalies from non-IID data in healthcare applications with rare-disease conditions and unbalanced data. To achieve robust aggregation, deviations in data and model updates are mostly identified and handled to maintain stability. However, in heterogeneous settings such deviations may arise from rare cases, rather than anomalous or unreliable data. This may result in instability and ambiguity in model aggregation and bias amplification.

In order to handle client level uncertainty, oort [64] provides a client selection framework that filters the participating clients before aggregation. Instead of aggregating all updates, Oort allows the selection of high-utility and reliable clients, effectively reducing the impact of noisy, slow, or low-quality data sources and thereby stabilizing the aggregated global model under heterogeneous healthcare settings.

Model level uncertainty is addressed via probability-based ensemble aggregation, FedBE [41] incorporates uncertainty in decision-making by placing a distribution over model weights and marginalizing these models to form a whole predictive distribution. In this case, local models are treated as samples from a posterior distribution rather than deterministic parameter estimates. The server aggregates these model updates through Bayesian ensemble, capturing model-level uncertainty and improving robustness against variability across hospitals, patient demographics, and data collection protocols. Another approach to mitigate model level uncertainty is TM-FAL [107], which uses temporal uncertainty strategy instead of relying on a single model prediction to avoid inconsistent anomalies and suitable for long-tailed datasets.

These methods collectively indicate that uncertainty in FL arises at multiple stages of the learning pipeline, including client participation, model estimation, and aggregation. Existing methods tries to handle specific aspects of uncertainty in highly heterogeneous healthcare settings and still remain as an open challenge.

3.2.2. Privacy

Privacy in FL refers to the protection of sensitive patient information such as medical images, electronic health records, and genomic data from being exposed during training, even though raw data never leaves the hospital. Institutions share model updates instead of sharing raw data. However, these updates indirectly may leak private information, making privacy a key concern rather than something that being guaranteed.

One of the major threats is model inversion, where an intruder may exploit access to model parameters or gradients to reconstruct sensitive patient information, such as medical images or underlying clinical features, effectively reversing the learning process to approximate original health records. Similarly, membership inference is when the attacker attempts to determine whether a specific patient’s data was part of the training dataset. In healthcare, this can expose whether an individual was treated for a particular condition at a hospital, leading to serious privacy violations. Attribute inference on the other hand, is when sensitive hidden attributes such as disease status, genetic predisposition, or demographic factors are inferred from seemingly innocuous model outputs or intermediate representations, even when those attributes were never explicitly included in the training data.

Secure aggregation methods, such as SecAgg [108], homomorphic encryption [109], and secret sharing, where individual client updates are cryptographically masked so that the server can only observe the aggregated result, all mitigate reconstruction of patient-level information. Differential privacy [110], [70] provides a complementary statistical defense by adding calibrated noise to model updates, offering formal privacy guarantees bounded by a privacy budget that quantifies the trade-off between confidentiality and model accuracy [17].

Overall, privacy-preserved FL aggregation strategies provide important safeguards against information leakage, while enabling collaborative model training across healthcare institutions. However, achieving strong privacy guarantees seems to come at the cost of model utility, communication efficiency, and computational overhead, highlighting the need for methods that can effectively balance privacy protection and learning utility.

3.2.3. Data Imbalance

Data imbalance is common since each institution contributes differently in the size and composition of their local datasets. Large tertiary hospitals may contribute data from broad patient populations, whereas community hospitals or specialized clinics may contain fewer samples concentrated around particular diseases or demographic groups. Consequently, healthcare FL may exhibit imbalance at multiple levels, including unequal dataset sizes across institutions, skewed class distributions within individual clients, and long-tailed representation of rare diseases or patient populations.

One major issue is biased global model learning on long tailed datasets, where clients with larger or more diverse datasets dominate the aggregation process, causing the global model to be skewed toward majority populations while under performing on under-represented groups or smaller healthcare institutions. This leads to poor generalization across clients, especially in healthcare settings where rare diseases, minority populations, or small clinics may be systematically ignored.

Optimization based aggregation addresses the main statistical challenge associated with data imbalance is global objective drift. For example, FedProx [14], controls client drift during local optimization and stabilizes contributions from clients with heterogeneous data distributions, thereby improving convergence stability even when data distributions are uneven. In contrast, FedWeightedAvg [102] adjusts the contribution of individual clients during aggregation based on factors such as dataset size or estimated importance. Beyond differences in client-level contributions, imbalance may also occur within local datasets through unequal class representation, TM-FAL [107] addresses this class imbalance through class-aware, pseudo-label-guided active sample selection by promoting a more balanced representation of classes throughout training.

