Federated Tensor Factorization for Computational Phenotyping.

Area of Science:

• Computational health informatics
• Machine learning for healthcare

Background:

• Electronic health records (EHRs) are vast data sources for clinical concept discovery (phenotyping).
• Tensor factorization models require diverse data to avoid bias, but multi-institutional collaboration is hindered by data privacy concerns.
• Deriving phenotypes across hospitals without patient-level data sharing remains a significant challenge.

Purpose of the Study:

• To develop a federated tensor factorization method for computational phenotyping across multiple hospitals.
• To enable collaborative phenotype discovery without compromising patient data privacy.
• To address the limitations of centralized models in multi-institutional settings.

Main Methods:

• Developed a novel federated tensor factorization approach.
• Implemented secure data harmonization and federated computation using the alternating direction method of multipliers (ADMM).
• Hospitals iteratively update local tensors and share summarized information for central aggregation.

Main Results:

• The federated approach successfully generated clinical phenotypes across multiple institutions.
• Achieved accuracy and phenotype discovery comparable to centralized training models that use combined datasets.
• Demonstrated the model's ability to respect patient privacy by not sharing raw data.

Conclusions:

• Federated tensor factorization is a viable and effective method for multi-institutional computational phenotyping.
• The developed ADMM-based procedures ensure data security and privacy.
• This approach facilitates large-scale, collaborative analysis of EHR data for robust phenotype discovery.