Artificial intelligence for the assessment of distal radius fracture instability on wrist radiographs: A Diagnostic Study

Abstract Background Misinterpretation of radiographs of distal radius fractures (DRFs) can lead to malalignment and hand and wrist disability. The application of artificial intelligence (AI) in detecting and predicting fracture instability still needs to be explored. This study aimed to evaluate the diagnostic accuracy of AI in identifying correlated factors, as well as in predicting unstable DRF, Fernandez types 1 and 3. Methods DRF radiographs (Fernandez 1 and 3) of adults aged > 18 years were retrieved from a university hospital and a provincial hospital between January 2017 and May 2022. Radiographs with any concomitant fracture, inadequate imaging, or radiocarpal pathology were excluded. Unstable fracture indicated operative intervention one the basis of 1) the Lafontaine criteria, and 2) radiographic operative factors (metaphyseal comminution and radial shortening). Two AI distal end radius (AIDER) models, AIDER1, identified each radiographic Lafontaine criterion and operative factor, and AIDER2, which predicts Lafontaine criteria ≥ 3 or unstable fractures, were trained, validated, and tested (70:15:15 ratio). Results Among the 1,548 initial radiographs, 1,469 films were eligible. The patients’ characteristics were as follows: 694 (47%) were elderly, 984 (67%) were female, and 840 (57%) were Fernandez type 3. Diagnostic accuracy of the testing groups: AIDER1 achieved the accuracy threshold in dorsal comminution (74%), dorsal tilt > 20 degrees (76%), and metaphyseal comminution (74%). This model provided low accuracy for ulnar styloid fractures, intra-articular fractures, and radial shortening (67%-71%). For AIDER2, the detection of unstable fractures achieved an accuracy of 82%, with an area under the curve (AUC) of 0.63 (95% CI: 0.53, 0.73), suggesting moderate discriminative ability. Conclusion The AI models performed well for selected radiographic features but showed limited accuracy for several operation-related findings. These results indicate that the system may assist with preliminary assessment, but further refinement, broader training data, and external validation are needed before it can be used to support treatment decisions reliably.