Distributed cross-learning for equitable federated models - privacy-preserving prediction on data from five...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Data Privacy

Background:

• Medical predictive analytics enhance quality improvement, clinical research, and patient care.
• Integrating diverse patient records (horizontally or vertically) can improve predictive model performance.
• Existing methods for data integration often risk patient privacy.

Purpose of the Study:

• Introduce Distributed Cross-Learning for Equitable Federated models (D-CLEF) for privacy-preserving medical data integration.
• Evaluate D-CLEF's performance against centralized, siloed, and federated learning approaches.
• Demonstrate D-CLEF's utility in real-world healthcare scenarios.

Main Methods:

• Developed D-CLEF to incorporate horizontally or vertically partitioned data without disseminating patient-level records.
• Compared D-CLEF with centralized, siloed, and federated learning in horizontal and vertical partitioning scenarios.
• Utilized multi-center datasets including COVID-19, surgical, and heart disease data.

Main Results:

• D-CLEF performance closely approximated centralized solutions across different data partitioning strategies.
• D-CLEF significantly outperformed siloed learning approaches.
• D-CLEF achieved performance equivalent to federated learning, with a trade-off in synchronization time.

Conclusions:

• D-CLEF offers a viable solution for collaborative healthcare analytics without compromising patient data privacy.
• The D-CLEF framework facilitates secure data sharing and model building across institutions.
• D-CLEF acts as a promising accelerator for healthcare systems seeking to leverage collective data.