Fig. 1: Federated Target Trial Emulation with Distributed Observational Data for Treatment Effect Estimation.

a Selection Flowchart for Federated Learning-based Target Trial Emulation (FL-TTE). The study cohorts were from five sites within INSIGHT CRN and 192 sites in eICU and MIMIC-IV database, with applications of estimating different drug repurposing signals for Alzheimer’s disease and sepsis, respectively. b Overview of the FL-TTE Framework. Step 1: Cohorts were constructed from INSIGHT and eICU-MIMIC datasets, respectively. Step 2: Federated propensity score calculation adjusted for differences in patient covariates between treated and control groups with inverse probability of treatment weighting (IPTW) for achieving the global covariate balancing. Step 3: Federated Cox proportional hazards model estimated the treatment effects of target drugs for achieving less-biased global time-to-event outcome estimates. The optimizations are regularized by the proximal term which can ensure local updates align with the global model, limit the impact of over-large local updates that can induce overfit, and finally address the data heterogeneity among sites.