FairFedMed: Fair Federated Medicine

• FairFedMed is a holistic approach in federated learning for medicine that integrates client-level, group, and participation fairness to address heterogeneous data challenges.
• It employs privacy-preserving methods combined with fairness-aware aggregation, uncertainty calibration, and robust optimization to balance performance and equity.
• By mitigating disparities across institutions and patient subgroups, FairFedMed aims to enhance clinical outcomes and trust in collaborative medical AI.

FairFedMed denotes a fairness-centric view of federated learning for medicine in which hospitals, clinics, or other health systems collaboratively train models without sharing raw patient data while explicitly controlling disparities across institutions or patient subgroups. Recent work uses the term as a building block or organizing label for privacy-preserving medical FL that combines client-level fairness, group fairness, robustness to domain shift, trust-aware participation, and fairness-aware uncertainty quantification (Li et al., 2024, Wang et al., 2024, Wu et al., 12 Mar 2025, Kahenga et al., 23 Sep 2025). Taken together, these works suggest that FairFedMed is best understood as an umbrella research program rather than a single fixed algorithm.

1. Conceptual scope

The starting point is standard federated learning, in which the server minimizes an average objective

$\min_w F(w) = \sum_{i=1}^N p_i F_i(w),$

with $F_i(w)$ the client-specific loss and $p_i$ aggregation weights. In medical FL, this objective is routinely stressed by non-IID data, because hospitals differ in patient demographics, outcome prevalence, imaging hardware, acquisition protocols, and local practice. In that regime, a global model can perform well on sites close to the dominant mixture while underperforming on minority or atypical sites (Wang et al., 2024).

FairFedMed spans at least three fairness axes. The first is client-level performance fairness, where the concern is whether institutions receive comparable utility from the shared model. The second is algorithmic group fairness, where the concern is whether patient subgroups such as sex, race, or age bands receive comparable predictive quality across the federation. The third is participation and representation fairness, where the concern is whether smaller or noisier institutions are systematically excluded from optimization and thus from the benefits of collaboration (Kahenga et al., 23 Sep 2025).

A central conceptual distinction is that fairness across institutions is not the same as fairness across demographic groups. FedMinMax formalizes group-fair FL through

$$\min_{h \in \mathcal{H}} \max_{a \in \mathcal{A}} r_a(h),$$

and proves that client-fairness does not generally imply group-fairness unless special structural conditions on client group distributions hold (Papadaki et al., 2021). In medical FL, this distinction is consequential because hospitals can differ sharply in subgroup composition; a system can equalize hospital-level performance and still leave one demographic group systematically worse off overall.

Medical FL also introduces fairness problems tied to domain shift and quality shift. Fed-LWR treats domain shift as feature shift across hospitals, estimated through layer-wise differences in learned representations (Yan et al., 2024). FedISM+ instead focuses on quality shift, where a minority of imaging clients hold corrupted or lower-quality data and standard FL biases the global model toward the more common high-quality images (Wu et al., 12 Mar 2025). These formulations broaden FairFedMed beyond purely demographic fairness.

2. Fairness objectives and metrics

Client-level fairness is often operationalized as uniformity of performance across clients. FedMABA defines fairness through the variance of client losses,

$$Var(F)= \sqrt{\frac{1}{N}\sum_{i=1}^{N}\left(F_{i}(w)-\frac{1}{N}\sum_{j=1}^{N}F_{j}(w)\right)^2},$$

and studies the constrained problem

$$\min_{w} \quad \frac{1}{N}\sum_{i=1}^{N} F_i(w) \quad \text{s.t.} \quad Var(F)\le \rho.$$

In experiments it evaluates fairness through the variance of client accuracies, together with worst-5% and best-5% client accuracies (Wang et al., 2024).

Group fairness in medical FL is frequently expressed through demographic parity and equalized odds. FairFML adopts the standard definitions

$$P[\hat{Y} = 1 \mid X_{d+1} = a] = P[\hat{Y} = 1], \quad \forall a$$

for demographic parity, and

$$P[\hat{Y} = 1 \mid X_{d+1} = a, Y = y] = P[\hat{Y} = 1 \mid Y = y], \quad \forall a, y$$

for equalized odds. It evaluates four derived metrics: Demographic Parity Difference (DPD), Demographic Parity Ratio (DPR), Equalized Odds Difference (EOD), and Equalized Odds Ratio (EOR) (Li et al., 2024).

