Contribution of Data Categories to Readmission Prediction Accuracy

Identification of patients at high risk for readmission could help reduce morbidity and mortality as well as healthcare costs. Most of the existing studies on readmission prediction did not compare the contribution of data categories. In this study we analyzed relative contribution of 90,101 variables across 398,884 admission records corresponding to 163,468 patients, including patient demographics, historical hospitalization information, discharge disposition, diagnoses, procedures, medications and laboratory test results. We established an interpretable readmission prediction model based on Logistic Regression in scikit-learn, and added the available variables to the model one by one in order to analyze the influences of individual data categories on readmission prediction accuracy. Diagnosis related groups (c-statistic increment of 0.0933) and discharge disposition (cstatistic increment of 0.0269) were the strongest contributors to model accuracy. Additionally, we also identified the top ten contributing variables in every data category. Introduction Hospital readmission is an important metric of inpatient care quality and a major contributor to healthcare costs. Nearly 20% of hospitalized Medicare beneficiaries are readmitted to the hospital within 30 days of discharge. For example, in 2009, the Medicare Payment Advisory Commission (MedPAC) found that Medicare spends about $12 billion per year on preventable readmissions. In order to solve this problem, the Centers for Medicare and Medicaid Services (CMS) uses readmission rate as a criterion to penalize the hospitals whose readmission rate exceeds the expected threshold. In fiscal year 2017, more than half of the nation’s hospitals (2597 hospitals) will be penalized for excessive readmission rates. The average penalty is 0.71% of the hospital’s total Medicare reimbursement, and it can be as high as 3% depending on how far the rate of readmission exceeds the threshold. Additionally, readmission rate is also an important metric to evaluate the quality of healthcare. In the United States, readmission rate was selected as a significant indicator to measure the quality of healthcare in the National Quality Forum of 2008. In the United Kingdom, readmission rate within 28 days after discharge is used to measure the quality of healthcare. There are three categories of approaches to decreasing readmission rates: pre-discharge interventions, post-discharge interventions and bridging interventions. Pre-discharge interventions may include patient education, discharge planning, medication reconciliation and follow-up appointment scheduling before discharge. Post-discharge interventions may include close follow-up, timely PCP communication, follow-up telephone call, patient hotline and home visits. Bridging interventions may include transition coach, patient-centered discharge instructions and provider continuity. However, it is very expensive to apply these interventions to every patient. Thus, it is significant to identify the patients at high risk for readmission in order to make these interventions cost-effective. Electronic Medical Record (EMR) data provide abundant information to predict readmission risk. There are many published studies in this field. Hasan, Omar, et al. performed logistic regression analysis to identify significant predictors of unplanned readmission within 30 days of discharge and developed a scoring system for estimating readmission risk. Shadmi, Efrat, et al. developed a prediction score based on before admission electronic health record and administrative data using a preprocessing variable selection step with decision trees and neural network algorithms. Yu, Shipeng, et al. proposed a generic framework for institution-specific readmission risk prediction, which takes patient data from a single institution and produces a statistical risk prediction model optimized for that particular institution and, optionally, for a specific condition. Chen, Robert, et al. implemented a cloud-based predictive modeling system via a hybrid setup combining a secure private server with the Amazon Web Services Elastic Map-Reduce platform. Greenwald, Jeffrey L., et al. designed a 30-day readmission risk prediction model through identification of physical, cognitive, and psychosocial issues using natural language processing. However, these investigations provide little information on the comparative importance of different data categories on the accuracy of the readmission prediction models. We therefore conducted a study to analyze the contributions of different data categories to the discriminative ability of readmission prediction.