Predicting cancer patients' risk of potentially avoidable hospitalization

With advances in treatment cancer has become a chronic disease. People with chronic illnesses including cancer have a disproportionate need for health services including hospital care. The Regione Emilia-Romagna (Italy) longitudinal health care database includes demographic information, death registry data, hospital data, specialty physician, laboratory and radiology utilization, and outpatient pharmacy data. We selected adults (age 18+) who were residents in 2000-2005, alive at the end of 2005, and identified as having cancer (n=182,318). Our goal was to predict risk of hospitalization for problems that are potentially avoidable in cancer patients who have diseases or problems amenable to disease management. Independent variables included demographic information, morbidity indicators, and quality of care indicators as well as location of cancer, stage, radiation or chemotherapy, use of selected medications, and treatment at an academic medical center. Risk models were developed using boosted regression trees (BRT), which can accommodate large sets of predictors and are able to include interactions and nonlinear terms in the model adaptively in order to improve predictive performance. Of 182,318 individual identified as having cancer, 13.8% were hospitalized or died in 2006. The model was successful at stratifying patients by risk (C-statistic = 0.79). Of patients in the highest three risk deciles, 48%, 26%, and 19% were hospitalized or died in 2006. These models, which allow the identification of the relative influence of factors predictive of risk of hospitalization, will assist local health authorities in the identification of residents who may benefit from interventions to reduce the risk.