Classification of Musculoskeletal Radiograph Requisition Appropriateness Using Machine Learning

Objective: Poor quality imaging requisitions lower report quality and impede good patient care. Manual control of such requisitions is time consuming and can be a source of friction with referring physicians. The purpose of this study was to determine if poor quality requisitions could be identified automatically using machine learning and natural language processing techniques in order to allow for more efficient workflow. Methods: Exam indications from 50 000 musculoskeletal radiograph requisitions were manually classified, reviewed and deemed 'appropriate' or 'inappropriate' by two staff radiologists based on ACR appropriateness criteria. The requisitions were divided into training and test groups (80/20 split). The training set was pre-processed, converted to a bag-of-words model and used to train a Multinomial Naïve Bayes classifier which was then applied to the test set. Results: Out of 50 000 requisitions, 12 253 (24.5%) were deemed to contain an inappropriate indication. A Naive Bayes model correctly classified requisitions with an accuracy of 98%. In the test set, 107 of 7561 (1.4%) appropriate requisitions were incorrectly flagged and 92 of 2439 (3.8%) inappropriate requisitions were not flagged. Conclusions: Accurate automated identification of inappropriate indications on musculoskeletal requisitions is feasible using machine learning and natural language processing.