Paper 2026/2784

Privacy-Preserving Multi-Institutional Clinical Prediction via Multiparty Encrypted Newton-CG Logistic Regression

Abstract

Multi-institutional clinical prediction can benefit from larger and more diverse patient populations, but privacy and data-governance constraints often prevent pooling patient-level records. This study aims to enable collaborative logistic-regression training while keeping such records local. We developed a framework combining multiparty Cheon–Kim–Kim–Song (CKKS) homomorphic encryption with Newton–conjugate gradient optimization. Each institution computes local gradient and damped Hessian contributions, while only encrypted fixed-dimensional statistics are aggregated. The global Newton system is solved by encrypted preconditioned conjugate gradient with diagonal preconditioning and delayed division, making the dominant homomorphic cost depend mainly on model dimension. We evaluated the framework on nine clinical-prediction tasks from NHANES 2017–2018 and on the PIMA Diabetes and Breast Cancer Wisconsin datasets. Across the nine NHANES tasks, mean absolute deviations from pooled-data plaintext Newton training were below 0.002 for area under the receiver operating characteristic curve, area under the precision–recall curve, and average precision. Five-fold cross-validation accuracy was (77.21 ± 1.28)% on PIMA Diabetes and (96.42 ± 1.44)% on Breast Cancer Wisconsin. Under a common 20-iteration budget, the proposed method achieved approximately 3.85× faster training and nearly 90% lower peak memory use than a reproduced multiparty encrypted first-order baseline, while attaining a lower final relative gradient and closer agreement with plaintext Newton. These results show that exploiting locally computed Newton statistics can preserve predictive utility while substantially reducing encrypted training cost, supporting privacy-preserving multi-institutional clinical prediction without pooling patient-level data.