Learning deep patient representations for the teleICU

Abstract

This thesis presents a method of extracting deep robust representations of teleICU clinical data using Transformer networks, inspired by recent machine learning literature in language modeling. The utility of these representations is evaluated in various prediction outcome tasks, in which they were able to outperform linear and neural baselines. Also examined are the probability distributions of various patient characteristics across the learned patient representation space; where corresponding high-level spatial structure suggests potential for use as a similarity metric or in combination with other patient similarity metrics. Finally, the code for the models developed is publicly provided as a starting point for further research.