Federated Clinical Meta-Learning for Personalized Treatment-Response Prediction in Heterogeneous Oncology Data

Abstract

Artificial intelligence can support treatment-response prediction in oncology, but centralized learning is constrained by institutional data heterogeneity and restrictions on sharing patient-level records. This study presents a federated meta-learning framework that combines learnable representations of mixed clinical variables with client-level adaptation and a lightweight personalization mechanism. De-identified oncology records from the Genomic Data Commons are used to simulate a multi-client environment in which raw records remain local and only model updates are aggregated. The framework is intended to improve adaptation under non-independent and identically distributed client data while preserving data locality. The experimental conclusions, including comparative accuracy, area under the receiver-operating-characteristic curve, and communication efficiency, must be finalized from the archived evaluation outputs before submission. The current evidence therefore supports the proposed architecture and evaluation protocol, but not yet a definitive numerical superiority claim.