FairFML and FedIDA use a convex fairness penalty originating in Berk et al. For two groups $S_1$ and $S_2$, FairFML defines

$F_i(w)$0

with

$F_i(w)$1

and optimizes the fairness-regularized loss

$F_i(w)$2

FedIDA uses the same convex penalty and extends it to multiple sensitive attributes and intersectional groups through fairness-aware regularization plus group-conditional oversampling (Li et al., 2024, Wu et al., 14 May 2025).

A more explicitly Rawlsian formulation appears in FedISM+, which targets Max-Min fairness in expected loss across latent image-quality groups: $F_i(w)$3 Because image-quality labels are not observable, the paper proves that client-level Max-Min fairness is equivalent under mild assumptions (Wu et al., 12 Mar 2025).

3. Algorithmic families

FairFedMed methods differ less in their high-level motivation than in the layer at which fairness is enforced: the local objective, the server aggregation rule, the client-selection policy, or the uncertainty-calibration layer. The current literature covers all four.

Method	Fairness target	Core mechanism
FairFed (Ezzeldin et al., 2021)	Group fairness	Server-side aggregation weights updated from local/global fairness gaps
FedMinMax (Papadaki et al., 2021)	Demographic worst-group risk	Federated min-max optimization over group weights $F_i(w)$4
FairFML (Li et al., 2024)	Algorithmic subgroup fairness	Local fairness-regularized loss integrated into FedAvg or Per-FedAvg
FedIDA (Wu et al., 14 May 2025)	Demographic disparity and imbalance	Fairness-aware regularization plus group-conditional oversampling
FedMABA (Wang et al., 2024)	Client performance fairness	Entropy-constrained adversarial multi-armed bandit reweighting
FedCE (Jiang et al., 2023)	Collaboration fairness and performance fairness	Contribution estimation in gradient space and data space
Fed-LWR (Yan et al., 2024)	Site-level fairness under domain shift	Layer-wise CKA-based re-aggregation from feature shift
FedISM+ (Wu et al., 12 Mar 2025)	Client fairness under quality shift	Progressive state matching with SALT and SAGA
FedFiTS (Kahenga et al., 23 Sep 2025)	Participation fairness and trustworthiness	Fitness-selected, slotted client scheduling
FedCF (Srinivasan et al., 26 Sep 2025)	Fair coverage in uncertainty sets	Federated conformal fairness via threshold optimization

Aggregation-centric methods alter how the server combines client models. FairFed updates client weights according to each client’s deviation from the global fairness metric, remaining agnostic to the local debiasing method (Ezzeldin et al., 2021). FedMABA instead models clients as arms in an adversarial multi-armed bandit and derives weights from an entropy-constrained robust objective (Wang et al., 2024). FedCE estimates client contribution in gradient space through directional novelty and in data space through validation error of a model excluding that client; those contribution estimates become aggregation weights (Jiang et al., 2023). Fed-LWR moves the same idea to representation space, assigning higher layer-specific weights to hospitals whose local and global feature representations differ more strongly (Yan et al., 2024).

Objective-centric methods modify the loss used during local training. FairFML inserts a convex fairness regularizer into the local loss while leaving server aggregation unchanged (Li et al., 2024). FedIDA adds the same kind of fairness regularizer but couples it with FairnessAwareROSE, which balances local data across sensitive-attribute–outcome groups (Wu et al., 14 May 2025). FedISM+ changes both local optimization and aggregation: locally it performs sharpness-aware training at a progressive search radius $F_i(w)$5, and globally it weights clients by sharpness or perturbed loss so that worse-off clients receive more influence (Wu et al., 12 Mar 2025).

