Solving the Revolving Door Problem: Machine Learning for Readmission Risk Assessment

Abstract

In 2012 the United States passed legislation, penalizing hospitals for readmission of patients discharged within 30 days. However, many unknowns mean that hospitals cannot predict when each patient is appropriate to discharge. Through researching readmissions across the Thomas Jefferson University Hospital enterprise, we found that staff must make judgement calls based on their own clinical perspectives. Rather than expecting doctors to somehow intuit the interaction effects from thousands of variables, we surface trends and present strategies for mitigating readmission risks through machine learning (ML). Commonly, ML models are trained against data aggregated from various sources. This method of sourcing interferes with responding to population-based risk factors and variables that are specific to the hospital of interest. However, creating a custom model presents its own set of hurdles. The work of our team provides hospitals everywhere with an end-to-end pipeline to create a readmissions assessment tool, using their own data.