Constrained elastic net based knowledge transfer for healthcare information exchange

Abstract

Transfer learning methods have been successfully applied in solving a wide range of real-world problems. However, there is almost no attempt of effectively using these methods in healthcare applications. In the healthcare domain, it becomes extremely critical to solve the “when to transfer” issue of transfer learning. In highly divergent source and target domains, transfer learning can lead to negative transfer. Most of the existing works in transfer learning are primarily focused on selecting useful information from the source to improve the performance of the target task, but whether the transfer learning can help and when the transfer learning should be applied in the target task are still some of the impending challenges. In this paper, we address this issue of “when to transfer” by proposing a sparse feature selection model based on the constrained elastic net penalty. As a case study of the proposed model, we demonstrate the performance using the diabetes electronic health records (EHRs) which contain patient records from all fifty states in the United States. Our approach can choose relevant features to transfer knowledge from the source to the target tasks. The proposed model can measure the differences between multivariate data distributions conditional on the predicted model, and based on this measurement we can avoid unsuccessful transfer. We successfully transfer the knowledge across different states to improve the diagnosis of diabetes in a certain state with insufficient records to build an individualized predictive model with the aid of information from other states.