Machine learning–based prediction of retear after arthroscopic rotator cuff repair: a multicenter cohort study Authors: Jae-Yoon Baek, Eun-Ji Yoon, Jinyoung Jeong, Yang-Soo Kim, Hyun Seok Song, Hyo-Jin Lee, Hyungsuk Kim, Sung-Hyun Cho, Min Jeon, Do-Hyun Jung, Youn Kyu Lee, Tae Hyung Kim, Jong-Ho Kim Published: 2026 ABSTRACT AbstractBackgroundRotator cuff retears following arthroscopic rotator cuff repair (ARCR) remain a critical clinical challenge, influenced by patient, tear, and rehabilitation factors. Conventional assessment methods struggle to capture this multifactorial complexity, limiting accurate retear prediction and personalized management. The purpose of this study was to develop a machine learning–based predictive model to stratify the risk of structural retear at 6 months post-operatively.MethodsThis study analyzed 1,428 patients with complete datasets from an initial cohort of 3,090 patients who underwent ARCR between January 2011 and May 2024 across four centers. A total of 79 clinical variables were collected, including patient demographics, tear size, range of motion, muscle strength, and pain and functional scores at baseline and at 2-4 months post-operatively. Retear was defined as Sugaya classification type IV or V on standardized 6-month post-operative magnetic resonance imaging. Five machine learning models (Logistic Regression, k-Nearest Neighbors, Support Vector Machine, Convolutional Neural Network, and Multi-Layer Perceptron; [MLP]) were evaluated for retear prediction. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve.ResultsThe MLP model showed the most favorable numerical performance among the evaluated models, with a retear-class F1-score of 0.76 and precision of 0.92 under a random 80:20 split, and the largest area under the receiver operating characteristic curve among the models tested. However, under temporal validation the retear-class F1-score fell to 0.30, indicating limited generalizability over time.ConclusionThe MLP model demonstrated potential for stratifying the risk of magnetic resonance imaging–defined 6-month structural retear after ARCR and may serve as an adjunctive tool for early post-operative risk stratification. Because it relies on post-operative predictors and has been evaluated only within a single hospital network, external validation is required before clinical use.Keywords: Arthroscopy, Rotator cuff repair, Retear prediction, Machine learning, Personalized rehabilitation, Multicenter registry, MLP classifier, F1-score Rotator cuff tears represent one of the most prevalent musculoskeletal disorders in the adult population, with incidence rates increasing substantially with age and affecting up to 40% of individuals over 60 years old.24,48 These injuries impose a significant clinical and economic burden, often resulting in chronic pain, functional limitation, and substantial health-care costs related to both conservative management and surgical intervention.1,41 Despite advances in arthroscopic repair techniques and post-operative rehabilitation strategies, outcomes remain unpredictable, with wide variability in tendon healing and functional recovery—even among patients with seemingly similar clinical profiles.27,44 This heterogeneity arises from complex interactions between patient-related factors (eg, age, comorbidities, activity level), tear characteristics (eg, size, chronicity, muscular atrophy), surgical variables, and rehabilitation protocols. Such multifactorial influences make it difficult to identify patients most likely to benefit from surgery or to anticipate post-operative recovery trajectories.23,26,37 Traditional assessment methods often lack the capacity to fully account for this complexity, underscoring the need for more sophisticated tools that can guide personalized treatment decisions and improve prognostic accuracy. Conventional approaches to rotator cuff outcome prediction rely predominantly on clinical examination findings, imaging-based assessments, and standardized scoring systems to guide treatment decisions and prognostic evaluation.25,30,31,40 Magnetic resonance imaging (MRI) parameters, including tear size and retraction, fatty infiltration, and muscle atrophy, serve as primary determinants for surgical candidacy and outcome prediction.29,47 Clinical scoring systems such as the American Shoulder and Elbow Surgeons (ASES) score and Constant Murley Shoulder Score attempt to quantify functional status and monitor recovery.2,10,11 However, these methods struggle to process multidimensional clinical data, cannot capture non-linear variable interactions, and do not adequately account for the complex temporal dynamics of tissue healing.5,9,16,40 Together with the subjective nature of clinical assessment, these limitations contribute to substantial variability in outcome prediction, which is particularly problematic in arthroscopic rotator cuff repair (ARCR), where accurate prediction of retear risk could improve patient outcomes and health-care resource utilization.7,32,41 Machine learning approaches have shown promising results in predicting retear rates and functional