Title: ODAC: learning from local to global - an efficient algorithm for integrating time to event healthcare system data

We propose a One-shot Distributed Algorithm to conduct multi-center survival analysis via the Cox proportional hazard model (ODAC) without sharing patient-level information across sites. Within each clinical site, we construct a surrogate likelihood function which is considered as an approximation of the combined Cox partial likelihood function when patient-level data are pooled together. The surrogate likelihood function is constructed using the patient-level data from a site and summary-level information from other sites. Each site can obtain a local estimator by maximizing its surrogate likelihood function, and the ODAC estimator is constructed as a weighted average of all the local estimators. The performance of the proposed ODAC algorithm is evaluated by a simulation study and an application to four datasets in the Observational Health Data Sciences and Informatics (OHDSI) network. In both simulation and real data studies, ODAC achieves almost identical estimation results compared to the pooled estimator. The proposed ODAC algorithm can achieve result close to the pooled estimator and greatly reduced the bias of the meta estimator caused by rare outcomes. Our algorithm is privacy-preserving and communication-efficient.

Jason a postdoctoral fellow in the Department of Biostatistics, Epidemiology and Informatics at UPenn. Before joining UPenn, he got his PhD degree of Statistics from University of Connecticut. and is currently working on healthcare data analysis (EHR, genetics, pharmacovigilance, meta analysis, etc). He's interested in developing novel methodologies for statistical inference, missing data, distributed learning, dimension reduction/variable selection.