Predicting risk of hospitalization in patients with chronic conditions in the Regione Emilia-Romagna, Italy

As the population ages in Italy, more people are living with chronic diseases and are at higher risk of more intensive use of medical care, including hospitalization. The goal of this project is to develop a model to predict risk of hospitalization for patients with conditions that may be amenable to disease management or case management. The 2000-2006 population-based longitudinal healthcare database of Regione Emilia-Romagna, Italy, contains data for the 4.5 million residents of the region including demographic, hospital discharge abstract, outpatient pharmacy, specialty care, and home health data. We included a total of 424 demographic, clinical, and disease severity indicators based on data from 2000-2005 to predict risk of hospitalization or death in 2006. In our dependent variable, we included only hospitalizations (or deaths) for problems that are potentially avoidable in patients with diseases or problems amenable to disease management. For model development, we used boosted regression trees, which can accommodate large sets of predictors and are able to include interactions and nonlinear terms adaptively. We used an enriched case-control training sample of 50,000 individuals, and assessed predictive performance and calibration of the model on an independent validation sample of equal size. We applied the optimized model to the entire population and developed risk categories using deciles of the predicted risk. We evaluated the performance of the model for the entire population and within each risk decile with several outcomes indicative of increasing comorbidity or disease severity. The C-statistic (area under the receiver operating characteristics curve) on validation data was 0.80. The two most influential predictors of hospitalization or death were the patient's number of chronic conditions and age; potentially inappropriate medication prescribing, polypharmacy, and the presence of cardiovascular conditions were prominent predictors. The overall rate of hospitalization or death was 5.9%; in the two highest risk deciles of the entire population, 27% and 11% of the patients were hospitalized or died. These two groups, the top 20% by predicted risk, accounted for 58% of the hospital expenditures in the region. Boosted regression tree methods provide a promising approach to the complex challenge of modeling risk of hospitalization. Our model predicts one-year risk of hospitalization or death for individuals using indicators from the preceding six years. Using this model, healthcare managers at the regional and local level can identify patients at risk of using costly levels of care who may benefit from case management or disease management programs.