outcomes after ARCR,3,4,9,17 but existing models typically focus on single end points and do not provide comprehensive guidance for clinical decision-making and rehabilitation planning.1,13,44 This study addresses this gap by developing a machine learning model focused on risk stratification of MRI-defined 6-month structural retear. Stratifying patients according to 6-month structural retear risk during the early post-operative period may support patient counseling and post-operative monitoring.36 Although the present study focused on retear risk prediction, the model's integration of pre-operative and post-operative variables may help identify patients who warrant closer post-operative surveillance. Leveraging these insights in future analyses may assist in guiding personalized rehabilitation approaches. We therefore hypothesized that integration of multidimensional clinical and imaging variables using a machine learning framework would enable risk stratification of MRI-defined structural retear at a standardized 6-month post-operative time point. Materials and methods Data collection A clinical dataset was collected from patients and was used to predict retear occurrence after ARCR. The dataset includes both pre-operative and post-operative clinical information of patients who underwent ARCR for rotator cuff tears between January 1, 2011, and May 7, 2024, across four centers within Department of Orthopedic Surgery, St. Mary's Hospital, College of Medicine, The Catholic University of Korea network. ARCR procedures were performed by experienced orthopedic surgeons. Patients who underwent revision surgery or procedures due to post-operative infection were excluded from this study. Retear occurrence was identified on follow-up imaging examinations and defined as Sugaya classification type IV or V (Fig. 1). Workflow of suggested machine learning model. VAS, visual analog scale; BMI, body mass index; ROM, range of motion; MTT, manual muscle test; KSS, Korean Shoulder Score; CSS, Constant Murley Shoulder Score; MRI, magnetic resonance imaging. To avoid imputing large amounts of missing data in the primary analysis, only patient records with complete information for all selected parameters were included in model training and evaluation. Among the 3,090 available records, 1,662 were excluded for insufficient clinical documentation, based on predefined inclusion criteria requiring complete longitudinal clinical data and consistent outcome assessment. Missing values were concentrated in post-operative functional scores and clinical indicators, and also affected the outcome (postoperative [POD] 6M retear status); the full pattern is given in Table I. After these exclusions, 1,428 cases met the inclusion criteria, having complete datasets and a minimum post-operative follow-up of 6 months. The detailed list of parameters used in this study is provided in the Variables section below. Table IFrequency and percentage of missing clinical variables among the 1,662 patients excluded for incomplete data.Clinical variableNumber of excluded patients with missing valuesPercentage of missing values among excluded patientsPOD 4M CSS1,57694.8%POD 4M ASES1,57694.8%POD 4M KSS1,57694.8%POD 3M KSS1,57594.7%POD 3M CSS1,57594.7%POD 4M VAS (resting)1,57594.7%POD 4M VAS (activity)1,57394.6%POD 3M ASES1,57194.5%POD 4M MMT sec1,56494.1%POD 4M MMT grade1,56494.1%Pre ASES1,56394.0%Pre KSS1,56394.0%Pre CSS1,56293.9%Pre VAS (resting)1,56293.9%Pre VAS (activity)1,55693.6%POD 3M VAS (resting)1,55293.3%POD 3M VAS (activity)1,54893.1%POD 2M MMT sec1,54392.8%POD 2M MMT grade1,54392.8%POD 6M VAS (resting)1,54292.7%POD 6M KSS1,52992.0%POD 6M CSS1,52491.7%POD 6M VAS (activity)1,52191.5%POD 6M ASES1,52191.5%POD 3M MMT grade1,49489.8%POD 3M MMT sec1,49489.8%POD 6M MMT sec1,48289.1%POD 6M MMT grade1,48289.1%POD 4M add1,46888.3%POD 4M ER-abd1,46788.2%POD 4M ER-side1,46788.2%POD 4M IR1,46488.0%Pre MMT sec1,46287.9%Pre MMT grade1,45587.5%POD 4M FF1,44186.7%POD 3M add1,33380.2%POD 2M add1,32779.8%POD 2M IR1,32579.7%POD 3M IR1,32179.4%Tear size (AP, cm)1,31278.9%POD 3M ER-abd1,30078.2%POD 3M ER-side1,30078.2%POD 2M ER-abd1,28977.5%POD 2M ER-side1,28977.5%POD 6M add1,28977.5%POD 6M ER-side1,28277.1%POD 6M ER-abd1,28277.1%POD 3M FF1,26676.1%POD 6M IR1,26576.1%Pre add1,26375.9%Pre ER-abd1,24775.0%Pre ER-side1,24474.8%POD 6M FF1,24074.6%POD 2M FF1,23774.4%Tear size (ML)1,21172.8%Tear size (retraction)1,21172.8%Pre IR1,18671.3%Pre FF1,16670.1%Surgery side (Rt./Lt.)1,12467.6%BMI26115.7%Height25415.2%Weight784.6%Age90.5%Sex80.4%BMI, body mass index; AP, anteroposterior; ML, mediolateral; FF, forward flexion; ER-side, external rotation with the arm at the side; ER-abd, external rotation with the arm in 90° abduction; add, adduction; IR, internal rotation; VAS, visual analog scale; MMT, manual muscle test; KSS, Korean Shoulder Score; CSS, Constant Murley Shoulder Score; ASES, American Shoulder and Elbow Surgeons; POD, postoperative; Rt., right; Lt., left.Percentages are calculated with the number of excluded patients (n = 1,662) as the denominator Frequency and percentage of missing clinical variables among the 