Privacy-first health research with federated learning

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Epidemiological Research

Background:

• Traditional health research often centralizes sensitive data, posing privacy risks.
• Centralized data analysis requires full access to private patient information.
• Need for privacy-preserving methods in distributed health studies.

Purpose of the Study:

• To evaluate the efficacy of federated learning (FL) in health research.
• To compare FL models with traditional centralized models for accuracy and privacy.
• To demonstrate the application of FL with differential privacy in clinical and epidemiological studies.

Main Methods:

• Applied modern federated learning techniques incorporating differential privacy.
• Trained machine learning models in a distributed manner across multiple sites/devices.
• Utilized diverse health studies, varying units of federation, model architectures, and learning task complexity.

Main Results:

• Federated models achieved comparable accuracy, precision, and generalizability to centralized models.
• Privacy protection was considerably stronger with federated learning, as data remained on local devices.
• Computational costs were not significantly increased by employing federated learning methods.

Conclusions:

• Federated learning with differential privacy is a viable and effective approach for health research.
• This methodology allows private data to remain local, empowering participants' data control.
• It enables robust scientific discovery without compromising patient privacy, bridging a critical gap in health research.