194 Privacy-Preserving Federated Learning (FL) Across the Modern International Data Sets of TRACK-TBI and CENTER-TBI: Developing the IMPACT-FL Prognostic Model

INTRODUCTION: Early prognosis in TBI is essential to support clinical decision-making, inform patients and relatives, and advance research. The IMPACT prognostic model for 6-month outcomes was originally developed from an international dataset of moderate-severe TBI patients (GCS =12; studies conducted 1984-1997). Multiple external validations have been performed but global data privacy regulations have made international sharing of human subject data largely infeasible, preventing validation and updating in combined contemporary datasets. METHODS: We implemented a data infrastructure that enables privacy-preserving analyses across the large-scale US-based TRACK-TBI and European CENTER-TBI observational datasets. Subject data were stored separately on secure servers in San Francisco, US and Leiden, the Netherlands respectively. Using local instances of Opal, an open-source data warehouse software, and DataShield, an R-programming framework for FL, analyses were iteratively performed locally and only the abstracted results were shared with the remote users. RESULTS: Using this infrastructure, we first created study-specific IMPACT models, fitted locally using logistic regression. A between study joint IMPACT-FL core model was derived by pooling estimates of the study-specific models. In addition, an iterative IMPACT-FL model was trained, where estimates are jointly learned across studies. The across studies IMPACT-FL model weighted predictors as pooling estimates from study-specific models. CONCLUSIONS: We demonstrated that privacy-preserving statistical analyses across international studies can be executed without transferring data between study centers.