Selection-centric methods treat fairness as a participation problem. FedFiTS operationalizes trustworthiness and participation fairness through a three-phase protocol—Free-for-All, Natural Selection, and Slotted Team Participation—with a client fitness score

$F_i(w)$6

and a threshold

$F_i(w)$7

This does not implement explicit demographic fairness constraints, but it directly addresses starvation of smaller or minority sites (Kahenga et al., 23 Sep 2025).

Finally, uncertainty-aware methods extend fairness from point predictions to prediction sets. FedCF transports Conformal Fairness to federated learning by computing fairness-specific coverage bounds and then selecting the smallest conformal threshold $F_i(w)$8 whose coverage gap across groups is at most a user-specified closeness parameter $F_i(w)$9 (Srinivasan et al., 26 Sep 2025).

4. Clinical case studies and empirical record

The most explicit FairFedMed-style case study is FairFML’s out-of-hospital cardiac arrest program. Using 7,425 eligible cases from the Resuscitation Outcomes Consortium Epistry-Cardiac Arrest registry, partitioned into four simulated cross-site scenarios, FairFML reports that it improves model fairness by up to 65% compared to the centralized model while maintaining performance comparable to both local and centralized models as measured by receiver operating characteristic analysis. Across all scenarios, the maximum drop in mean AUC for FairFML versus the central model is reported as $p_i$0 (Li et al., 2024).

In medical image segmentation, Fed-LWR reports strong gains on two standard cross-silo benchmarks. On ProstateMRI, FedAvg yields Avg Dice 88.78 and Std 4.54, whereas Fed-LWR reaches Avg Dice 93.25 and Std 0.97. On RIF, FedAvg yields Avg 84.63 and Std 7.79, while Fed-LWR reaches Avg 87.17 and Std 5.08. The paper attributes these improvements to feature-shift-aware, layer-wise re-aggregation (Yan et al., 2024).

FedCE studies retinal fundus segmentation and prostate MRI segmentation with six clients in each task. On retinal segmentation, FedAvg achieves average Dice 78.27 with std 18.66, while FedCE (Multi.) reaches average Dice 83.08 with std 12.70. On prostate MRI, FedAvg achieves average Dice 87.82 with std 2.90, while FedCE (Multi.) reaches average Dice 88.31 with std 2.22. FedCE also reports substantially better agreement between estimated client contributions and leave-one-out “ground truth” contributions, with Pearson correlation above 0.94 in its main analyses (Jiang et al., 2023).

FedISM+ addresses imaging fairness under corruption rather than hospital identity per se. On RSNA ICH and ISIC 2019, it improves corrupted-client performance and reduces the standard deviation of client-level AUC to 1.39%, compared with 3.23% for FedAvg, 1.49% for the best fair-optimization baseline, and 1.59% for FedISM. Its reported gains come from progressively matching multiple convergence states rather than only training loss or sharpness at a single radius (Wu et al., 12 Mar 2025).

FedFiTS supplies a complementary trustworthy-healthcare result on X-ray pneumonia. For $p_i$1 clients, the method reports normal-mode accuracy 0.968 versus 0.946 for FedRand and 0.950 for FedPow, and attack-mode accuracy 0.940 versus 0.913 for both baselines. It also reports client participation ratios rising from 45% for FedAvg and 52% for FedPow to 82% for FedFiTS with dynamic $p_i$2, positioning participation breadth as a fairness proxy (Kahenga et al., 23 Sep 2025).

5. Privacy, uncertainty, and evaluation infrastructure

A defining property of FairFedMed is that fairness interventions must coexist with privacy constraints rather than assume centralized access to sensitive attributes or raw data. FairFML keeps all OHCA cases local and only exchanges model parameters or gradients (Li et al., 2024). FedMABA notes that its server needs only scalar client loss values in addition to standard FL communication (Wang et al., 2024). FedISM+ adds only one scalar per client per round for $p_i$3 or $p_i$4, making the communication overhead effectively identical to ordinary FL for ResNet-18-scale imaging models (Wu et al., 12 Mar 2025). A healthcare-specific benchmark methodology for fairness-aware and privacy-preserving FL additionally evaluates AES-GCM encryption and reports that doing so under privacy-preserving conditions can create significant increases in network communication cost and latency compared to the typical federated learning scheme (Annapareddy et al., 2023).

