Confederated learning in healthcare: Training machine learning models using disconnected data separated by...

Area of Science:

• Health Informatics
• Machine Learning
• Data Science

Background:

• Patient health data is fragmented across multiple providers and settings.
• Privacy concerns and regulatory barriers hinder data centralization for healthcare analysis.
• Existing federated learning methods struggle with vertically separated data and patient ID matching across institutions.

Purpose of the Study:

• To develop "confederated machine learning" methods for training models on data separated by multiple dimensions.
• To enable machine learning on horizontally and vertically separated health data without patient ID matching.
• To address limitations of traditional federated learning in complex healthcare data scenarios.

Main Methods:

• Proposed and evaluated a confederated learning approach for training machine learning models.
• Applied the method to stratify disease risk using data separated by individual, data type, and identity.
• Leveraged representation learning, generative models, imputation, and data augmentation within the confederated framework.

Main Results:

• Achieved an Area Under the Curve Receiver Operating Characteristic (AUCROC) of 0.787 for diabetes prediction.
• Obtained an AUCROC of 0.718 for psychological disorder prediction.
• Reached an AUCROC of 0.698 for Ischemic heart disease prediction using nationwide health insurance claims.

Conclusions:

• The proposed confederated learning method successfully trained machine learning models on multi-dimensionally separated health insurance data.
• This approach offers a viable solution for leveraging fragmented health data for predictive modeling.
• Demonstrated the effectiveness of confederated learning in overcoming data silos and privacy challenges in healthcare.