ReFIT: Federated Transfer Learning for Sequential Prediction and Uncertainty Quantification Using Streaming EHR Data.

Modern biomedical data are increasingly collected across multiple institutions and time periods, creating opportunities for improved statistical inference through knowledge transfer, but also posing challenges for privacy, scalability, and distributional heterogeneity. We propose a Renewable Federated Incremental Transfer framework, termed ReFIT, for sequentially integrating information from streaming source datasets to improve model estimation and prediction in a target population with limited samples. ReFIT builds upon a density ratio model to account for covariate shift between the source and target populations and employs a renewable updating strategy that allows model parameters to be incrementally refined as new source data become available, using only summary-level information from prior sources. This framework ensures privacy preservation and computational efficiency while adapting to evolving data environments. Beyond improving predictive performance, ReFIT also quantifies predictive uncertainty within a conformal prediction framework, yielding valid prediction intervals that adapt as new information accumulates. Extensive simulation studies demonstrate that ReFIT achieves higher predictive accuracy and better uncertainty quantification than models trained on target or source data alone. The method remains robust under nonlinear model misspecification and varying degrees of source-target shift. Moreover, as ReFIT incrementally integrates additional source data, the conformal prediction intervals become progressively narrower without sacrificing coverage, evidencing improved statistical efficiency with growing information. In an electronic health record application for breast cancer prediction, ReFIT substantially improves prediction for the Hispanic population by sequentially leveraging information from non-Hispanic White patients collected over multiple time periods. These results highlight the potential of ReFIT as a general and practical framework for privacy-preserving, adaptive, and scalable learning from distributed and periodically updated biomedical data.