Decomposing the hazard function into interpretable readmission risk components

Abstract

Hospital decision-makers use predictive models to proactively manage risk of readmission for discharged patients. While predictions from classification models are easily integrated into decision-making processes, it is unclear how to best integrate predictions of the evolution of risk from time-to-event models. We propose a method for summarising predictions of risk over time that produces interpretable components for use in a variety of decision-making processes. The proposed method summarises predictions of risk over time (hazard functions) by approximating them with a parametric smoother. The components of the smoothed approximation can then serve as the basis for decision-making. To demonstrate the proposed summarisation method, we apply it in the specific case of a previously published model for patients discharged from a large teaching hospital on the Gold Coast, Australia. In this context, we describe how the summaries produced by the method could be used to estimate time until a patient reaches a stable, persistent risk level or to stratify patients according to risks of readmission in excess of patient-specific baselines. Our method is anticipated to be valuable in and outside of healthcare for settings where the evolution of risk is important, with specific examples including post-transplantation risk and reinjury risks.