BackgroundSufficiently accurate predictions of hospital readmissions are necessary for the allocation of scare clinical resources to reduce preventable readmissions. We describe the use of a data-driven approach that relies on machine learning algorithms to predict readmission at the time of discharge.MethodsWe employ random forests to clinical and administrative electronic health record data available from a cohort of 103,688 patients discharged from the acute inpatient settings of the University of Pennsylvania Health System between June 25th, 2011 and June 30th, 2013. We predict both 30-day all-cause readmissions and 7-day unplanned readmissions using only predictors available by the time of discharge. Using oversampling and undersampling of the different outcome classes of readmission and no readmission, we incorporate into our models the asymmetric costs of a false negative relative to a false positive from the perspective of a hospital. We calculate variable importance scores for included predictors. Our approach was derived and validated using split-sample internal validation.ResultsWe developed a machine learning-based model using random forests with a 5:1 relative cost ratio for 30-day all-cause readmissions that achieves a sensitivity of 65% and specificity of 71% on validation data, as well as a random forests model with a 20:1 cost ratio for 7-day unplanned readmissions that achieves a sensitivity of 62% and specificity of 66% on validation data. Prior health system utilization, clinical discharging service, and vital sign information were most predictive of readmissions.ConclusionBy modeling the complex relationships between many predictor variables and readmission data for a large health system, we demonstrate successful predictive models that can be used upon discharge to flag patients at high risk of readmission.
