Predicting intensive care unit patient outcomes through patient similarity

Abstract

An ICU stay involves invasive treatments, and frequently, the decision to continue therapy is made with limited information based on the physician's personal experience. This thesis proposal describes a tool to assist this decision by identifying similar patients and using their outcomes for prediction. We used the eICU Collaborative Research Database (eICU-CRD) v2.0 for the project. Different time varying and time constant features about the patient's demographics and clinical trajectory was used as input data, such as patient age and longitudinal blood pressure measurement. Using this information, a Cox Proportional Hazards model was built to map the multivariate time series of input data to a univariate time series, which was used to match the patient to a cohort of similar patients. Based on the cohort, this model predicted the probability of a healthy discharge by using the aggregate outcome of the cohort for prediction.