JMIR Publications

ABSTRACT

Background:

Machine learning is increasingly used in building models for disease prediction and capturing complex relationships for clinical decision making. Effective and generalizable predictive models need diverse samples to avoid potential population bias and address algorithmic fairness. A major challenge is retrieving learning models across multiple institutions without bringing in local biases and inequity while preserving individual patients' privacy at each site. We designed and implemented a bias mitigation process within the blockchain-empowered federated learning framework based on a novel architecture design that enables multiple medical institutions to jointly train predictive models using their privacy-protected data effectively and efficiently, and ultimately achieving fairness of decision making in the healthcare domain. System evaluation indicated that the proposed architecture is capable of providing accurate prediction while preserving fairness with acceptable overhead. We further presented research implications and strategic future directions for stakeholders and algorithm development in the social-technical system context.

Objective:

The main objectives of this study are to understand the issues of bias and fairness in the machine learning process used in predictive healthcare domain and suggest that a software architecture integrated with federated learning and blockchain improves fairness with acceptable prediction accuracy and overhead.

Methods:

We used the design science research method consisting of 2 design cycles to implement our overall design. We innovated over existing federated learning platforms with blockchain integration using iterative design approach that improves predictive healthcare in terms of fairness and accuracy.

Results:

Here, we identified key challenges to bias mitigation and improving fairness in disease prediction. Next, using federated learning and recent innovations in distributed ledge technology, we proposed a decentralized and fairness aware federated learning system. Our design and prototype implementation using Aplos smart contract, microservices, Rahasak blockchain, and Apache Cassandra-based distributed storage helped us address several challenges facing bias mitigation process in the federated learning platforms. We demonstrated how our solution, which combines federated learning and blockchain smart contract with improved fairness mechanism, solves the challenges of bias mitigation and sets the stage for improving patient care and accuracy of predictive diagnosis, using 20,000 local model training iterations and 1,000 federated model training iterations in five simulated medical centers as peers in the Rahasak blockchain network.

Conclusions:

We identified technical limitations that bring in bias and reduce the fairness of predictive models. We presented several design innovations using federated learning and blockchain to prototype a system. We also presented the implementation details of a unique distributed architecture for a fairness aware system. We illustrated how this design can overcome privacy, security, prediction accuracy, and scalability limitations. We illustrated how improving these factors sets the stage for improving fairness and standardized application of federated learning and blockchain to predictive healthcare data.