Accuracy of Key Sonographic Markers for Juxta-Articular Fractures: A Prospective Study Using Explainable Machine...

To systematically evaluate sonographic features of long bone juxta-articular fractures and identify key diagnostic predictors using machine learning-based feature selection.

This prospective single-center diagnostic accuracy study enrolled 121 patients with clinically suspected juxta-articular fractures. Ten predefined sonographic features were assessed, with computed tomography (CT) serving as the reference standard. Key features were identified using shapley additive explanations (SHAP), and the 3 top-ranked features were used to construct a decision tree-based sonographic triad model, hereafter referred to as the sonographic triad model. This model generated a final binary ultrasound diagnosis for each suspected fracture site. Agreement between the final binary ultrasound diagnosis and CT was quantified using Cohen's κ with 95% confidence intervals (CIs). Inter-observer and intra-observer agreement was evaluated using blinded dual-reader review across the cohort. Diagnostic performance was quantified by sensitivity, specificity, predictive values, and the area under the receiver operating characteristic curve (AUC).

Analysis of 209 suspected juxta-articular fracture sites revealed that subcutaneous soft-tissue edema (88.5%), cortical separation (50.7%), and heterogeneous intra-articular effusion (56.9%) were the most prevalent findings. The SHAP analysis identified a specific predictive triad. The 3 primary predictors were subperiosteal hematoma (43.8% contribution), arc-shaped fracture end (28.3%), and heterogeneous intra-articular effusion (26.6%). These 3 features collectively accounted for 98.7% of the total feature importance and demonstrated high reproducibility, with good inter-observer agreement (κ = 0.833-0.980) and excellent intra-observer agreement (κ = 0.987-1.000). The final binary ultrasound diagnosis generated by the sonographic triad model demonstrated near-perfect agreement with CT at the suspected fracture-site level (κ = 0.904, 95% CI: 0.847-0.962). In addition, the sonographic triad model achieved excellent diagnostic performance, with a sensitivity of 93.0%, specificity of 97.2%, positive predictive value (PPV) of 96.8%, negative predictive value (NPV) of 93.8%, and an AUC of 0.966.

The sonographic triad model, based on subperiosteal hematoma, arc-shaped fracture end, and heterogeneous intra-articular effusion, demonstrated high diagnostic accuracy and strong concordance with CT for suspected juxta-articular long-bone fractures. This radiation-free and rapid approach may assist early screening and triage.