betaFeedback

WorksMathematics

GENMETA-Lite: A Fast and Flexible Approach to Federated Learning with Disparate Covariate Information Across Studies

Abstract

Meta-analysis is foundational to federated learning, where one learns parameters by combining information across studies without sharing individual-level data. Handling missing and disparate covariates across studies is a major issue due to potential introduction of bias and decreased efficiency. We previously proposed a generalized meta-analysis estimator (GENMETA) to build a unified regression model using information on parameter estimates from sub-models and a reference population to estimate the joint covariate distribution. Building on that work, we present a linear estimator (GENMETA-Lite) as a practical, non-iterative method for multivariate meta-analysis, extended to handle heterogeneity in covariate distributions across studies, coarsened covariates, and model selection. Simulations demonstrate that our estimator has robust performance compared to estimates from naïve inverse-variance weighted meta-analysis or multiple imputation with individual-level data (approximately 4 to 100-fold reduction in mean squared error), even when underlying assumptions of linearity and homogeneous distributions are violated. We apply the method to fit a breast cancer risk prediction model with data from five studies (N = 313,851). Our study shows GENMETA-Lite is a robust federated learning approach for building regression models by combining information across disparate studies. An R package implementing these methods is made publicly available.

Paper details

vlnc:W_63bdapdc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…