A deep learning model using pre-reduction radiographs and clinical-demographic data predicted hand surgeons' surgical recommendations for distal radial fractures, reaching 87.14% accuracy and 97% sensitivity on the test set

The study demonstrates that pre-reduction injury radiographs plus clinical and demographic data can predict whether a fellowship-trained hand surgeon would recommend operative intervention for a distal radial fracture. Whereas operative decisions have relied on clinical judgment and radiographic parameters, this work treats the surgeon's recommendation itself as a predictable target and reports quantitative performance for a combined model. Single-institution feasibility study with 1,040 patients, of whom 884 were used for training and 156 for hold-out testing; test-set accuracy 87.14%, sensitivity 97%, specificity 73%, AUC 0.96, Brier score 0.10.

A combined architecture of a convolutional neural network and a random forest can use both imaging and structured clinical data. The CNN outputs from radiographs were combined with clinical and demographic data in a random forest model, rather than relying on an imaging-only or tabular-only model. The methods state that a CNN was trained on pre-reduction injury radiographs and its outputs were combined with clinical and demographic data in a random forest model, evaluated on a hold-out test data set.

Interpretability analyses pointed to clinically understandable bases for predictions. Grad-CAM heatmaps and SHAP explanations attributed predictions to specific image regions and clinical features, rather than only producing a classification. Grad-CAM visualizations indicated the CNN focused on clinically relevant features such as fracture displacement, and SHAP analysis of the random forest highlighted age and lateral wrist radiographs as key contributors.

The result is intended for the initial point-of-care setting, using pre-reduction injury radiographs and clinical and demographic data to predict whether a group of fellowship-trained hand surgeons at one institution would recommend operative versus nonoperative management of a distal radial fracture; the authors position it as a pilot feasibility study and state that future work will focus on external validation, expanding data sets, and incorporating additional imaging features to optimize performance and generalizability. For readers, this suggests the approach could inform patient counseling and reinforce timely follow-up with specialists, with applicability bounded by the institution and surgeon group studied.

Readers should keep in mind that this is a single-institution feasibility study of one surgeon group, and the authors themselves list external validation, expanded data sets, and additional imaging features as future work; how the model performs at other institutions, with other surgeon groups, and under different imaging conditions, and how predicting a surgeon's recommendation relates to actual treatment decisions, remain open questions. In addition, this is the full text without figure details, so the specific Grad-CAM and SHAP visualizations cannot be further checked here.