Presented by Yuanlin Feng
Masters Candidate in Biostatistics
Plan B Adviser: Dr. Jue Hou

Risk models developed on large research cohorts must be validated on external healthcare systems before guiding local clinical practice, yet validation sites often cannot share individual-level data and repeated interactive evaluation is often impractical. We propose a privacy-preserving, communication-efficient federated validation framework to assess the transferability of risk models to external populations using only low-dimensional summaries computed at validation sites. Each site shares the scalar value, gradient, and Hessian of a possibly smoothed target functional evaluated at its locally fitted model. From these summaries, a quadratic surrogate can be constructed to rapidly approximates external performance for many candidate models without repeated site-side re-evaluation or iterative exchanges. Simulations under varying degrees of distributional shift show that the surrogate closely tracks site-computed the area under the receiver operating characteristic curve (AUC) and generally preserves cross-site ranking, with degraded agreement under weak predictive signal. We illustrate the framework by validating rheumatoid arthritis risk models trained on the All of Us Research Program across geographically distributed validation sites.