Uncertainty calibration adds another layer. FedCF starts from the standard conformal guarantee

$p_i$5

and extends Conformal Fairness to federated learning by estimating fairness-specific conditional coverage for each group $p_i$6 and label $p_i$7, then selecting the smallest threshold $p_i$8 whose coverage gap across groups is at most $p_i$9. The same machinery can be used as a federated fairness auditor: clients retain their data locally and send only summary terms needed to evaluate the fairness-related coverage gap of a deployed conformal predictor (Srinivasan et al., 26 Sep 2025).

Evaluation methodology is itself part of the FairFedMed problem. MEDFAIR standardizes medical-imaging fairness assessment across 10 public datasets, 11 algorithms, and three model-selection criteria, and shows that the under-studied issue of model selection can have a significant impact on fairness outcomes, while state-of-the-art bias mitigation algorithms do not significantly improve fairness outcomes over empirical risk minimization in both in-distribution and out-of-distribution settings (Zong et al., 2022). This finding is directly relevant to FairFedMed because it implies that federated fairness claims should not be interpreted independently of checkpoint selection, validation protocol, and test-shift design.

6. Limitations and open problems

Taken together, the literature indicates that FairFedMed remains fragmented across fairness definitions. Some methods optimize client-level variance or worst-client risk (Wang et al., 2024, Wu et al., 12 Mar 2025); others target demographic parity or equalized odds (Li et al., 2024, Wu et al., 14 May 2025); others still focus on participation fairness and trustworthiness without explicit demographic guarantees (Kahenga et al., 23 Sep 2025). No single method in the provided literature simultaneously covers demographic fairness, institution-level fairness, trustworthy selection, and calibrated uncertainty.

Several limitations recur. FedMABA explicitly notes that malicious clients could misreport their losses to gain more influence, which is especially problematic when higher loss yields higher aggregation weight (Wang et al., 2024). FairFML currently focuses on a single sensitive attribute, gender, even though its own subgroup analysis shows that gender disparities vary substantially across race groups (Li et al., 2024). FedMinMax assumes

$$\min_{h \in \mathcal{H}} \max_{a \in \mathcal{A}} r_a(h),$$0

so clients differ only in group mixtures and not in the group-conditional data distribution itself; that assumption can be restrictive in real multi-hospital systems (Papadaki et al., 2021). FedISM+ addresses quality shift at the client level, but if a single hospital contains both very clean and very corrupted images, intra-client fairness is not directly addressed (Wu et al., 12 Mar 2025). FedCF becomes conservative when calibration slices are small, and its guarantees depend on exchangeability between calibration and deployment data (Srinivasan et al., 26 Sep 2025).

Benchmarking results also warn against overinterpreting nominal fairness improvements. MEDFAIR finds that model-selection strategy alone can materially change fairness outcomes (Zong et al., 2022). The healthcare FAFL benchmark reports considerable variation in how different fairness-aware schemes respond to high heterogeneity and shows that privacy-preserving deployment can impose substantial communication and latency costs (Annapareddy et al., 2023). A plausible implication is that future FairFedMed systems will need joint optimization over fairness, robustness, privacy, calibration, and systems cost rather than treating fairness as a single post hoc metric.

The trajectory of the field points toward that broader synthesis. Recent work already sketches the ingredients: explicit group-fair objectives (Papadaki et al., 2021), server-side fairness-aware aggregation (Ezzeldin et al., 2021), fairness-regularized local optimization (Li et al., 2024), imbalance-aware subgroup handling (Wu et al., 14 May 2025), representation- and geometry-aware fairness under medical shift (Yan et al., 2024, Wu et al., 12 Mar 2025), trustworthy scheduling (Kahenga et al., 23 Sep 2025), and federated fairness-aware uncertainty auditing (Srinivasan et al., 26 Sep 2025). In that sense, FairFedMed names an emerging architecture for medical FL in which equity is treated as a first-class systems and optimization constraint rather than as an afterthought.