Back to 2026 Abstracts
Deep Learning to Provide Surgical Recommendations for Distal Radius Fractures: A Feasibility Study
Omar Shareef, BS
1, Hailey P. Huddleston, MD
2, Seong Jang, MD
2, Emma T Smolev, MD
3; Duretti T Fufa, MD
2(1)Perelman School of Medicine at the University of Pennsylvania, Philidelphia, PA, (2)Hospital for Special Surgery, New York, NY, (3)Case Western Reserve University, Cleaveland, OH
Introduction: Surgical decision-making for distal radius fractures (DRFs) is nuanced, requiring integration of clinical judgment and radiographic assessment at the time of injury and during follow-up. As one of the most common fractures encountered in emergency departments, DRFs represent a key opportunity for improving triage and referrals. Tools that can assist patients and urgent care providers in determining fracture severity and the potential need for surgical intervention may facilitate timely referral to surgical specialists. This study investigates the feasibility of using artificial intelligence (AI) to predict surgical recommendations for DRFs at initial presentation, based on pre-reduction radiographs and demographic data.
Materials & Methods: A convolutional neural network (CNN) model was trained on pre-reduction injury radiographs, and its outputs were combined with demographic data in a random forest model. The final model was trained on a dataset of 3,330 images from 1,040 patients which were divided into training, validation, and test datasets using a 70-15-15 split. To enhance interpretability, Grad-CAM heatmaps and SHapley Additive exPlanations (SHAP) were employed to identify image regions and clinical features contributing to model predictions.
Results: There was a total of 2,086 images from 632 patients in the operative group and 1,244 images from 408 patients in the non-operative group. The combined model achieved an accuracy of 87.14%, sensitivity of 97%, AUROC of 0.96, and Brier score of 0.10, thus demonstrating its ability to predict surgical recommendations for DRFs. Grad-CAM visualizations indicated that the CNN focused on clinically relevant features (
Figure 1) and SHAP analysis of the RF model highlighted age and lateral wrist radiographs as key contributors to predictions (
Figure 2).
Conclusion: This study demonstrates the feasibility of using AI to predict surgical recommendations for distal radius fractures based on pre-reduction radiographs and demographic data. The CNN model focused on clinically relevant information used in operative decision making, such as age and lateral radiographs as demonstrated on SHAP feature importance. Further work will seek to build upon this study by increasing the interpretability of the models as well as include a greater set of features to be used for inference.
Back to 2026 Abstracts
