Validation of an Electronic Health Record Model for Unplanned Readmission in Patients With Heart Failure

Risk scores aim to predict 30-day heart failure (HF) readmissions.

Current models, including Epic's Risk of Unplanned Readmission for a 30-Day Event, demonstrate modest to moderate discriminatory ability (area under the curve: 0.59-0.72).

The goals of this study were to create an electronic health record (EHR)-automatable 30-day HF readmission risk calculator; to test the accuracy of Epic's model to predict 30-day HF readmission risk; and to compare the 2 models. To derive our EHR-automatable model using machine learning, we first performed a systematic review of 30-day HF risk prediction models. We then compared the predictive capabilities of the Epic model with our model. The Epic model had a C-statistic of 0.74, and the derived EHR risk score had a C-statistic of 0.76 (P = 0.41), suggesting comparable performance.

Risk scores are fundamental for personalization of care to accelerate resources for those who need them most. Although both models had moderate discrimination in predicting 30-day HF readmission, the independently derived model provides a platform for future enhancement, such as the inclusion of electrocardiogram-based artificial intelligence diagnostics.