Optimal policy responses and targeting of interventions to reduce 30-day hospital readmissions
Griffith, Kevin N.
MetadataShow full item record
Nearly 1 in 5 Medicare inpatients are readmitted within 30 days of discharge, costing the Medicare program approximately $15 billion per year. The Centers for Medicare & Medicaid Services (CMS) implemented the Hospital Readmission Reduction Program (HRRP) in 2012, which penalizes hospitals with higher than expected 30-day readmissions for patients with certain conditions. In the first study, we evaluated whether the HRRP was associated with lower readmission rates for targeted conditions. Overall, we find that HRRP implementation led to a 1.4 percentage-point reduction in readmission rates at penalized hospitals. Hospitals were responsive both to a “labeling effect” of receiving any penalty, as well as to an “incentive effect” associated with the size of the penalty. The HRRP is intended to penalize hospitals based on the quality of care they provide to patients, but not characteristics of the communities they serve. However, the program does not account for the availability of post-discharge care within hospitals' service areas. In study 2, we examined the association between post-discharge care supply (e.g., PCPs, nursing homes, skilled nursing facilities, hospices) and hospitals' readmission rates. We find that readmissions were positively associated with the per capita supply of home health agencies and nurse practitioners, and negatively associated with hospices, PCPs, and palliative care. Our results suggest potential modifications to the HRRP's risk adjustment, in order to avoid punishing hospitals that lack access to certain community resources. Hospitals have engaged in a variety of activities to reduce readmissions such as redesigned discharge processes, improved coordination with post-discharge sites of care, or through specific quality-of-care interventions. In the final study, we sought to enhance our ability to predict these patient readmissions, using cutting-edge techniques developed in the field of machine learning. We used the Nationwide Readmissions Database to estimate twelve individual machine learning algorithms and then combine them using mathematical optimization. The resulting 'super learner' predicts readmissions better than what's possible with the individual algorithms, or traditional regression methods. To the extent that patients at high risk of readmission may be identified, interventions and healthcare resources may be targeted towards them in a cost-effective manner.