The Effectiveness of Transfer Learning in Electronic Health Records Data

Sebastien Dubois, Nathanael Romano, Kenneth Jung, Nigam Shah, and David C. Kale

TL;DR: We apply transfer learning to the problem of training disease classifiers from EHR data where labels are scarce.

Abstract: The application of machine learning to clinical data from Electronic Health Records is limited by the scarcity of meaningful labels. Here we present initial results on the application of transfer learning to this problem. We explore the transfer of knowledge from source tasks in which training labels are plentiful but of limited clinical value to more meaningful target tasks that have few labels.

Conflicts: stanford.edu, usc.edu.isi.edu