Federated Learning... and Blockchain Technology?

Recently, we attended a keynote presentation at the BIO-IT Conference wherein several thought-leaders shared their experiences with—and the promises of—federated learning for advancing big data models in healthcare. In contrast to centralized learning, federated learning allows siloed data centers to train a model locally, without the need for traditional data-sharing. If you’re at all familiar with the realities—and aspirations—of blockchain technology, this thesis may sound familiar. While we don’t pretend to be experts on the nitty gritty of algorithm development or learning structures, we’ll offer some thoughts on what federated learning is, how it may relate to blockchain technology, and what it could mean for our increasingly data-enabled society.

First, let’s start with centralized learning. According to this structure, data centers (think: large health systems and research institutions) share their valuable data with a central actor, who then aggregates this data and uses it to train the model in development. Ringing any alarm bells? This type of learning structure faces several well-documented challenges that come with sharing data as a research tool, despite data having become a sort of currency (for example, patients refusing to grant data-sharing access, institutions keeping their data private). The keynote presenters shared their own experience with the difficulty of gathering sufficient data to train an AI model for brain tumor recognition in radiation oncology; namely, the international Brain Tumor Segmentation (BraTS) challenge that, through centralized data-sharing, gathered data from 2,000 patients over the course of ten years.

Rather than sharing data, in a federated learning structure each data center trains the model locally, producing model updates that are then aggregated together into the “final” model. In other words, the model, not the data, is moving. Following the BraTS challenge, the aforementioned researchers developed a federated learning approach for their radiation oncology AI brain tumor model: The Federated Tumor Segmentation (FeTS) initiative. Incredibly, they found that FeTS, the largest real-world federation in medicine, actually outperformed BraTS in a feasibility study. Federated learning powerfully unlocked access to data, with the FeTS study reaching more than 6,300 patients in only ten months.

The promise of a federated learning structure—that individual actors retain ownership of their data while simultaneously using it to advance a shared program or model—feels inherently allied with the principles of blockchain, or distributed ledger technology (DLT). At a birds-eye view, blockchain networks rely on nodes that secure and upload immutable data according to consensus algorithms. Blockchain networks promise interoperability by being inherently distributed (ie. the ledger is continuously synchronized across nodes rather than requiring an input/export into a central database to share information across entities) and by having zero-knowledge proof, whereby a message can be approved and added to the chain without revealing the actual contents of a message. Although unmentioned in the keynote presentation, these characteristics among others make us wonder: how does, can, or will blockchain technology enable federated learning, particularly in the health system?

Sources

1. Bakas, S and Martin, J. "Federated Futures: How the Largest Federated Learning Effort in Medicine Will Inform Our Next Steps." BIO-IT Conference 2023.