The effects of data sources, cohort selection, and outcome definition on a predictive model of risk of thirty-day hospital readmissions

摘要: Display Omitted Hospitals seek to predict readmissions, but there are many proposed models.The study aims show the impact of varying factors model design on performance.Targeting reason for readmission improves discrimination. ROCs range from 0.68 0.92.Patient visit and laboratory results contribute most prediction.Performance is highly cohort-dependent. Comparing models across studies may be hard. BackgroundHospital risk prediction remains a motivated area investigation operations in light hospital readmissions reduction program through CMS. Multiple have been reported with variable discriminatory performances, it unclear how affect performance. ObjectivesTo effects three development based health record data: (1) (primary diagnosis); (2) available data types (e.g. history, results, etc); (3) cohort selection. MethodsRegularized regression (LASSO) generate predictions using prevalence sampling. Support Vector Machine (SVM) used comparison selection testing. Calibration by refitting outcome prevalence. ResultsPredicting multiple reasons resulted ROC areas ranging 0.92 congestive heart failure 0.71 syncope all-cause readmission. Visit history tests contributed predictive value; contributions varied diagnosis. Cohort definition affected performance both parametric nonparametric algorithms. Compared all patients, limiting patients whose index admission diagnoses matched decrease average 0.78 0.55 (difference 0.23, p value 0.01). plots demonstrate good calibration low mean squared error. ConclusionTargeting impacted In general, prediction; source had large performance, these difficulty comparing different modeling.