An Analytics Approach To Reducing Hospital Readmission

One of the significant sources of waste in the Unites States health care systems is preventable hospital readmission. About 2.3 million Medicare fee-for-service beneficiaries are re-hospitalized within 30 days after discharge which incurs an annual cost of $17 billion. However, it is reported by the Medicare Payment Advisory Commission that about 75% of such readmissions can and should be avoided because they are the results of factors such as poor planning for follow up care transitions, inadequate communication of discharge instructions, and failure to reconcile and coordinate medications. Hence, reducing unnecessary rehospitalization through care transition and systems engineering principles has attracted policymakers and health organizations as a way to simultaneously improve quality of care and reduce costs. In this dissertation we investigated predictive and prescriptive analytics approaches for discharge planning and hospital readmission problem. Motivated by the gaps in research, we first develop a new readmission metric based on administrative data that can identify potentially avoidable readmissions from all other types of readmission. The approach is promising and uses a comprehensive risk adjustment, Diagnostic Cost Group Hierarchical Condition Category, to assess the clinical relevance between a readmission and its initial hospitalizations. Next, we tackle the difficulties around selecting an appropriate readmission time interval by proposing a generic Continuous Time Markov Chain (CTMC) approach conceptualizing the movements of patients after discharge. We found that cutoff point defining readmission time interval must not depend on the instantaneous risk of readmission but rather it has to be based on quality of inpatient or outpatient care received. We further assert that the government endorsed 30 day time window which has been used for profiling hospitals and public reporting is not appropriate for chronic conditions such as chronic obstructive pulmonary disease. Thus, we propose a special case of the CTMC method and obtain the optimal cut point that best stratifies among inpatient and outpatient care episodes. Third, we proposed a novel tree based prediction method, phase time survival forest (PTSF), for patient risk of readmission that combines good aspects of traditional classification methods and timing based models. The method is simple to implement and can be able to (1) model the effect of partially known information (censored observations) into the risk of readmission, and (2) directly incorporate patient\u27s history of readmission and risk factors changes over time. The latter property is highly favorable especially when repeated measurements of patient factors or recurrent readmissions are likely. The basic idea is quite generic and it works by modifying the traditional replicate based bootstrap samples to account for correlations among repeated records of a subject. We demonstrated the superiority of our model over current solutions with respect to various accuracy and misclassification criteria. Further, to confirm that the high discrimination ability of our proposal is irrespective to overfitting, we performed internal and external validation with 2011-12 Veterans Health Administration data from inpatients hospitalized for heart failure, acute myocardial infarction, pneumonia, or chronic obstructive pulmonary disease in the Mid West facilities. Results indicated improved discrimination power compared to the literature (c statistics greater than 80%) and good calibration. Overall, the current research outlined a successful multifaceted analytics framework that enables medical decision makers to systematically characterize, predict, and reduce avoidable readmissions and contribute to patient care quality improvements