Predictive modeling to identify scheduled radiology appointments resulting in non-attendance in a hospital setting

Abstract

No-show appointments are a troublesome, but frequent, occurrence in radiology hospital departments and private practice. Prior work in medical appointment no-show prediction has focused on general practice and has not considered features specific to the radiology environment. We collect data from 16 years of outpatient examinations in a multi-site hospital radiology department. Data from the radiology information system (RIS) are fused with patient income estimated from U.S. Census data. Features were categorized into three groups: Patient, Exam, and Scheduling. Models based on the total feature set and separately on each feature group were developed using logistic regression to assess the per-appointment likelihood of no-show. After five-fold cross-validation, no-show prediction using the total feature set from 554,611 appointments yielded an area under the curve (AUC) of 0.770 ± 0.003. Feature groups that were most informative in the prediction of no-show appointments were those based on the type of exam and on scheduling attributes such as the lead time of scheduling the appointment. A data-driven no-show prediction model like the one presented here could be useful to schedulers in the implementation of an automated scheduling policy or the assignment of examinations with a high risk of no-show to lower impact appointment slots.