Search

Radiologists use an imaging order from the ordering physician, which includes a radiology title, to select the most suitable imaging protocol. Inappropriate radiology titles can disrupt protocol selection and result in mistaken or delayed diagnosis. The objective of this work is to develop an algorithm to predict correct radiology titles from incoming exam order data. The proposed instrument is an ensemble of five decision tree-based machine learning (ML) techniques (Light Gradient Boosting Machine, eXtreme Gradient Boosting Machine, Random Forest, Adaptive Boosting, and Random UnderSampling Boosting Model) trained to recommend radiology titles of computed tomography imaging examinations based on electronic medical records. Issues of imbalanced data and generalization were addressed. The tuned models were used to predict the top three radiology titles for the radiologist revision. The models were evaluated using a 10-fold cross-validation method, yielding an approximate average accuracy of $80.5\% \pm 2.02\%$ and F1-score of $80.3\% \pm 1.67\%$ for all models, while the ensemble classifier (~83% F1-score) outperformed individual models. An accumulated average accuracy of ~92% was obtained for the top three predictions. ML techniques can predict radiology titles and identify highly important features. The proposed system can guide physicians toward selecting appropriate radiology titles and alert radiologists to inconsistencies between the radiology title in the exam order and the patient’s underlying conditions, thereby improving imaging utility and increasing diagnostic accuracy, which favors better patient outcomes.