A Decentralized Framework for Biostatistics and Privacy Concerns.

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Multi-centric studies are crucial for large medical datasets but face challenges in data centralization due to practical, legal, and ethical concerns.
• Federated learning offers a solution by enabling data analysis across multiple institutions without data aggregation.
• Existing federated learning methods often rely on a central coordinating entity, which can be a bottleneck or a privacy risk.

Purpose of the Study:

• To propose and evaluate a fully decentralized federated learning framework for medical data analysis.
• To apply this framework to logistic regression, including the computation of confidence intervals.
• To demonstrate the feasibility and accuracy of the decentralized approach compared to centralized methods.

Main Methods:

• Developed a fully decentralized federated learning algorithm where all participating centers have equivalent roles.
• Applied the algorithm to logistic regression models for clinical data analysis.
• Incorporated confidence interval computation within the decentralized framework.

Main Results:

• The proposed decentralized federated learning framework achieved results comparable to a traditional centralized approach.
• The algorithm was successfully tested on two distinct clinical datasets distributed across multiple centers.
• The framework demonstrated the potential for accurate statistical modeling without data pooling.

Conclusions:

• A fully decentralized federated learning framework is a viable alternative for multi-centric medical studies.
• This approach maintains data utility and analytical accuracy while addressing privacy concerns.
• Further research into privacy-preserving mechanisms within decentralized federated learning is warranted.