Federated Target Trial Emulation using Distributed Observational Data for Treatment Effect Estimation.

Area of Science:

• Health Informatics
• Epidemiology
• Machine Learning

Background:

• Target trial emulation (TTE) uses real-world data to simulate clinical trials for treatment effect estimation.
• Applying TTE across distributed datasets is limited by privacy and data-sharing constraints.
• Federated learning offers a solution for collaborative analysis without centralizing sensitive patient data.

Purpose of the Study:

• To propose and validate a Federated Learning-based Target Trial Emulation (FL-TTE) framework.
• To enable privacy-preserving TTE across multiple, distributed, and heterogeneous datasets.
• To overcome limitations of traditional TTE methods in multi-site settings.

Main Methods:

• Developed FL-TTE, a framework integrating federated protocol design, federated inverse probability of treatment weighting, and a federated Cox proportional hazards model.
• Applied FL-TTE to emulate Sepsis trials using eICU and MIMIC-IV data from 192 hospitals.
• Validated FL-TTE on Alzheimer's trials using the INSIGHT Network across five NYC health systems.

Main Results:

• FL-TTE successfully enabled TTE across distributed and heterogeneous data without sharing patient-level information.
• The framework produced less biased treatment effect estimates compared to traditional meta-analysis methods when validated against pooled results.
• FL-TTE demonstrated theoretical support and practical applicability in real-world scenarios.

Conclusions:

• FL-TTE provides a robust and privacy-preserving method for federated treatment effect estimation.
• The framework enhances the generalizability and power of TTE by leveraging distributed data.
• FL-TTE represents a significant advancement for real-world evidence generation in multi-institutional research.