Distributed learning from multiple EHR databases: Contextual embedding models for medical events.

Area of Science:

• Health Informatics
• Machine Learning
• Data Privacy

Background:

• Electronic health records (EHR) offer potential for personalized treatments and clinical predictions.
• EHR data present challenges due to irregularity, complexity, and privacy concerns, hindering multi-institutional analysis.
• Existing predictive models for EHR data often require data sharing, which is frequently infeasible.

Purpose of the Study:

• To develop a novel method for learning predictive models from multiple EHR databases in a distributed manner.
• To address privacy issues associated with sharing sensitive EHR data among research institutions.
• To enable the generalization of predictive models across different institutions without compromising data privacy.

Main Methods:

• Proposed a novel method based on Distributed Noise Contrastive Estimation (Distributed NCE) for multi-database learning.
• Integrated differential privacy techniques to protect intermediary information during the distributed learning process.
• Implemented the proposed methods as a stand-alone Python library available on GitHub.

Main Results:

• Demonstrated the ability to build predictive models in a distributed fashion while ensuring privacy protection.
• Showcased that the proposed method effectively preserves model structure.
• Achieved comparable prediction accuracy to centralized models using a real-world EHR dataset.

Conclusions:

• The developed Distributed NCE method enables privacy-preserving, distributed learning from EHR data.
• The approach facilitates the creation of generalizable clinical prediction models across institutions.
• The open-source implementation provides a practical tool for researchers and clinicians.