Evaluating Therapeutic Outcomes in Childhood Interstitial Lung Disease Through Privacy-Preserving Bayesian Federated Inference

Abstract

Introduction: Childhood interstitial lung disease (chILD) encompasses rare, diffuse lung disorders with high morbidity
and mortality. No pharmacological therapy has received regulatory approval, leaving management dependent on off-
label treatments. The extreme rarity of chILD prevents individual centers from gathering sufficient data for well-powered
studies, while data governance regulations restrict traditional multicenter data pooling.
Objectives: This study aims to establish a privacy-preserving Bayesian Federated Inference (BFI) framework for
treatment stratification in chILD across a multinational network. Specifically, it seeks to: (1) compare the effectiveness of
off-label therapies on clinical trajectories without moving individual patient data; (2) implement local Explainable AI (XAI)
modules to provide interpretable treatment-response stratification by chILD subtype; and (3) evaluate a fixed-noise
differential privacy mechanism to protect shared posterior estimates while maintaining clinical utility.
Methods: A retrospective multicenter cohort study will enroll patients aged 0–18 with a chILD diagnosis and a minimum
of 12 months of follow-up. The BFI framework enables privacy-preserving analysis, where centers compute local
Bayesian hierarchical models and share only summary-level posterior parameter estimates. Federated aggregation
combines these estimates into a global inference. Institutional heterogeneity and non-IID data distributions will be
managed via BFI extensions. A fixed Local Differential Privacy mechanism will add calibrated noise to shared posteriors
prior to aggregation. Local XAI post-hoc modules (SHAP) will generate interpretable explanations of treatment-response
predictions using the global model, with only aggregated feature-importance patterns shared across the network.
Results: A pilot proof-of-concept across European centers is expected to demonstrate the feasibility of collaborative
therapeutic evaluation without centralizing sensitive data. The privacy-protected BFI models are projected to yield
interpretable stratifications of treatment response, with XAI enhancing clinical trust and actionable insight.
Conclusion: Ultimately, this framework establishes a scalable paradigm for collaborative medical AI in rare diseases,
overcoming data fragmentation to directly inform clinical practice while rigorously preserving patient privacy at every
contributing institution.