Using electronic medical records to develop a predictive model of 30‐day hospital readmission for people with ADRD

Abstract

AbstractBackgroundHospitals are insufficiently equipped for Alzheimer’s disease and related dementia (ADRD) patients. Thus, readmission incidence is much higher and costlier among ADRD patients than the general population. Hospital discharges often occur without adequate preparation for the complex care management needs of ADRD patients and their caregivers. This study’s objective was to develop a risk‐assessment tool for hospitalized patients with ADRD. By supporting timely and better identification of who is at risk and why, already‐scarce resources can be allocated more efficiently to reduce readmissions.MethodsWe used 2016‐2019 EMR data from the University of Michigan health system (Michigan Medicine) and applied machine learning techniques (Random Forest, XGBoost, and Logistic LASSO) to develop a readmission risk‐assessment tool. We identified 2,899 individuals with ADRD who had at least one index hospital admission. All data features available in EMR – demographics, lab results, prior counts of healthcare use, and characteristics of index hospitalization – were included in our predictive models. Additionally, we geocoded the street address of place of residence using the National Neighborhood Data Archive (NaNDA) using the U.S. Census tract‐level information to include two composite measures of socioeconomic status: disadvantage and affluence.ResultsThe readmission rate for ADRD patients was 22% versus 17% for the general population. The best predictive model was the Random Forest (area under the receiver operating characteristic curve = 0.66; sensitivity = 0.64; specificity = 0.61). The accuracy of our model (0.61) was 42% higher than the LACE score (0.43), which is currently used by the hospital for all patients. The top 5 predictors of 30‐day readmission among people with ADRD included length of hospital stay, frailty index, living in a disadvantaged neighborhood, and total prior‐year healthcare charges.ConclusionADRD patients are highly vulnerable and require many resources, with substantially greater readmission risk and elevated rates of other adverse health events. Leveraging EMR data in a readmission risk‐assessment tool can help inform appropriate and efficient coordination and transitions of care for ADRD patients. Our risk‐assessment tool can identify ADRD patients at high risk for readmission and why they are at higher risk. This can enable better decision‐making upon discharge.