Federated learning in healthcare

Challenge

Some research questions require data from various sources. For example, when studying rare cancers, the volume of patients in single country may not be enough to determine which treatments result in the best outcome. Combining datasets from various organisations is often infeasible due to legal and privacy concerns. This inaccessibility of sensitive data sources is quietly holding back research on the problems that matter.

Solution

By using the latest advancements in federated learning and secure computation, we can combine insights from data sources that are maintained by multiple organisations without the need to share record-level data. Instead of combining data in a single place, we send algorithms to wherever the data is located. This creates a new opportunity for a collaborative research, while significantly reducing privacy risks.

Results

As these technologies are still quite novel, they require further development to be able to use them in practice. I have been involved in the following:

• Development of a Secure Cox Proportional Hazards model for vertically-partitioned data. Paper published in BMC Medical.

• Development of protocols for vertically-partitioned data, including secure approximate matching, secure exploratory data analysis, and a secure logistic regression - more info.

• Set-up a federated learning infrastructure using vantage6 at 4 cancer registries in South-East Asia.