A Sequence Based Approach for Predicting Clinical Events

Abstract

The data associated to each patient increases almost linearly as the patient flows through the continuum of care. Analysis of the data collected during a patient’s admission to the hospital reveals that it grows vertically as well as horizontally as a variety of readings are taken for the patient. In general, ma- chine learning techniques are designed and evaluated to predict clinical events at one particular time point during this process (on admission to the hospital, or on discharge). This highlights one of the key challenges of making predictive solutions applicable to the real world setting, as it limits the interventions that can be taken while the patient is at the hospital, to avoid undesirable clini- cal outcomes down the road. To address this challenge, we have proposed a novel framework of at-admit and sequence based models that predict clinical outcomes accurately at different time points of a patient’s hospital stay and perform consistently better than a retrospectively designed solution. Hospitalizations account for about half of all healthcare expenses, and it has been estimated that 13% of the inpatients in the United States use more than half of all hospital resources through repeated admissions. Therefore, the clinical outcome chosen for this work is predicting thirty day readmissions for the “all cause” population. We compare our proposed approach to the state of the art readmission modeling approach of retrospective feature creation, and see an average improvement of 7% in the area under the curve as well as significant improvements in precision, accuracy and recall.