Using medical knowledge to improve learned representations of patient health trajectories.

Most state of the art physiological processing models are unstructured, neural models. These models achieve good performance, but cannot benefit from human curated medical knowledge. Prior work in this domain has found that simpler models can benefit from clinical domain knowledge, but in the context of neural models over physiological data, it is not clear how we can even encode this structure while retaining the flexibility of neural representations, and no works that we know of have shown a significant benefit from the inclusion of clinical knowledge.

We experiment with several mechanisms of encoding such structure into a recurrent neural framework for the analysis of physiological data from the MIMIC-III database. Preliminary results show that simple structural modifications to the standard LSTM framework do induce performance improvements, though as of yet these do not appear to be due to the addition of clinical knowledge. Additional experiments involving more complex network structures and encodings of clinical knowledge are underway.