An End-to-End Predictive and Intervention Framework for Reducing Hospital Readmissions

Abstract

Hospital readmissions within a given days of discharge are an enduring issue to the healthcare systems of the world. Early readmissions are largely viewed as indicators of poor quality of care, and they impose high economic costs and penalties to providers, valued at billions of dollars of avoidable expenditure annually. This paper presents an end-to-end solution that merges predictive analytics with accurate clinical interventions in a bid to reduce 30-day hospital readmission rates. The model consists of data importation from electronic health records, a predictive module based on machine learning to forecast individual readmission risk at discharge, and an intervention engine that accepts predicted risk and translates it into actions such as enhanced discharge planning and post-discharge follow-up. The system also consists of closed-loop monitoring to learn from patient outcomes and iteratively refine its predictions and interventions. This work validates the framework on a large clinical dataset of diabetic patient admissions, demonstrating how careful feature engineering and a variety of machine learning models (k-nearest neighbors, logistic regression, and random forests) can identify high-risk patients. The results suggest that the prediction model has good discriminative performance (with area under the ROC curve of around 0.70) and that the addition of an intervention strategy based on model predictions could decrease observed readmission rates considerably in our assessment. This end-to-end solution showcases an effective roadmap for hospitals to prospectively identify high-risk patients and intervene promptly, using evidence-based actions to improve patient outcomes and avoid unwanted readmissions.