Identifying Sub-Phenotypes of Acute Kidney Injury using Structured and Unstructured Electronic Health Record Data with Memory Networks.

We propose a deep learning model architecture to identify sub-phenotypes of AKI using longitudinal structured and unstructured EHR data. An ICU stay representation v is derived by integrating structured and unstructured EHR data. Time-dependent structured EHR data (clinical sequences) is push into the memory by using embedding matrices A and B. Unstructured EHR data (clinical notes) is represented as a vector u by using hierarchical LSTM, which is combined with input memory vectors to retrieve important information from output memory vectors to form a vector o. Static information is integrated with o and u to form an ICU stay representative vector v.