An international multi-centre study to develop and validate federated learning-based prognostic models for anal
Stelios Theophanous1,2, Per-Ivar Lønne3, Ananya Choudhury4
1Leeds Teaching Hospitals NHS Trust, Leeds, UK. stelios.theophanous@nhs.net.
Nature Communications
|March 15, 2026
Summary
Federated learning enables robust prognostic models for rare cancers like anal cancer by training models across multiple institutions without sharing patient data. This approach ensures data privacy and supports international collaboration for precision oncology.
Area of Science:
- Oncology
- Medical Informatics
- Biostatistics
Background:
- Precision oncology requires high-quality data for rare cancer subgroups.
- Federated learning offers a potential solution for data access challenges.
Purpose of the Study:
- To investigate the utility of federated learning for prognostic modeling in anal cancer, a rare cancer.
- To assess the performance of federated multivariable Cox models using real-world data.
Main Methods:
- Trained federated multivariable Cox models across 14 international centers (1428 patients).
- Externally validated models in two additional centers (277 patients).
- Utilized leave-one-centre-out cross-validation for robust assessment.
Main Results:
- Achieved consistent calibration and discrimination (c-indices 0.68-0.79) in validation.
- Identified prognostic factors for overall survival, locoregional control, and distant metastasis.
- Demonstrated the effectiveness of federated learning for rare cancer modeling.
Conclusions:
- Federated learning enables privacy-preserving, robust prognostic modeling for rare cancers.
- Supports international collaboration and real-world data utilization in oncology.
- Facilitates precision oncology by leveraging distributed data resources.
