Developing and validating EHR-integrated readmission risk prediction models for hospitalized patients with diabetes

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Hospital readmission is an undesirable, costly outcome that may be preventable. Hospitalized patients with diabetes are at higher risk of readmission within 30 days (30-d readmission) than patients without diabetes, and >1 million readmissions occur among diabetes patients in the US annually. Certain interventions can reduce readmission risk, but applying these interventions widely is cost prohibitive. One approach for improving the efficiency of interventions that reduce readmission risk is to target high-risk patients. We previously published a model, the Diabetes Early Readmission Risk Indicator (DERRITM), that predicts the risk of all-cause 30-d readmission of patients with diabetes. The DERRI, however, has modest predictive accuracy (C-statistic 0.63- 0.69), and requires manual data input. Recently, we demonstrated that adding variables to the DERRI substantially improves predictive accuracy (DERRIplus, C-statistic 0.82). However, using this larger model to predict readmission risk based on manual input of data would be too labor intensive for clinical settings. Indeed, most readmission risk prediction models are limited by the trade-off between accuracy and ease of use; lack of translation to a tool that integrates with clinical workflow; modest accuracy; lack of validation; and dependence on data only available after hospital discharge. The objectives of the current proposal are: 1) To develop more accurate all-cause unplanned 30-d readmission risk prediction models using electronic health record (EHR) data of patients with diabetes (eDERRI); 2) To translate the models to an automated, EHR-based tool that predicts % readmission risk of hospitalized patients; and 3) To prospectively validate the eDERRI models and tool. The new eDERRI models will expand upon the variables in the DERRIplus based on availability in EHR data (e.g., sociodemographics, encounter history, medication use, laboratory results, comorbidities, and length of stay). To develop the models, we will leverage data from the PaTH Clinical Data Research Network (CDRN), a multi-center, 40-plus hospital member of the National Patient-Centered Clinical Research Network (PCORnet). We will apply state-of-the-art deep-learning methods to develop optimal predictive models. This project will analyze a large, multi-center cohort of nearly 340,000 discharges with cutting-edge techniques to develop better models and translate them to an automated tool that predicts readmission risk for individual patients with diabetes. The proposed tool would identify higher risk patients more likely to benefit from intervention, thus improving care and reducing costs.