Architectural Design of a Blockchain-Enabled, Federated Learning Platform for Algorithmic Fairness in Predictive...

Area of Science:

• Artificial Intelligence
• Health Informatics
• Blockchain Technology

Background:

• Developing generalizable predictive models for healthcare requires diverse data to mitigate bias and ensure fairness.
• A key challenge is cross-institutional model training without compromising patient privacy or introducing local biases.

Purpose of the Study:

• To investigate bias and fairness issues in machine learning for predictive healthcare.
• To propose a software architecture integrating federated learning and blockchain to enhance fairness, maintain accuracy, and minimize costs.

Main Methods:

• An iterative design approach using design science research, involving two cycles: federated learning for bias mitigation and decentralized architecture.
• Development of a bias-mitigation process within a blockchain-empowered federated learning framework.
• Implementation using Aplos smart contract, microservices, Rahasak blockchain, and Apache Cassandra.

Main Results:

• Demonstrated enhanced accuracy of predictive diagnosis through an improved fairness mechanism.
• Successfully simulated joint training of privacy-protected models across 5 medical centers using 20,000 local and 1000 federated training iterations.
• Validated the effectiveness of the blockchain-integrated federated learning approach.

Conclusions:

• Addressed technical challenges of prediction biases in healthcare AI.
• Presented an innovative design for a fairness-aware system using federated learning, blockchain, and a distributed architecture.
• Showcased the system's ability to tackle privacy, security, accuracy, and scalability issues, promoting fairness and equity in predictive healthcare.