POCs within the testbed
Federated Mortality Prediction
Region Halland, together with Halmstad University, has devised an algorithm predicting emergency care patients' survival rate up to 30 days after they have visited the emergency care unit. This allows them to better follow-up and prevent patients from unnecessary illness and cost savings for the hospital.
In the testbed project, Region Halland, Halmstad University, and HPe will replicate the algorithm on a simulated Federated learning setting, to learn how the performance of the algorithm is affected by having the data decentralized in a various number of nodes. The next step is to transfer these learnings to a different Swedish hospital, in another region of Sweden, e.g. the Karolinska Institute, where a common model will be built on each other's data.
Here, challenges include not only the technical, e.g. which framework is most suitable to what use-case but also how do you guarantee that the data is structured in the same way on the other side, and the quality of the data? Remember, you are not allowed to see the data, only monitoring the result of the trained algorithm.
To enable this, the privacy of the data has to be guaranteed and the model will be immune to, for instance, linkage attacks (a way to discover who the “user” is by comparing the model to another data source).
An aspect that should not be forgotten is the legal side. Novel legal documents have to be constructed to allow the various hospitals to share a model, devised on the premise of federated learning.
