Representation Learning of Patient States

In this deliverable, a report introducing the development of representation learning on electronic health records is presented.

Electronic health records (EHRs) contain many types of information such as vital signs, laboratory measurements, clinical notes, demographics as well as other entries related to the patients. The proper use of EHRs holds the promise to advance different clinical research topics and one of the most recent and promising approaches to extract useful information is to apply representation learning methods.  These methods map EHRs into suitable data representations, which leads to better utilization of advanced machine learning methods and improves the performance of downstream clinical applications. This report presents the development of EHRs representation learning method for medical complication prediction tasks, from target definition to feature extraction.  A new strategy using hidden Markov model to learn a meaningful representation of the patients is also introduced.