Personalized Federated Learning for Gastric Cancer: Adaptive Inference from Large Heterogeneous Piecewise Electronic Health Record Based Biomedical Data

In large cancer databases, it is common to collect electronic health records that have varying origins - for example, records often come from rural and urban settings or from academic and community based hospitals. It is increasingly becoming commonplace to use large databases such as these to guide evidence based decision making. Two problems in this regard are privacy concerns and heterogeneity due to varied data sources. Federated learning has come up as a possible tool of choice in addressing privacy concerns, especially when there are piecewise datasets. We propose a novel technique that addresses inferential aspects of federated learning when dealing with heterogeneous data. Our methods are motivated by data from gastric cancer, one of the most debilitating forms of cancer. We show, in particular, heterogeneity that exists when considering piecewise data in gastric cancer and superiority of our techniques when heterogeneity is adaptively addressed in such settings.