TRA2-1 REAL-TIME PREDICTIVE MODELING TO STRATIFY RISK FOR ALL ADULT INPATIENTS TO REDUCE HOSPITAL READMISSIONS

Thursday, October 18, 2012: 10:30 AM
Regency Ballroom C (Hyatt Regency)
INFORMS (INF), Health Services, and Policy Research (HSP)

Eduard E. Vasilevskis, MD1, Henry J. Domenico2, Daniel W. Byrne2, Neal R. Patel2, Julianne M. Morath2 and Laura Beth Brown2, (1)Vanderbilt University and the VA - Tennessee Valley, Nashville, TN, (2)Vanderbilt University, Nashville, TN

Purpose:  Effectiveness of interventions to reduce hospital readmissions is limited by inadequate risk-stratification at hospital admission.  The aim of this research was to develop and validate a 30-day all-cause readmission model using electronic medical records (EMR) data available within 24 hours, followed by integration of readmission risk into the electronic medical record.

Methods:   We performed a retrospective cohort study among patients at Vanderbilt University Medical Center (VUMC) who were discharged alive. Patients were included if ≥ 18 years of age and admitted to a medical or surgical unit from 7/1/2009 to 6/30/2010. The outcome was readmission within 30-days from hospital discharge.  388 variables were assessed as independent predictors, obtained exclusively from electronic databases, including: demographics, admission source, number of hospital admissions in the 6 months prior, and routine laboratory tests (e.g., CBC, BMP) from the first 24 hours of admission. We developed a logistic regression model of the relationship between independent variables and all-cause 30-day readmission using modern data reduction methods.  Bootstrap validation was performed with 200 replicates. We assessed discrimination and calibration with the c-statistic, Brier's score, and Hosmer-Lemeshow statistic. Finally, we tested feasibility of real-time risk calculations in the EMR.

Results:   A total of 20,718 patients met the inclusion criteria, 3172 (15.3%) were readmitted to VUMC within 30 days.  Overall, patients were: 53.2% male, mean age 53.5, median LOS 3.6 days (IQR 2.0 to 6.3).  The final model variables included: age, emergency department admission, number of hospital admissions in the prior 6 months, hemoglobin, MCV, RDW, WBC, CO2, Cl, and BUN.  The model with 10 variables had a c-statistic of 0.646 and a Brier of 0.125.  The model Hosmer-Lemeshow statistic was significant (p < .0001), however this could be due to large sample size as visual calibration appeared excellent. The bootstrap validation with 200 replicates indicated minimal bias due to overfitting (slope optimism =.019). Finally, incorporation into the EMR was successfully demonstrated (See Figure).

Conclusions: Development and implementation of an all-cause real-time predictive model for 30-day hospital readmission based on data available within the first 24 hours is feasible for the entire adult hospital population.  Our future work will assess whether using this model to focus interventions leads to reduced hospital readmissions.