Early Detection of Readmission Risk for Decision Support Based on Clinical Notes

Abstract

Hospital readmission shortly after discharge is contributing to rising medical care costs. Attempts have been exerted to reduce readmission rates by predicting patients at high risk of this episode on the basis of unstructured clinical notes. Discharge summary as part of the clinical\n prose is effective at modeling readmission risk. However, the predictive value of notes written upon discharge offers few opportunities to reduce the chance of readmission because the target patient might have already been discharged. This paper presents the use of early clinical notes in\n building a machine learning model to predict readmission at 48 h immediately after a patient s admission. Extensive feature engineering, testing multiple algorithms, and algorithm tuning were performed to enhance model performance. A risk scoring framework that combines the data- and knowledge-driven\n feature scores in risk computation was developed. The proposed predictive model showed better prognostic capability than the machine learning model alone in terms of the ability to detect readmission. In specific, the proposed algorithm showed improvements of 11%–28% in sensitivity and\n 1%–3% in the area-under-the-receiver operating characteristic curve.