Closed reduction of anterior shoulder dislocations (ASDs) carries a risk of iatrogenic humeral neck fractures. However, due to the lack of pre-reduction computed tomography (CT), previous studies failed to distinguish true de novo iatrogenic humeral neck fractures from the iatrogenic displacement of pre-existing occult anatomical neck fractures (ANFs). This study aimed to develop an interpretable “dual-pipeline” machine learning framework to independently predict the risks of these two distinct iatrogenic complications. We retrospectively analyzed patients with ASD treated at a trauma center between July 2019 and December 2025. Based on pre-reduction CT, patients were rigorously stratified into two cohorts. The primary cohort (Model A) included patients with a simple ASD or an isolated greater tuberosity fracture (GTF) to predict de novo iatrogenic humeral neck fractures using multiple machine learning algorithms. Model A was chronologically divided into a training set (July 2019–December 2024, employing the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance) and a temporal validation set (January–December 2025). The subgroup cohort (Model B) comprised patients with pre-existing occult or non-displaced ANFs, utilizing leave-one-out cross-validation (LOOCV) to build a Random Forest model for predicting iatrogenic displacement. Performance was evaluated via metrics, such as the area under the curve (AUC), followed by SHapley Additive exPlanations (SHAP) for interpretability. Model A included 937 patients, of whom 49 (5.2%) sustained de novo iatrogenic humeral neck fractures. The CatBoost model achieved optimal performance (Temporal validation AUC = 0.862, 95% CI 0.794–0.930). SHAP analysis identified a concomitant GTF, advanced age, high-energy trauma, junior operators, osteoporosis, and the Hippocratic maneuver as primary risk factors. Model B included 54 patients, of whom 26 (48.1%) experienced iatrogenic displacement of ANFs. The Random Forest model performed well (AUC = 0.858), with SHAP revealing that advanced age, the Hippocratic maneuver, junior operators, and reduction without anesthesia significantly increased the risk of displacement, whereas axillary manipulation demonstrated a protective effect. This interpretable dual-pipeline machine learning framework successfully predicts the risk factors for two distinct iatrogenic complications during ASD reduction. Clinically, our findings highlight the potential benefit of considering pre-reduction CT scans for elderly ASD patients subjected to high-energy trauma and presenting with a GTF. If an occult ANF is identified, patients should be fully informed of the displacement risk, and gentle axillary manipulation under adequate anesthesia is advised as a preferential option to maximize protection and prevent catastrophic displacement. However, prospective multicenter validation is required before establishing firm clinical recommendations.
Chen et al. (Sat,) studied this question.