Federated learning improved radiation pneumonitis prediction across lung cancer datasets

A center-specific federated learning model predicted grade 2 or higher radiation pneumonitis with stable cross-dataset performance.

KEY POINTS

  • This retrospective TRIPOD type 3 prediction-model study included 1238 lung cancer patients from 4 datasets, including institutional chemoradiotherapy cohorts and RTOG 0617-derived datasets.
  • The endpoint was symptomatic radiation pneumonitis, defined as grade 2 or higher by Common Terminology Criteria for Adverse Events version 4.0, occurring within 6 months after radiotherapy.
  • The proposed Federated Cross-Center Adaptive Alternating Model used planning computed tomography and dose images, sharing only global feature-extractor weights while keeping center-specific adaptation parameters local.
  • FCAAM achieved stable test area under the curve values across datasets: 0.77, 0.71, 0.76, and 0.75, compared with mean area under the curve 0.58 for single-center models and 0.70 for standard federated averaging.
  • In robustness testing, reducing a client’s training sample size caused lower area-under-the-curve degradation with FCAAM than with federated averaging: 16.4% vs 32.2%.

CLINICAL TAKEAWAY

This study addresses a real radiation oncology problem: toxicity models often fail when moved across centers with different patients, imaging, dose distributions, and systemic therapy patterns. FCAAM improved cross-center stability while preserving data locality, making it technically relevant for future multi-institutional radiation pneumonitis prediction. But the work remains retrospective, lacks a fully independent external institutional validation cohort, and the web platform is explicitly a research prototype rather than a clinical decision-support device.

SOURCE

International Journal of Radiation Oncology, Biology, Physics