Using Deep Learning to Suggest Treatment for Proximal Humerus Fractures.

Abstract

Proximal humeral fractures are among the most common fractures seen in emergency departments. Accurately diagnosing and selecting the most appropriate treatment for these fractures can be challenging, and consultation with a senior orthopedic surgeon can be time-consuming for both the patient and the emergency unit. We developed a machine learning model for predicting the type of treatment based on injury radiographic images. The model distinguishes between nonoperative and operative treatment options, achieving an accuracy of 86% and an interobserver reliability (kappa) of 0.722 for test-dataset, which is more than the interobserver agreement between shoulder surgeons. This model has the potential to serve as a therapeutic decision support system for the practitioners in the emergency departments to expedite treatment decisions and to reduce patients' waiting time.