Heterogeneous Datasets for Federated Survival Analysis Simulation

Alberto Archetti, Eugenio Lomurno, Francesco Lattari, André Martin, Matteo Matteucci

Abstract

Survival analysis studies time-modeling techniques for an event of interest occurring for a population. Survival analysis found widespread applications in healthcare, engineering, and social sciences. However, the data needed to train survival models are often distributed, incomplete, censored, and confidential. In this context, federated learning can be exploited to tremendously improve the quality of the models trained on distributed data while preserving user privacy. However, federated survival analysis is still in its early development, and there is no common benchmarking dataset to test federated survival models. This work provides a novel technique for constructing realistic heterogeneous datasets by starting from existing non-federated datasets in a reproducible way. Specifically, we propose two dataset-splitting algorithms based on the Dirichlet distribution to assign each data sample to a carefully chosen client: quantity-skewed splitting and label-skewed splitting. Furthermore, these algorithms allow for obtaining different levels of heterogeneity by changing a single hyperparameter. Finally, numerical experiments provide a quantitative evaluation of the heterogeneity level using log-rank tests and a qualitative analysis of the generated splits. The implementation of the proposed methods is publicly available in favor of reproducibility and to encourage common practices to simulate federated environments for survival analysis.

Introduction

Survival analysis—predicting the time until an event like a mechanical failure or a medical complication—is critical in fields where data is often siloed and highly confidential. While Federated Learning (FL) offers a way to train models on this distributed data without compromising privacy, the field lacks standardized, realistic datasets for testing.

In real-world scenarios, data is rarely "Identically and Independently Distributed" (IID). For instance, one hospital might have thousands of patients (quantity skew), while another might specialize in high-risk cases with much shorter survival times (label skew). Most existing FL simulations use uniform data splits that fail to reflect these complexities. Without a method to simulate this heterogeneity, the performance of federated survival models in diverse, real-world environments remains uncertain.

Methodology: Dirichlet-Based Splitting

This work introduces a reproducible framework for generating heterogeneous federated datasets from existing centralized survival analysis data. We propose two splitting algorithms based on the Dirichlet distribution, using a similarity parameter ($\alpha$) to control the degree of skew.

Quantity-Skewed Splitting

This algorithm assigns a different number of samples to each client. By tuning $\alpha$, researchers can simulate a federation where some clients are "data-rich" and others are "data-poor." As $\alpha \to \infty$, the distribution becomes uniform; as $\alpha \to 0$, the imbalance becomes extreme.

Distribution of sample sizes across clients for varying levels of quantity skew (α).

Label-Skewed Splitting

Label skewing in survival analysis is more complex than in standard classification. We divide the survival timeline into discrete "bins" and use the Dirichlet distribution to ensure different clients receive varying proportions of samples from these bins. This allows for the simulation of clients with distinct survival profiles (e.g., one client having mostly early-event samples, while another has mostly censored data).

Experimental Evaluation

The splitting techniques were applied to five well-known centralized survival datasets: GBSG2, METABRIC, AIDS, FLCHAIN, and SUPPORT. These datasets vary in size (686 to 9,105 samples) and censoring rates (28% to 68%), providing a diverse testing ground for the framework.

Visualizing Heterogeneity

The impact of the $\alpha$ parameter is clearly visible through Kaplan-Meier plots, which visualize survival probability over time. As $\alpha$ decreases, the survival curves for different clients diverge significantly, mimicking real-world silos where patient outcomes vary by location or demographic.

Survival curves (Kaplan-Meier plots) for different clients as the heterogeneity parameter α decreases.

Quantitative Insights

We introduced a heterogeneity score ($h$), representing the percentage of client pairs with statistically significant differences in their survival distributions (determined via log-rank tests).

  • Quantity Skew vs. Distribution Skew: Interestingly, quantity skew alone does not significantly alter underlying survival curves, with $h$ scores remaining below 5%. This suggests quantity skew tests resilience to imbalance rather than population differences.
  • Label Skew Impact: In contrast, label skewing dramatically increases heterogeneity. For the SUPPORT dataset with 5 clients, the $h$ score rose from 0.5% ($\alpha=1000$, near-uniform) to 94.0% ($\alpha=0.1$, highly heterogeneous).

Conclusion

This framework provides a robust tool for benchmarking federated survival models under realistic conditions. While currently focused on right-censoring, future work could extend these algorithms to handle left- or interval-censoring and incorporate system-level heterogeneity (e.g., hardware differences) for more comprehensive simulations. The implementation is publicly available to encourage standardized practices in the field.

Summary updated

Abstract

Survival analysis studies time-modeling techniques for an event of interest occurring for a population. Survival analysis found widespread applications in healthcare, engineering, and social sciences. However, the data needed to train survival models are often distributed, incomplete, censored, and confidential. In this context, federated learning can be exploited to tremendously improve the quality of the models trained on distributed data while preserving user privacy. However, federated survival analysis is still in its early development, and there is no common benchmarking dataset to test federated survival models. This work provides a novel technique for constructing realistic heterogeneous datasets by starting from existing non-federated datasets in a reproducible way. Specifically, we propose two dataset-splitting algorithms based on the Dirichlet distribution to assign each data sample to a carefully chosen client: quantity-skewed splitting and label-skewed splitting. Furthermore, these algorithms allow for obtaining different levels of heterogeneity by changing a single hyperparameter. Finally, numerical experiments provide a quantitative evaluation of the heterogeneity level using log-rank tests and a qualitative analysis of the generated splits. The implementation of the proposed methods is publicly available in favor of reproducibility and to encourage common practices to simulate federated environments for survival analysis.
BibTeX
@inproceedings{archetti2023heterogeneous,
  author = {Alberto Archetti and Eugenio Lomurno and Francesco Lattari and André Martin and Matteo Matteucci},
  title = {{Heterogeneous Datasets for Federated Survival Analysis Simulation}},
  booktitle = {ICPE (Companion)},
  year = {2023},
  doi = {10.1145/3578245.3584935},
}
Figure