Meta learning strategies such as FedMeta[58], Per-FedAvg[60] and pFedMe[59] address the imbalance by learning a shared initialization that can quickly adapt to each client’s local data distribution, making the global model less sensitive to skewed or uneven datasets.

Attention-based strategies such as FedAtt[57] and FedAMP[53] dynamically assign higher weights to more informative or under-represented client updates. This allows the aggregation process to selectively emphasize useful contributions, while reducing the dominance of large or redundant datasets.

Ensemble-based aggregation such as FedSDC [111] addresses class imbalance by introducing shuffled multi-head classifier and incorporate ensemble strategy on top of a shared feature extractor for decision stability.

Overall, existing approaches such as clustering, meta-learning, attention-based aggregation, and ensemble strategies have demonstrated their effectiveness in addressing the consequence of data imbalance in FL. However, maintaining unbiased model utility across highly heterogeneous and long-tailed healthcare datasets remains a significant challenge, particularly when client data distributions differ substantially in size, quality, and class representation.

3.2.4. Fairness and Bias

Fairness in healthcare FL concerns avoiding systematic disparities in model utility across participating institutions and patient populations, while ensuring that the aggregation process does not systematically disadvantage particular clients or underrepresented groups.

Maintaining fairness in healthcare FL is a challenging task as it involves data from multiple clients of different geographical regions and demographic groups. Such difference among these data sources creates biases in model utility through either over or under-representation.

In order to improve personalization, clustered based aggregation approaches such as clusterFL [34] and Flexifed [29] group institutions with similar characteristics, reducing bias introduced by aggregating fundamentally different populations.

Representation based aggregations address heterogeneity by aligning the information learned across clients. For example, MOON [46], uses contrastive learning to constrain local representations relative to global and previous local representations, whereas FedProto [47] exchanges and aggregates class-level prototypes to promote consistency in representation spaces across clients. By encouraging semantically related information to remain aligned across heterogeneous sites, representation-based approaches can improve personalization when local feature distributions differ substantially [27].

Personalization strategies, including Ditto [112] and pFedMe [59], are especially valuable in healthcare as they allow each institution to adapt the global model to its local patient population, accounting for variations in disease prevalence, equipment, and clinical protocols. Similarly, approaches based on adversarial training such as [113], propose individual fairness on the global model as well as on client-side through robust optimization where a model is trained on both clean and adversarially perturbed inputs to improve resistance against worst-case perturbations.

Collectively these types of aggregation address different sources of unfairness in FL healthcare systems. Achieving fairness requires more than balanced participation during aggregation. It requires aggregation and optimization strategies that account for both institutional heterogeneity and population-level differences throughout the learning process. Robust approaches further supports this objective by improving model resilience to variations across clients and patient populations diversity.

4. Discussion

In this article, we provided a systematic survey and a taxonomy of FLAg strategies designed for healthcare, while taking into account their relation to distinct technical and data-related FL challenges. Significant progress is recognized in FLAg strategies to improve model utility and privacy guarantees, while mitigating these challenges (Figure 3), e.g., stragglers, system heterogeneity, fairness and biases, uncertainty, security, and data imbalance. With this focused content, this article intends to serve as a guide to readers navigating and interpreting the current scattered body of relevant literature, towards developing and deploying scalable, equitable, and clinically reliable FLAg strategies.

The following trade-off patterns are recognized across current FLAg strategies:

• Robustness vs Fairness: FLAg strategies designed to resist anomalous or outlier updates also tend to down-weight updates that deviate from the majority. In healthcare, such outliers are not necessarily malicious, but they may reflect long-tailed data distributions, including rare disease populations or minority cohorts. Thus, improving robustness should be done with care as it is often a byproduct of suppressing underrepresented features, compromising healthcare equity.
• Personalization vs Generalization: Personalized FL approaches adapt models to local hospital data distributions or particular individual as a target utility. In contrast, generalization targets single global model that works consistently throughout. The key challenge here is that one must completely choose one side. In practice, it is not alright and many approaches try to balance both by using Shared global backbone and local fine-tuning or clustering similar clients.
• Stability vs Convergence rate: Strategies that stabilize training through fewer local epochs with smaller updates, regularization, and conservative aggregation can reduce high-frequency fluctuations in model utility during training and improve convergence stability across heterogeneous clients. However, these strategies may slow down convergence rate, increasing training time and computational costs. In clinical settings, unstable models are unreliable. The goal here is not to choose stability over speed, it is to avoid naive stability that kills learning efficiency.
• Computational expensiveness vs Communication overhead: Reducing communication to fewer rounds or sending compressed updates can reduce communication overhead. However, performing more local epochs between federated rounds and the additional processing required for compression and decompression may increase the local computation burden on participating hospitals. In contrast, reducing local computation leads to more frequent communication among the clients. Hospitals have uneven infrastructure with some able to handle complex computing, whereas others cannot handle it. Hence, design choices made with only one collaborating institution in mind could penalize the other(s).
• Traceability vs trustworthiness: Improving aggregation traceability by documenting can improve transparency and auditability. However, it can expose vulnerabilities such as enabling adversarial manipulation or privacy leakage. In healthcare, trust is not just about visibility, but about guarantees under adversarial and sensitive conditions.

