Skip to content

federated-survival

federated-survival is an extensible Python toolkit for running multiple survival objectives through one federated execution path. The current release provides adapters for CoxPH, DeepSurv, CoxCC, CoxTime, LogisticHazard, PC-Hazard, and DeepHit. Four protocol adapters separate the communication and server-update rule from model-specific behavior: FedAvg, FedProx, FedOpt, and a linear-Cox-only WebDISCO-style risk-set-statistics protocol.

The software contribution is the common model contract, high-level API, validation, and reproducible experiment infrastructure. FSA-MVAE is retained as an optional augmentation component and is not presented as a new method in the software package.

Installation check

pip install federated-survival
federated-survival doctor

Design goals

  • one public workflow across heterogeneous survival objectives;
  • explicit, testable model-specific adapters;
  • explicit, testable federated protocol adapters;
  • fail-fast validation of invalid survival predictions;
  • machine-readable configurations, metrics, failures, and manifests;
  • one YAML workflow for real-data loading, client partitioning, paired Center/Federated/Local experiments, and automatic figures;
  • backward compatibility with FSAConfig and FSARunner.

Compose an experiment

For real or simulated data, partitioning, paired Center/Federated/Local comparisons, and selectable plots without editing YAML, see Composable Python experiments. For MVAEC/MVAES comparison plots and experimental update-perturbation configuration, see Augmentation and privacy calls. For recorded experiments with a manifest, see Configuration-driven experiments.

Privacy scope

The optional update-perturbation utilities do not currently provide a formal record-level, multi-round differential-privacy guarantee. See the privacy page before using them with sensitive data.