Development of a data-driven method for selecting candidates for case management intervention in a community's medically indigent population
The Indigent Care Collaboration (ICC), a partnership of Austin, Texas, safety net providers, gathers encounter data and manages initiatives for the community's medically indigent patients. One such initiative is the establishment of a care management program designed to reduce avoidable hospitalizations. This study developed predictive models designed to take year-one encounter data and predict inpatient utilization in the following two years. The models were calibrated using 2003 through 2005 data for the 41,260 patients with encounters with ICC partner providers in all three years. Predictor variables included prior inpatient admissions, age, sex, and a summary measure of overall health status: the relative risk score produced by the Diagnostic Cost Groups prospective Medicaid risk-adjustment model. Using the 44,738 patients with encounter data in each of years 2004 through 2006 data, the performance of the predictive models was cross-validated and compared against the performance of the "common sense" method of choosing candidate patients based on prior year chronic disease diagnoses and high utilization, referred to herein as the Utilization Method (UM). The 620 patients with three or more 2005 through 2006 inpatient admissions were considered the actual high use patient subset. Each model's highest-risk 620 patients comprised its high-risk subset. Only 344 high-risk patients met the UM’s criteria. Prediction accuracy was described in terms of positive predictive value (PPV), i.e., the proportion of identified high-risk patients who were high-use patients. Three of the predictive models had a PPV of near 25% or greater, with the highest, the linear model using the DCG relative risk score, at 26.8%. The PPV of the UM was 17.1%, lower than that of all predictive models. When all high-risk subsets were limited to 344 patients (the number identified by the UM), the performance of the UM and the predictive models was similar. This study demonstrated that “common sense” targets for case management can be identified via simple filter as effectively as through empirically-based predictive models. However, once the supply of easily identifiable targets is exhausted, predictive models using a measure of health status identify high-risk patients who could not be easily identified by other means.