Purpose: Patients scheduled for outpatient infusion sometimes may be deferred for treatment after arriving for their appointment. This can be the result of a secondary illness, not meeting required bloodwork counts, or other medical complications. The ability to generate high-quality predictions of patient deferrals can be highly valuable in managing clinical operations, such as scheduling patients, determining which drugs to make before patients arrive, and establishing the proper staffing for a given day.
Methods: In collaboration with the University of Michigan Comprehensive Cancer Center, we have developed a predictive model that uses patient-specific data to estimate the probability that a patient will defer or not show for treatment on a given day. This model incorporates demographic, treatment protocol, and prior appointment history data. We tested a wide range of predictive models including logistic regression, tree-based methods, neural networks, and various ensemble models. We then compared the performance of these models, evaluating both their prediction error and their complexity level.
Results: We have tested multiple classification models to determine which would best determine whether a patient will defer or not show for treatment on a given day. We found that a Bayesian additive regression tree model performs best with the University of Michigan Comprehensive Cancer Center data on the basis of out-of-sample area under the curve, Brier score, and F1 score. We emphasize that similar statistical procedures must be taken to reach a final model in alternative settings.
Conclusion: This article introduces the existence and selection process of a wide variety of statistical models for predicting patient deferrals for a specific clinical environment. With proper implementation, these models will enable clinicians and clinical managers to achieve the in-practice benefits of deferral predictions.