Comparing pooled data, meta-analysis, and federated learning approaches to emulate EHR-based target trials across health systems

Max I Sunog, Colin Magdamo, Marie-Laure Charpignon, Jeffrey G. Klann, Hbid youssef, Antonio Surcel, Haoyang Li, Chengxi Zang, Mohit Manoj Sharma, Anish Kasam, Thibaut Horel, Ran Abuhasira, Munther A. Dahleh, Deborah Blacker, Fei Wang, Ioanna Tzoulaki, Michael Hogarth, Bella Vakulenko-Lagun, Mark W. Albers

Keywords: Federated Learning, Target Trial Emulation, Meta-analysis, Enclaves, Dementia

TL;DR: We assess federated methods and meta-analysis for target trial emulation using pooled data from two EHR sources in an enclave.

Track: Findings

Abstract: Many forms of healthcare research, including studies using the target trial emulation (TTE) framework, rely on data from the electronic health record (EHR). Because data from multiple EHRs often cannot be combined, studies that require multiple data sources often collaborate by combining independent analyses via meta-analysis (MA). However, MA can be ineffective when there is high heterogeneity among sites or when the outcome of interest is rare, two common scenarios in TTE. Alternatively, de-identified data from multiple health systems have been aggregated in secure enclaves. While valuable, the setup and maintenance of these platforms present significant administrative challenges, and restrictions on uploading sensitive data hinders their utility. To address these limitations, federated learning (FL) methods facilitate collaboration when MA and data enclaves are insufficient. To illustrate the advantages of FL, we used data in the ENACT enclave from EHRs in Massachusetts and California to empirically compare TTE results obtained via MA and FL to reference results derived from pooled data. FL consistently produced results closer to the reference than MA, with a larger effect for rarer outcomes. These findings motivate the creation of DRIAD-FL, our platform to expand methods for federating TTEs across five diverse health systems located in the US, the United Kingdom, and Israel.

General Area: Applications and Practice

Specific Subject Areas: Survival Analysis, Causal Inference & Discovery

Data And Code Availability: Yes

Ethics Board Approval: Yes

Entered Conflicts: I confirm the above

Anonymity: I confirm the above

Submission Number: 248