UC Riverside

Electronic Health Records (EHRs) consist of sparse, noisy, incomplete, heterogeneous, and

unevenly sampled clinical data of patients. Such data can be treated as a temporal stream

of events of varied types occurring at irregularly spaced time points. Our first approach

to model such data is a Bayesian structure learning based marked temporal point process

model. We model the event streams, including vital signs, laboratory results contained

in two different datasets using a piecewise-constant conditional intensity model (PCIM),

a type of marked point process. Next, we model the stream of discrete clinical events

in continuous time using a neurally self-modulating marked temporal point process model

that uses continuous-time long short-term memory (LSTM) cells as its building blocks.

Our methods improve prediction performance in multiple tasks, including in-hospital mortality prediction, while providing suitable regularization and bypassing the need for data

imputation.