Abstract WP108: Predictive Accuracy of Multivariable Models for 30-day and 1-year All Cause Readmission After Stroke Using Different Combinations of Registry, Hospital and Claims Based Data Sources

Abstract

Introduction: Hospital readmissions are often used as an indicator of quality of care. However, identifying patients at risk of readmission after stroke is challenging and predictive models have historically not performed well, in part because they often rely on single data sources. Data linkage might offer a solution. Methods: We probabilistically linked data from the Michigan’s Get With The Guidelines Stroke registry and Michigan Value Collaborative multipayer claims database from Medicare and Blue Cross Blue Shield beneficiaries discharged alive following acute stroke (ICD-10 I61-I63) between 2016-2020. The registry dataset included 64 variables covering demographics, stroke presentation, medical history, procedures, and complications. The hospital dataset included 20 variables from the American Hospital Association’s database. The claims dataset included payer and 79 HCC comorbidity codes. Using combinations of the 3 data sources, we examined the performance of multivariable LASSO logistic regression models to predict all cause readmission at 30-days and 1-year post discharge. We generated hospital-specific testing and training models and reported the mean model discrimination (AUC) of all combinations of the 3 datasets. Results: Of 19,382 linked stroke discharges, 2,724 (14.1%) and 8,169 (42.2%) were readmitted within 30-days and 1-year, respectively. For 30-day readmission, the model based on only registry data produced the best performance (M1, Table). However, for prediction of 1-year readmission, the combination of registry and claims data produced the best performing model (M13, Table). Hospital level characteristics did not have any significant impact on prediction accuracy. Conclusions: Clinical registry data was the best data source for predicting 30-day readmission. However, claims based data were additive when predicting readmission within 1-year, probably because HCC codes add information about total comorbidity burden.