Federated ADMM-IRLS: A privacy-preserving framework for high-dimensional multicenter genomic analysis with binary outcomes

Many clinically relevant multicenter genomic studies involve binary outcomes and high-dimensional data distributed across multiple institutions. However, standard sparse logistic regression (LR) methods typically require centralized access to patient-level data, which is often infeasible under privacy constraints. To address this limitation, a privacy-preserving federated ADMM-IRLS framework is proposed for sparse LR. In the proposed approach, local iteratively reweighted least squares updates are combined with ADMM-based global consensus and soft-thresholding, so that distributed coefficient estimation and variable selection are enabled without sharing raw patient-level data. The proposed method is evaluated through simulation studies and a real-data application to multicenter lung adenocarcinoma (LUAD) genomic data, in which GEO-LUAD and TCGA-LUAD are used as training centers and ORIEN is used as an external validation cohort. Across both moderate-dimensional and high-dimensional settings, the proposed federated ADMM-IRLS method performs very close to the pooled benchmark in predictive accuracy and coefficient recovery while generally outperforming the single-center models. In the real-data application, an apparent (in-sample) training AUC of 0.725 is achieved by the proposed method, compared with 0.735 for the pooled benchmark, whereas the corresponding AUC values for the single-center models are 0.695 and 0.705. On the independent ORIEN test set, an AUC of 0.752 is achieved by the proposed method, remaining close to the pooled benchmark (0.759) while outperforming LR-1 (0.609) and LR-2 (0.733). In addition, 13 of the 14 genes selected by the pooled benchmark are recovered by the proposed method, indicating strong concordance in gene selection without raw-data sharing across centers. The methodological contribution of the proposed framework is the integration of center-specific IRLS optimization for the nonlinear logistic likelihood, ADMM-based consensus estimation, and global LASSO regularization within a single federated optimization procedure, enabling sparse high-dimensional logistic regression across institutions while preserving patient-level data locality during iterative model fitting.