Fair Federated Survival Analysis

Md Mahmudur Rahman, Sanjay PurushothamAAAI 2025 · DOI

TL;DR

This paper introduces FairFSA, a novel framework for fair federated survival analysis that addresses bias in predictions arising from demographic and censoring disparities across decentralized datasets. By combining federated pseudo values with distributionally robust optimization and a fairness-aware aggregation strategy, FairFSA achieves equitable and accurate survival predictions without requiring direct data sharing.

What Problem Does This Paper Address?

Survival analysis is crucial in healthcare for predicting time-to-event outcomes like patient recovery or disease progression. However, training robust survival models is challenging due to censored data (where the event of interest isn't observed) and privacy concerns that limit data sharing. Federated Survival Analysis (FSA) offers a solution by enabling collaborative model training across multiple institutions without sharing raw data.

Despite FSA's promise, a significant challenge is the potential for bias in survival predictions. This bias can stem from disparities in sensitive attributes (like race or age) and censoring rates across different clients' datasets. These disparities can lead to unfair outcomes, particularly for underrepresented groups. Existing fairness approaches in survival analysis often require access to the entire dataset or complex pairwise comparisons, making them unsuitable for the decentralized nature of FSA. Furthermore, standard federated aggregation methods like FedAvg can amplify existing biases by giving more weight to clients with larger datasets, potentially neglecting the performance of smaller clients.

Methods

FairFSA tackles these challenges through a multi-pronged approach:

  • Federated Pseudo Values (FPV): To handle censored data and enable client-level regression, FairFSA reformulates survival analysis using FPVs. These pseudo values effectively represent the observed survival function, allowing for the use of standard regression techniques and making the loss function separable, which is crucial for federated learning.
  • Distributionally Robust Optimization (DRO) for Fairness: FairFSA integrates DRO into the FPV-based objective function. This approach optimizes for the worst-case scenario across different subpopulations within a client's data, ensuring fairness without explicitly requiring sensitive demographic attributes. This makes the framework more broadly applicable.
  • Fairness-Aware Concordance Index (CI)-Weighted Aggregation: To ensure global fairness, FairFSA enhances the traditional FedAvg algorithm. Instead of solely relying on sample size for aggregation, it incorporates a fairness-aware strategy that weights local models based on their validation Concordance Index (C-Index) performance. This ensures that clients with better performance contribute more, promoting equitable outcomes across all participants.

Datasets

The paper evaluates FairFSA on four publicly available real-world survival datasets:

  • FLChain
  • SUPPORT
  • SEER
  • MSK-MET

These datasets were used for both centralized and federated experiments, with a simulated non-IID federated environment created by assigning subjects to clients in a way that skewed event time distributions.

Key Insights and Ideas

  • Fairness in FSA is a critical, under-addressed problem: The paper highlights that existing FSA methods, while addressing privacy, often overlook fairness, which is paramount in healthcare applications.
  • FPVs are key to enabling fairness in FSA: By transforming survival data into a regression-friendly format, FPVs unlock the application of fairness techniques like DRO in a federated setting without needing raw data access.
  • DRO can achieve fairness without explicit sensitive attributes: The proposed PseudoDRO model demonstrates that fairness can be enforced by optimizing for worst-case scenarios across data distributions, even when sensitive attributes are not explicitly provided or are unknown.
  • Global fairness requires more than just local fairness: Simply having fair local models doesn't guarantee a fair global model. The proposed CI-weighted aggregation is crucial for balancing contributions and ensuring equitable performance across all clients.

Key Results

  • Superior Fairness and Accuracy: FairFSA consistently outperformed baseline federated survival analysis methods in both accuracy (measured by C-Index, IBS, MAE) and fairness (measured by Favg) across multiple datasets.
  • Competitive with Centralized Models: The proposed centralized PseudoDRO model also showed strong performance, often matching or exceeding existing fair survival models.
  • Robustness to Censoring: The pseudo-value-based approach (FPV) in FairFSA demonstrated significantly less performance degradation due to censoring compared to non-pseudo-value-based methods.
  • Improved Global Fairness: The CI-weighted aggregation strategy in FairFSA led to a lower standard deviation in performance metrics across clients compared to traditional FedAvg, indicating more equitable global fairness.

Limitations and Future Work

  • Censoring Dependency on Covariates: While FPVs handle non-uniform censoring, the paper notes that covariate-dependent FPVs (which could further address censoring dependency on sensitive attributes) are beyond the scope of this work.
  • Hyperparameter Tuning: The paper mentions selecting the trade-off parameter α for balancing fairness and accuracy, and future work could explore more automated or adaptive methods for this.
  • Calibration of Predicted Probabilities: The authors suggest that future work will investigate calibrating predicted probabilities to enhance Integrated Brier Score (IBS) performance while maintaining fairness.
  • Generalizability to More Complex Federated Settings: While experiments were conducted on non-IID data, exploring FairFSA in more complex federated scenarios with varying network conditions or client availability could be a future direction.

Summary updated

BibTeX
@inproceedings{rahman2025fair,
  author = {Md Mahmudur Rahman and Sanjay Purushotham},
  title = {{Fair Federated Survival Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i19.34214},
}
Figure