JBJS - 2026-09-28 - Journal Article
Ability of Deep Learning to Predict Surgical Recommendations for Distal Radial Fractures: A Feasibility Study.
Shareef O, Huddleston H, Jang SJ, Smolev E, Fufa DT
Topics
Key Takeaway
A combined CNN and random forest model predicted fellowship-trained hand surgeon operative recommendations for distal radial fractures with 87% accuracy, 97% sensitivity, and AUC 0.96 on pre-reduction radiographs.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This single-institution feasibility study trained a convolutional neural network on pre-reduction DRF radiographs, combined its outputs with clinical and demographic data in a random forest classifier, and evaluated performance on a 156-patient hold-out test set. The combined model achieved 87.14% accuracy, 97% sensitivity, 73% specificity, and AUC 0.96. Grad-CAM and SHAP analyses identified fracture displacement, patient age, and lateral wrist radiographs as the dominant predictive features.
Key Limitation
Ground truth is defined by operative recommendations from a single group of fellowship-trained hand surgeons at one institution, meaning the model learns institutional practice patterns rather than an evidence-based universal standard.
Original Abstract
BACKGROUND
Distal radial fractures (DRFs) are extremely common, and decisions regarding operative intervention rely on clinical judgment and radiographic parameters at and after injury. Understanding whether a recommendation will be made for operative versus nonoperative management in an initial point-of-care setting can assist in patient counseling and reinforce timely follow-up with general orthopaedic or hand specialists. This study investigated the feasibility of using artificial intelligence (AI) to predict whether a fellowship-trained hand surgeon would recommend operative intervention using pre-reduction radiographs and clinical and demographic data.
METHODS
A convolutional neural network (CNN) model was trained on pre-reduction injury radiographs, and its outputs were combined with clinical and demographic data in a random forest (RF) model. The final model was evaluated on a hold-out test data set. To enhance interpretability, gradient-weighted class activation mapping (Grad-CAM) heatmaps and SHapley Additive exPlanations (SHAP) were employed to identify image regions and clinical features contributing to model predictions.
RESULTS
Of 1,040 included patients, 884 were used for training and 156 for testing model performance. On the test data set, the combined model achieved an accuracy of 87.14%, sensitivity of 97%, specificity of 73%, area under the receiver operating characteristic curve of 0.96, and Brier score of 0.10. Grad-CAM visualizations indicated that the CNN focused on clinically relevant features, such as fracture displacement, and SHAP analysis of the RF model highlighted age and lateral wrist radiographs as key contributors to predictions.
CONCLUSIONS
This pilot study demonstrates the feasibility of using AI to predict the recommendations of a group of fellowship-trained hand surgeons at 1 institution regarding operative versus nonoperative treatment for a DRF on the basis of pre-reduction injury radiographs and clinical and demographic data. Future work will focus on external validation, expanding data sets, and incorporating additional imaging features to optimize performance and generalizability.
LEVEL OF EVIDENCE
Prognostic Level III . See Instructions for Authors for a complete description of levels of evidence.