@unpublished{pittir35782,
           month = {January},
           title = {Conducting a retrospective cohort to evaluate the accuracy of a real-time multiple logistic regression tool for predicting the likelihood of patient readmission},
          author = {Cory Hayes},
            year = {2019},
        keywords = {readmission, HRRP, ACA, readmissions reductions, CMS, pay-for-performance, P4P},
             url = {http://d-scholarship-dev.library.pitt.edu/35782/},
        abstract = {Introduction:
Hospital readmissions have a great deal of public health significance, as they are burdensome and costly to providers, hospitals, and patients. The quality-of-care provided by hospitals is evaluated by comparing hospital readmission rates to national averages. Underperforming hospitals are penalized by the Centers for Medicare \& Medicaid Services (CMS). In 2018, 2,597 hospitals are being penalized a total of \$564 million. Published studies have demonstrated a multitude of approaches that were successful in reducing readmission rates, but they are too expensive for systemic implementation within a hospital. The University of Pittsburgh Medical Center (UPMC) Mercy Hospital Clinical Analytics team has constructed a multiple logistic regression prediction model that scores patient risk factors in order to flag high risk patients who are most likely to experience readmission.
Objectives/Aims:
Primarily, this study aims to evaluate the accuracy with which the UPMC Mercy multiple logistic regression model correctly predicts readmission risk in a clinical application. Once validated, it is our secondary aim to initiate a discharge intervention specifically for patients who are flagged by the model.
Methods:
The predictive logistic regression tool has been in use for slightly over one year at UPMC Mercy; daily reports score patients as ?lowest?, ?lower?, ?medium?, ?higher?, or ?highest? risk for readmission. Based on sample size calculations, about 200 individuals per predicted risk group were retrospectively recruited and followed for the next month in order measure readmission status over 7 days and 30 days. Chi-Square testing for independence and stratified one-sample proportion testing allowed for validation of the model?s accuracy.
Results:
Chi-square testing of independence demonstrated that not all risk quintiles had distinct mean readmission rates, contrary to our hypothesis. One-sample proportion testing further illustrated the poor fit for predicting 7-day readmission, with only 1/5 risk groups following expected mean rates of readmission. However, one-sample proportion analysis for 30-day readmission prediction resulted in 3/5 of the risk strata exhibiting similar mean rates of readmission as was expected, as well as a clinically relevant increasing trend across risk strata.
Discussion:
This multiple logistic regression model is not an accurate predictor of 7-day readmission. However, it appears that the model could be clinically relevant for predicting 30-day readmissions. The ?highest? risk strata displayed 36\% 30-day readmission, which is 16-18\% higher than the national average 30-day readmission rate. Though the model could benefit from optimization, it likely could be utilized in its current state to target high risk groups for 30-day readmission rather reliably.}
}