Background: Unplanned hospital readmissions (HRAs) are very common in cancer patients. These events can potentially impair the patients' health-related quality of life and increase cancer care costs. In this study, data-driven prediction models were developed for identifying patients at a higher risk for HRA.
Methods: A large dataset on cancer pain and additional data from clinical registries were used for conducting a Bayesian network analysis. A cohort of gastrointestinal cancer patients was selected. Logical and clinical relationships were a priori established to define and associate the considered variables including cancer type, body mass index (BMI), bone metastasis, serum albumin, nutritional support, breakthrough cancer pain (BTcP), and radiotherapy.
Results: The best model (Bayesian Information Criterion) demonstrated that, in the investigated setting, unplanned HRAs are directly related to nutritional support (p = 0.05) and radiotherapy. On the contrary, BTcP did not significantly affect HRAs. Nevertheless, the correlation between variables showed that when BMI ≥ 25 kg/m2, the spontaneous BTcP is more predictive for HRAs.
Conclusions: Whilst not without limitations, a Bayesian model, combined with a careful selection of clinical variables, can represent a valid strategy for predicting unexpected HRA events in cancer patients. These findings could be useful for calibrating care interventions and implementing processes of resource allocation.
Keywords: Bayesian analysis; breakthrough cancer pain; cancer pain; hospitalization; predictive models; quality of life.