Assessing Clinical Usefulness of Readmission Risk Prediction Model

Abstract

Readmission manifests signs of degraded quality of care and increased healthcare cost. Such adverse event may be attributed to premature discharge, unsuccessful treatments, or worsening comorbidities. Predictive modeling provides useful information to identify patients at a higher risk for readmission for targeted interventions. Though many studies have proposed readmission risk predictive models and validated their discriminative ability with performance metrics, few examined the net benefit realized by a predictive model. We compared traditional logistic regression against modern neural network to predict unplanned readmission. An added value of 7% on discriminative ability is observed for modern machine learning model compared to regression. A cost analysis is provided to assist physicians and hospital management for translating the theoretical value into real cost and resource allocation after model implementation. The neural network model is projected to contribute 15× more savings by reducing readmissions. Aside from constructing better performing models, the results of our study demonstrate the potential of a clinically helpful prediction tool in terms of strategies to reduce cost associated with readmission.