Federated Learning and Explainable AI in Healthcare

Abstract

Data-driven innovation is opening new opportunities for AI in healthcare, leading to impactful solutions. However, increasing amounts of data are stored across silos, facing issues related to security, privacy, custodianship, and usage for research. Data sharing, which is crucial to building robust AI models, is still challenging. Federated learning (FL), an emerging machine learning technique that trains AI models across decentralized devices or sites holding local datasets without moving them, could help address some of the challenges. With FL, the limitations and constraints concerning data availability for research could be reduced, as it enables sites to jointly work towards a common research goal while keeping data local. Moreover, FL can support model surveillance and in-product learning to improve an already deployed model by relying on large and distributed real-life user datasets. Here the power of the federation is harnessed to prevent unwanted degradation and address edge cases (i.e., outliers at local or global level). With the adoption of FL, there is a need to apply model explainability in this new context so that sites can get a deeper understanding of their model, enriched with information relevant for the FL way of training. In FL, features or data samples can contribute at one site and be absent at other sites. Therefore, explainability should be integrated in FL to share insights about the federated and local models so that each participant can decide how to leverage the training and inference processes.

Ähnliche Arbeiten