Related health data is sometimes dispersed over multiple data silos, each controlled by a different entity (GP, hospital, lab, health insurer, pharma company, …). While each of these entities can apply their own machine learning on their data, their models could potentially benefit from the data held in other silos. However, for practical, business, IP or legal reasons, directly sharing the data (such as with federated data approaches) or the models (such as with federated learning) is often difficult. This talk explains the concept of federated analytics, that lets each participating entity build their own model using only locally available data whilst indirectly incorporating information from other data silos in a way that doesn't compromise privacy. We illustrate this with two cases in a clinical setting. In the first we introduce amalgamated machine learning (PAML), a federated analytics approach that only calculates and shares PAML features, that are meant to preserve privacy and IP of the underlying data as well as models. We show how we can apply PAML in application on predicting acute kidney injury (AKI) prediction models where early results using the MIMIC-III data set show that the performance of federated analytics is significantly better than purely local models, and close that of models built by ignoring privacy and pooling all data. Next we also introduce the Athena research project where multiple leading hospitals in Flanders, pharma industry and research partners are developing a federated analytics platform for clinical oncology research.
Share this page on your timeline