Federated Bias Mitigation

Ensures the model's prediction rate is independent of sensitive attributes like race or gender across all client sites.

• Global Constraint: The difference in positive prediction rates between groups is bounded across the entire federated network.
• Local Enforcement: Each client computes group-specific prediction rates locally and shares only aggregate statistics.
• Clinical Example: A diagnostic model must recommend further testing at similar rates for both male and female patients, regardless of which hospital trained the data.
• Implementation: Uses Lagrangian multipliers in the federated optimization objective to penalize violations without exposing individual patient demographics.

Guarantees that the model's error rates—both false positives and false negatives—are balanced across protected groups.

• Dual Constraint: Enforces parity in both true positive rates and false positive rates simultaneously.
• Per-Client Auditing: Each institution computes confusion matrices stratified by sensitive attributes locally.
• Clinical Significance: A sepsis predictor must not systematically miss cases in minority populations while over-alerting for majority groups.
• Federated Aggregation: Secure multi-party computation combines local fairness statistics without revealing patient-level outcomes.

A relaxed variant of equalized odds that focuses exclusively on equalizing true positive rates across groups.

• Recall Parity: Ensures the model is equally likely to correctly identify positive cases for all demographic groups.
• Clinical Relevance: A cancer screening model must achieve similar sensitivity across racial groups to prevent delayed diagnoses in underserved populations.
• Federated Computation: Clients share only group-specific recall values, preserving patient privacy while enabling global fairness auditing.
• Trade-off: Less restrictive than equalized odds, allowing some variation in false positive rates to maintain overall model utility.

Quantifies the ratio of favorable outcomes between a protected group and a reference group, with a threshold of 0.8 typically indicating compliance with the 80% rule.

• Calculation: DIR = P(ŷ=1 | protected_group) / P(ŷ=1 | reference_group).
• Federated Estimation: Each client computes local DIR values; the global metric is aggregated via weighted averaging based on group sample sizes.
• Regulatory Alignment: Maps directly to legal standards for adverse impact in employment and lending, adapted here for clinical decision support.
• Threshold Monitoring: Alerts trigger when any client's DIR falls below 0.8, indicating potential discriminatory model behavior at that site.

Monitors the variance in fairness metrics across different clinical sites to detect localized bias that global averages might mask.

• Metric: Standard deviation of demographic parity or equalized odds across all participating clients.
• Detection: Identifies sites where the global model performs unfairly despite acceptable aggregate fairness scores.
• Root Cause: Often caused by severe label distribution skew or underrepresentation of certain groups in a specific hospital's training data.
• Mitigation: Triggers personalized federated learning strategies or targeted data augmentation for outlier clients.

Ensures that a model's prediction for an individual would remain the same if their sensitive attribute were different, all other features held constant.

• Causal Approach: Builds a structural causal model to generate counterfactual instances for each patient.
• Federated Auditing: Each client generates counterfactual pairs locally and reports the consistency of predictions without sharing the original sensitive attributes.
• Clinical Depth: A treatment recommendation model should prescribe the same intervention regardless of a patient's race, given identical clinical presentation.
• Implementation Complexity: Requires careful modeling of causal relationships between sensitive attributes and clinical features, which may vary across institutions.

A group fairness criterion requiring that a model's positive prediction rate is statistically independent of a protected attribute (e.g., race, gender). In federated learning, achieving this is complicated by the fact that protected attributes may be distributed non-uniformly across clients. A model that appears fair globally can still be locally discriminatory if one hospital's data skews the aggregation. Mitigation strategies often involve constrained optimization during local training or fairness-aware aggregation that re-weights client contributions to balance group outcomes.

A primary driver of bias in federated clinical models. This occurs when the prior probability of class labels varies significantly across clients.

• Example: A cardiac unit has 40% positive cases for heart disease, while a general practice has only 5%.
• Impact: Standard Federated Averaging (FedAvg) can cause the global model to overfit to the majority label distribution, under-serving minority populations at smaller sites.
• Mitigation: Techniques like FedProx (adding a proximal term) or clustered federated learning help stabilize training under severe label skew.

Mathematical constraints integrated into the local training objective to enforce fairness. Common approaches include:

• Lagrangian multipliers: Penalizing the local loss function when fairness violations exceed a threshold.
• Adversarial debiasing: Training a local adversary to predict the protected attribute from the model's representations, then removing that signal via a gradient reversal layer.
• Fair Federated Averaging (FairFed): Adjusting client aggregation weights based on local fairness performance relative to the global model, promoting equity without requiring raw data sharing.

An optimization framework that learns data representations which elicit the same optimal classifier across all training clients. Unlike standard empirical risk minimization, IRM seeks to discover causal relationships robust to spurious correlations.

In bias mitigation, IRM prevents the model from relying on proxy variables (e.g., zip code as a proxy for race) that vary across sites. By enforcing invariance across heterogeneous environments, the model learns genuinely relevant clinical features rather than site-specific artifacts.

The process of quantifying each client's marginal contribution to model performance using game-theoretic concepts like the Shapley value. In the context of bias mitigation, data valuation identifies clients whose data is critical for representing minority groups.

• Fairness-aware valuation: Extends standard valuation to reward clients whose data improves both accuracy and fairness metrics.
• Application: Ensures that clients serving underrepresented populations are not down-weighted or excluded during aggregation, preserving equitable representation in the final model.

A technique that uses a domain discriminator network to enforce fairness across clients. During local training:

1. A feature extractor learns representations of the input data.
2. A label predictor uses these representations to make clinical predictions.
3. A fairness adversary attempts to predict the protected attribute from the same representations.
4. The feature extractor is trained to fool the adversary, removing sensitive information from the learned features.

This results in client-invariant representations that are both useful for the task and free of discriminatory signals.