Predicting Patient No-Shows: Situated Machine Learning with Imperfect Data.

ABSTRACT

Patients who do not show up for scheduled appointments are a considerable cost and concern in healthcare. In this study, we predict patient no-shows for outpatient surgery at the endoscopy ward of a hospital in Denmark. The predictions are made by training machine leaning (ML) models on available data, which have been recorded for purposes other than ML, and by doing situated work in the hospital setting to understand local data practices and fine-tune the models. The best performing model (XGBoost with oversampling) predicts no-shows at sensitivity = 0.97, specificity = 0.66, and accuracy = 0.95. Importantly, the situated work engaged local hospital staff in the design process and led to substantial quantitative improvements in the performance of the models. We consider the results promising but acknowledge that they are from a single ward. To transfer the no-show models to other wards and hospitals, the situated work must be redone.