PREDICTING HOSPITAL READMISSIONS BEFORE THE INITIAL ADMISSION

Purpose:

   Prevention of hospital readmissions has been an ongoing focus for payers as part of quality improvement and cost containment efforts. Numerous approaches for identifying patients at risk for readmission have focused on using clinical diagnoses which can be difficult to assert from administrative data.  We approach readmission risk from a diagnosis perspective and developed a model that predicts readmission risk prior to the initial admission and does not require diagnostic information, either from comorbidities or the diagnosis from the initial admission. 

Method:    Using a set of 123,951 hospital discharges spanning nearly three years, we developed a model that predicts an individual’s 30-day readmission risk should they incur a hospital admission. The model is built using clustering and boosted decision tree machine learning techniques and uses administrative medical insurance claims data as input. Readmission rates were obtained from two separately generated models, depending upon whether or not the individual had at least 12 months of continuous insurance coverage or had less than 12 months but at least 6 months of continuous insurance coverage. The first stage of the 12 month model clusters the number of: ICD9 diagnosis, hospital admissions, medical consultations, emergency department visits, and prescriptions during the prior 6 months into 5 risk levels. The risk level plus all of the 12 month variables are fed into an ensemble of boosted decision trees, which yields the final prediction. A similar calculation is performed for those members with between 6 and 12 months of insurance coverage except that the clustering is based upon the prior 3 months and the boosted decision tree uses the 6 month aggregated data.     

Result:    The dominate predictors of readmission are the type of insurance plan (Medicare, Medicaid, Commercial), age, sex, and the number of prior: ICD9 diagnosis, hospital admissions, medical consultations, emergency department visits, and prescriptions. Using a 25% hold-out set containing 15% readmissions, the model captures 64% of all 30-day readmits with a true positive rate of over 27%. Moreover, many of the “false” positives are simply delayed true positives.  53% of the predicted 30-day readmissions are readmitted within 180 days.

Conclusion:    Because this model identifies an individual’s readmission risk prior to their initial admission, it provides a method for proactively managing readmissions during the initial hospitalization.