A Machine Learning Approach to Predicting Radiographic Outcomes of Nonsurgically Treated Distal Radius Fractures

INTRODUCTION: Several statistical models have been developed to predict the stability of distal radius fractures after closed reduction, but their findings have not been consistently reproduced. We aimed to develop a machine learning (ML) model to predict radiographic outcomes of nonsurgically treated distal radius fractures based on pre-reduction and postreduction radiographic parameters and demographic variables. METHODS: Adults with displaced distal radius fractures at a single institution between 2012 and 2024 were identified through retrospective chart review. Inclusion criteria required closed reduction in the emergency department, with radiographs obtained before reduction, immediately after reduction, and 6 weeks after reduction. At 6 weeks, treatment outcomes were classified as "success" or "failure" based on American Academy of Orthopaedic Surgeons acceptable reduction parameters. Five ML models were trained to predict 6-week outcomes using demographic data and pre-reduction and postreduction radiographic measurements. The 10 parameters with highest Shapley values for predictive ability were used to create an interpretable composite model. RESULTS: Among 1,227 patients, 152 met the inclusion criteria (mean age: 61.4 ± 20.2 years; 75.7% female). The composite model correctly predicted outcomes in 25 of 31 patients, achieving an accuracy, precision, and recall of 81%; area under the curve of 0.84; and F1 score of 0.81. Restoration of postreduction palmar tilt, radial height, and excellent reduction based on the Lindstrom score were most predictive of 6-week radiographic outcomes. The best performing decision tree showed the following cutoffs predictive of treatment failure: +4.7 mm of pre-reduction ulnar variance, 8° of postreduction dorsal tilt, and