Machine-Learning Modeling to Predict Hospital Readmission Following Discharge to Post-Acute Care

ObjectivesPrimary purpose was to generate a model to identify key factors relevant to acute care hospital readmission within 90 days from 3 types of post-acute care (PAC) sites: home with home care services (HC), skilled nursing facility (SNF), and inpatient rehabilitation facility (IRF). Specific aims were to (1) examine demographic characteristics of adults discharged to 3 types of PAC sites and (2) compare 90-day acute hospital readmission rate across PAC sites and risk levels.DesignRetrospective, secondary analysis design was used to examine hospital readmissions within 90 days for persons discharged from hospital to SNF, IRF, or HC.Settings and ParticipantsCohort sample was composed of 2015 assessment data from 3,592,995 Medicare beneficiaries, including 1,536,908 from SNFs, 306,878 from IRFs, and 1,749,209 patients receiving HC services.MeasuresInitial level of analysis created multiple patient profiles based on predictive patient characteristics. Second level of analysis consisted of multiple logistic regressions within each profile to create predictive algorithms for likelihood of readmission within 90 days, based on risk profile and PAC site.ResultsTotal sample 90-day hospital readmission rate was 27.48%. Patients discharged to IRF had the lowest readmission rate (23.34%); those receiving HC services had the highest rate (31.33%). Creation of model risk subgroups, however, revealed alternative outcomes. Patients seem to do best (i.e., lowest readmission rates) when discharged to SNF with one exception, those in the very high risk group. Among all patients in the low-, intermediate-, and high-risk groups, the lowest readmission rates occurred among SNF patients.Conclusions and ImplicationsThe proposed model has potential use to stratify patients’ potential risk for readmission as well as optimal PAC destination. Machine-learning modeling with large data sets is a useful strategy to increase the precision accuracy in predicting outcomes among patients who have nonhome discharges from the hospital.