Advancing Healthcare Solutions with Federated Learning

Abstract

As the COVID19 pandemic began spreading, there were only pockets of information available with hospitals across geographies. Researchers attempting to analyze information scrambled to collaborate. These efforts were hampered due to regulations and privacy protection laws in various nations, which frequently confine access to clinical information. On the one hand machine learning and AI were helping doctors to make quicker diagnostic decisions using predictive models and on the other pharmaceutical companies could leverage AI for advancing drug discovery and vaccine research. COVID19 is one example among many research efforts for advancing treatment for ailments concerning prominent health issues such as cancer treatment and rare diseases. Yet, these AI systems were being developed in silos and their capabilities were hampered by the lack of a collaborative learning mechanism, thus limiting their potential. In this chapter we describe how multiple healthcare services can collaboratively build common global machine learning models using federated learning, without directly sharing data and not running afoul of regulatory constraints. This technique empowers healthcare organizations harness data from multiple diverse sources, much beyond the reach of a single organization. Furthermore, we will discuss some engineering aspects of FL project implementation such as data preparation, data quality management, challenges over governance of models developed with FL, and incentivizing the process. We also cover some challenges arising from data as well as model privacy concerns, which could be addressed with solutions such as differential privacy, Trusted Execution Environments, and homomorphic encryption.