Enhancing healthcare data privacy and interoperability with federated learning.

Area of Science:

• Digital Health
• Machine Learning
• Health Informatics

Background:

• Vast amounts of digital health data, especially from wearable sensors, are underutilized due to privacy and interoperability issues.
• Existing systems struggle to analyze and exchange data across heterogeneous platforms, limiting the exploitation of electronic medical records and mobile health applications.

Purpose of the Study:

• To develop a novel platform integrating federated learning (FL) and Fast Healthcare Interoperability Resources (FHIR) to address healthcare data underutilization.
• To enable collaborative, privacy-preserving model training on wearable sensor data while ensuring data standardization and interoperability.

Main Methods:

• A converged platform combining FL and FHIR was developed, utilizing local model learning to enhance data privacy.
• An AutoML-powered web application compatible with both FL and centralized learning (CL) was created for practical demonstration.
• Empirical evaluation compared FL and CL models using standard classification and regression metrics.

Main Results:

• Federated learning models achieved classification accuracy comparable to or better than centralized learning models.
• FL models demonstrated equal performance to CL models in regression tasks.
• The platform successfully illustrated the feasibility of predictive modeling for physical activity and energy expenditure, adhering to FHIR standards.

Conclusions:

• The FHIR-integrated federated learning platform offers a practical framework for future digital health ecosystems.
• This approach optimizes the use of connected health data by ensuring interoperability and preserving user privacy.
• The study highlights the potential of FL and FHIR to unlock the full value of digital health data.