Predicting potentially avoidable hospitalizations

Abstract

Background: Hospitalizations due to ambulatory care sensitive conditions (ACSCs) are widely accepted as an indicator of primary care access and effectiveness. However, broad early intervention to all patients in a health care system may be deemed infeasible due to limited resources.

Objective: To develop a predictive model to identify high-risk patients for early intervention to reduce ACSC hospitalizations, and to explore the predictive power of different variables.

Methods: The study population included all patients treated for ACSCs in the VA system in fiscal years (FY) 2011 and 2012 (n=2,987,052). With all predictors from FY2011, we developed a statistical model using hierarchical logistic regression with a random intercept to predict the risk of ACSC hospitalizations in the first 90 days and the full year of FY2012. In addition, we configured separate models to assess the predictive power of different variables. We used a random split-sample method to prevent overfitting.

Results: For hospitalizations within the first 90 days of FY2012, the full model reached c-statistics of 0.856 (95% CI, 0.853-0.860) and 0.856 (95% CI, 0.852-0.860) for the development and validation samples, respectively. For predictive power of the variables, the model with only a random intercept yielded c-statistics of 0.587 (95% CI, 0.582-0.593) and 0.578 (95% CI, 0.573-0.583), respectively; with patient demographic and socioeconomic variables added, the c-statistics improved to 0.725 (95% CI, 0.720-0.729) and 0.721 (95% CI, 0.717-0.726), respectively; adding prior year utilization and cost raised the c-statistics to 0.826 (95% CI, 0.822-0.830) and 0.826 (95% CI,0.822-0.830), respectively; the full model was reached with HCCs added. For the 1-year hospitalizations, only the full model was fitted, which yielded c-statistics of 0.835 (95% CI, 0.831-0.837) and 0.833 (95% CI, 0.830-0.837), respectively, for development and validation samples.

Conclusions: Our analyses demonstrate that administrative data can be effective in predicting ACSC hospitalizations. With high predictive ability, the model can assist primary care providers to identify high-risk patients for early intervention to reduce ACSC hospitalizations.