These trade-offs arise from the interaction of statistical heterogeneity across institutions, such as differences in patient populations and disease distributions; system-level constraints, including computational capacity and communication bandwidth; and strict ethical and regulatory requirements related to privacy, fairness, and explainability. As a result, FLAg strategies should not be simply optimized for model utility but must balance the aforementioned criteria.

Frameworks such as NVIDIA FLARE [114], Flower [115], FATE [116], OpenFL [117], and IBM FL [118] play a critical role in enabling real-world deployment. Each platform is based on several critical features such as model compatibility, machine learning library support, expandability, privacy provisions, and commercial usage [119]. These frameworks provide end-to-end infrastructure for orchestrating distributed training across multiple institutions, handling tasks such as client-server coordination, secure data exchange, model versioning, and fault tolerance. They also support integration with hospital IT systems, allowing models to be trained directly on local clinical data while complying with privacy regulations like HIPAA. In practice, they simplify deployment by offering scalable architectures, APIs for customization, and support for heterogeneous environments where different hospitals may use different hardware or data formats. As a result, these frameworks act as the operational backbone that bridges theoretical FL methods with practical, large-scale, and regulation-compliant healthcare applications.

Benchmark datasets and organized computational challenges have played an important role in advancing FL research in healthcare. In medical imaging, the Federated Tumor Segmentation (FeTS) challenges provide a notable example, using multi-institutional multi-parametric brain MRI data to facilitate the development and evaluation of FLAg strategies for segmentation workloads [120,121]. These challenges have enabled the systematic evaluation of FLAg strategies across data originating from different healthcare institutions. However, much of the current literature still relies on curated benchmark datasets that are often preprocessed and standardized. While such settings support reproducible evaluation, they may not fully capture the noise, missing values, inconsistencies, and other complexities encountered in real-world clinical data, highlighting the need to further evaluate FLAg strategies under realistic clinical conditions.

Several gaps are identified in FL research between current methodological advances and real-world deployment in healthcare, including limitations in i) evaluation beyond model utility, ii) clinical risk assessment, and iii) governance requirements for transparent and accountable FL. Most FL studies focus on algorithmic improvements in isolation, assuming ideal infrastructure, stable connectivity, and homogeneous client capabilities. In reality, healthcare institutions differ widely in computational resources, data storage systems, and network reliability. The current literature also still continues to discuss the model utility centric as an evaluation metric, which is not sufficient for healthcare applications. There should be more emphasis on fairness, uncertainty, and clinical risk assessment for the model as a standard evaluation framework. Furthermore, there is a minimal understanding of regulatory and governance requirements for FL deployment in healthcare. Although privacy-preserving techniques are widely discussed, there is still a limited work to make sure there is transparency, auditability, and compliance with evolving healthcare regulations. Also, accountability, model validation, and cross-institutional trust remain largely not addressed, highlighting the need for frameworks that align technical design with legal and ethical standards.

Addressing these challenges requires a shift from purely algorithmic improvements toward system-aware designs that integrate adaptive aggregation, efficient communication protocols, and robust data standardization strategies tailored for clinical environments.

5. Conclusion

This review article offers a taxonomy of current FLAg strategies and how they strategically handle technical and data-related challenges during their deployment. It further also discusses the trade-off that these FLAg strategies bring. However, to have a more robust approach to clinical variability in real-world deployment, methodological innovation proven with system-level integration is necessary as a future direction. We hope that this survey will inspire fellow researchers to devote their efforts on investigating FL in the healthcare domain, targeting the deployment of scalable, equitable, and clinically reliable FLAg strategies.