FLORENCE: Federated Learning on Cancer Data
Colorectal cancer is the second most common cancer in Norway when looking at both sexes. In 2021, 4550 men and women were diagnosed with colorectal cancer. 1 in 4 colorectal cancer patients experience complications after surgery, and the same proportion experience a recurrence of the cancer within 3 years.
Better utilisation of registry and hospital data can give clinicians and patients a better basis for treatment and prevention of complications after surgery. Mapping data to the OMOP data model makes it possible to analyze data across countries without moving the data itself across borders.
In this project, we want to lay the foundation for an IT tool (decision support tool) that can be used by specialists to help with treatment choices for colorectal cancer patients.
The project is a collaboration with the Center for Surgical Science (CSS) at Zealand University Hospital in Denmark, and by using hospital and registry data from both countries, we will form the basis for a new decision support tool. In addition, it collaborates with Lund University in Sweden, Computerome at DTU and the unit for research projects at Zealand University Hospital.
In order to create a decision support tool that uses basic data from several countries, it is important that the data on which the analyses are based are comparable. As a result, data must be standardised and harmonised across countries before it can be compared. To achieve this, we will use a generic data model called OMOP-CDM from Observational Health Data Sciences and Informatics (OHDSI). The project will also use synthetic data for testing and development of algorithms that can be regularly improved via federated learning.
We have received funding for the project together with the project partners from Interreg Øresund-Kattegat-Skagerak (ØKS). Interreg ØKS is a European regional development fund that supports projects that focus on innovation, a green transition, transport or a borderless labour market.
Privacy is safeguarded, among other things, by using a common data model so that algorithms can be shared instead of health information. The local data on a hospital/registry will be used to train each version of the model, or decision support tool in this case. Changes in the model between each site can be shared, and in this way all treatment centres will have an equally good decision-making tool.