A mong patients discharged from the hospital, those discharged after undergoing percutaneous coronary intervention (PCI) have among the highest rates of early readmission, accounting for an estimated $359 million 1 cost each year to Medicare alone. Because of their importance, hospital 30-day readmission rates are now publicly reported on a volunteer basis on the Hospital Compare website. Identifying patients at high risk for readmission is critical to targeted interventions that improve value.
Editorial see p 463Nevertheless, despite a growing emphasis on reducing short-term hospital readmissions, predicting readmissions accurately has remained elusive.2 Risk prediction models derived from registry data have been developed, 3 but are limited by modest discrimination. The electronic health record (EHR) may be a repository of patient-level information, including both structured and unstructured data, that can improve the ability to risk-stratify patients based on their likelihood of readmission.Registry-derived risk models can benefit from large data sets and reflect broad ranges of patients. Nevertheless, registryderived risk models cannot account for unstructured data. In fact, Background-Early readmission after percutaneous coronary intervention is an important quality metric, but prediction models from registry data have only moderate discrimination. We aimed to improve ability to predict 30-day readmission after percutaneous coronary intervention from a previously validated registry-based model. Methods and Results-We matched readmitted to non-readmitted patients in a 1:2 ratio by risk of readmission, and extracted unstructured and unconventional structured data from the electronic medical record, including need for medical interpretation, albumin level, medical nonadherence, previous number of emergency department visits, atrial fibrillation/ flutter, syncope/presyncope, end-stage liver disease, malignancy, and anxiety. We assessed differences in rates of these conditions between cases/controls, and estimated their independent association with 30-day readmission using logistic regression conditional on matched groups. Among 9288 percutaneous coronary interventions, we matched 888 readmitted with 1776 non-readmitted patients. In univariate analysis, cases and controls were significantly different with respect to interpreter (7.9% for cases and 5.3% for controls; P=0.009), emergency department visits (