Improving Disease Progression Models With Clinical Notes

Abstract

This thesis improves upon RNN-based disease progression models in the vein of Choi et al.’s Dr. AI by adding clinical note data through two different fixed-vector representations. The first representation is a bag-of-words vector that indicates the presence of terms in the clinical notes examined. The second, more interesting representation is the dense representation, created using the final hidden state of an RNN that takes in clinical notes as an input. After evaluating on an independent test set, this thesis finds that there is a ~1% jump in efficacy after including clinical notes when measuring using the ‘recall @ n’ standard developed in Choi et al.’s paper. Additionally, this paper finds inconclusive evidence on which of the two representations is better, meaning that further study is likely needed to find the optimal fixed-vector representation of clinical notes.