The digital health sector's enthusiasm for federated learning as a quick-fix solution for data governance issues often overlooks its inherent limitations. In federated learning, a model moves across multiple independent datasets, learning locally and combining local models into a global one without centralizing data. The assumption that federated learning solves the ethical challenges of data sharing is prevalent among healthcare institutions and governing bodies, but it's a narrow perspective.
While some criticisms of centralized systems are valid, insufficient attention has been given to the drawbacks of decentralization. In their article "Federated learning is not a cure-all for data ethics," Bak M et al. express concern about the eagerness to fulfill digital health promises with federated learning. They point out that this approach presents challenges and limitations that need careful consideration. While federated learning avoids centralizing sensitive data, it doesn't inherently protect privacy. Organizations must uphold data protection responsibilities, and methods like differential privacy are needed to balance privacy and utility. Differential privacy's noise addition can affect model accuracy. Moreover, federated learning can lead to biased models, particularly when certain populations are under-represented in the source data, exacerbating fairness concerns. Transparency and explainability are also challenging due to the 'black box' nature of artificial neural networks and scattered datasets, posing accountability, ethical, and legal concerns.
Decision-making should involve stakeholders' discussions to balance each approach's ethical benefits and downsides, maintaining public trust in data-driven medicine. It's crucial to understand that federated learning isn't a cure-all. Properly de-identifying data allows training centralized datasets from various locations, addressing data diversity without federated learning's limitations.
Reference:
Bak, M., Madai, V.I., Celi, L.A. et al. Federated learning is not a cure-all for data ethics. Nat Mach Intell (2024). https://doi.org/10.1038/s42256-024-00813-x
.svg)
