| Abstract ID |
| 20260126 |
| Category |
| Shoulder: Rotator Cuff |
| Preferable Presentation |
| Oral Presentation |
| Title |
| MACHINE LEARNING-BASED RADIOMICS OF PREOPERATIVE PLAIN RADIOGRAPHS FOR PREDICTING RETEAR AFTER ARTHROSCOPIC ROTATOR CUFF REPAIR |
| Author |
|
| Presenter |
| Young Dae Jeon |
| Abstract |
| Background: Retear is a major complication of arthroscopic rotator cuff repair (ASRCR). Although numerous artificial intelligence models have been proposed to predict postoperative retear, prediction using preoperative and clinical radiographic variables has not been evaluated. Identifying high-risk patients using routine radiography remains a critical clinical gap. This study was to develop and validate a machine learning (ML) framework using preoperative plain radiographs and clinical variables to predict the risk of retear after ASRCR, without the need for advanced imaging. Objectives: We analyzed the preoperative data of 310 patients who underwent ASRCR performed by a single surgeon. The predictive models integrated clinical, radiographic, and 4,095 radiomic variables extracted from three anatomical regions of interest on plain radiographs: the greater tuberosity (GT), lesser tuberosity, and subacromial (SA) spur. Five ML classifiers were trained and evaluated with feature selection using the Least Absolute Shrinkage and Selection Operator within a double-loop nested 5-fold cross-validation framework to ensure unbiased performance measurement through cycling test sets. Shapley additive explanation (SHAP) analysis was used to interpret the model and identify the key predictors of retear. Study design: Cohort study, level of evidence III Results: Among the 310 patients, 26 experienced retear. Based on the validation performance, the CatBoost model was selected as the primary predictive model, yielding area under the receiver operating characteristic curves (AUROCs) of 0.826 ± 0.098 and 0.839 ± 0.089 for the validation and test datasets, respectively. SHAP analysis identified the SA spur wavelet-HLL-filtered first-order mean intensity, GT Gaussian filtered GLSZM gray-level non-uniformity, exponential filtered GLCM inverse variance, and the clinical interaction term between age and Hamada grade (Age × Hamada) as the top predictors of retear. Conclusion: An ML model integrating radiomics variables from preoperative plain radiographs with clinical variables could predict retear after ASRCR. The quantitative assessment of routinely obtained radiographs may provide additional preoperative prognostic information. |