1,662 patients excluded for incomplete data. BMI, body mass index; AP, anteroposterior; ML, mediolateral; FF, forward flexion; ER-side, external rotation with the arm at the side; ER-abd, external rotation with the arm in 90° abduction; add, adduction; IR, internal rotation; VAS, visual analog scale; MMT, manual muscle test; KSS, Korean Shoulder Score; CSS, Constant Murley Shoulder Score; ASES, American Shoulder and Elbow Surgeons; POD, postoperative; Rt., right; Lt., left. Percentages are calculated with the number of excluded patients (n = 1,662) as the denominator A total of 1,428 patient records were analyzed and used for the proposed artificial intelligence (AI) model development. The data were randomly split into a training set (1,142 cases) and a test set (286 cases) at a ratio of 8:2 to develop and evaluate the machine learning model. This retrospective multicenter cohort study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Institutional Review Board of The Catholic University of Korea, Yeouido St. Mary's Hospital (IRB No. SC24WISI0061). The requirement for informed consent was waived due to the retrospective nature of the study. De-identified data were extracted from the Clinical Data Warehouse and used for analysis. Post-operative imaging protocol All patients underwent post-operative MRI at approximately 6 months as part of the standard institutional follow-up protocol across all centers (mean, 6.7 ± 0.5 months; median, 6; interquartile range, 6-7). Tendon integrity was assessed by Sugaya classification, based primarily on coronal oblique and sagittal oblique T2-weighted images interpreted according to routine institutional practice. MRI examinations were independently interpreted at each participating hospital by board-certified musculoskeletal radiologists who were blinded to the model development process and clinical status. Retear was defined strictly as Sugaya type IV or V on post-operative MRI. Cases demonstrating partial thinning (Sugaya type III) were not classified as retears in order to maintain a conservative definition of structural failure. Variables and definitions The dataset encompassed a comprehensive set of clinical variables, which were categorized as follows: (1) basic patient information, including age, sex, height, weight, body mass index, comorbidities (diabetes mellitus, hypertension, coronary heart disease, cerebrovascular accident, dyslipidemia, hyperthyroidism, hypothyroidism, cancer, hepatitis B, and tuberculosis), and treating hospitals; (2) tear characteristics, such as surgery side (right or left), anteroposterior and mediolateral tear length, retraction degree; (3) pre-operative functional indicators, including range of motion (ROM; forward flexion [FF], external rotation with the arm at the side [ER-side], external rotation with the arm in 90° abduction [ER-abd], adduction [add], and internal rotation [IR]), pain scores (visual analog scale at rest and during activity), muscle strength (manual muscle test [MMT] grade and holding time, MMT grade was determined at 90° of FF in the scapular plane by comparison with the contralateral limb according to the conventional Medical Research Council 0-5 grading scale. Holding time was measured according to the endurance component of the Korean Shoulder Score (KSS), with patients standing and holding a 2-kg dumbbell at 45° of FF in the scapular plane, with the palm down and the elbow straight; holding the position for more than 10 seconds was assigned the maximum score.), and functional scores (the ASES score, the CSS, and the KSS); and (4) post-operative functional recovery indicators, including ROM (FF, ER-side, ER-abd, add, and IR) and muscle strength at POD 2M, with additional visual analog scale, ASES, CSS, and KSS scores at POD 3M and POD 4M. Proposed artificial intelligence model for retear risk prediction Fig. 2 illustrates the overall framework of the proposed retear risk prediction model, consisting of a training step (a) and an inference step (b). Pre-operative variables and post-operative clinical indicators collected through POD 4M were used as inputs to estimate POD 6M retear probability (RP). RP was defined as the predicted probability of MRI-defined structural retear, corresponding to Sugaya type IV or V, at 6 months after ARCR. In the training step, the classifiers learned the relationship between the input variables and the POD 6M retear label. In the inference step, the trained model generated an individualized RP for each patient. Overall framework of the proposed method for predicting rotator cuff retear risk after ARCR. The system consists of an AI-based model that estimates the retear risk probability at six months post-surgery (POD 6M RP). ARCR, arthroscopic rotator cuff repair; AI, artificial intelligence; POD, postoperative. Preprocessing For prediction of POD 6M RP, 79 features derived from rotator cuff patient data were used. These data included basic patient information and clinical indicators collected at POD 2M, 3M, and 4M. Preprocessing was conducted using objective and reproducible criteria. Specifically, samples containing at least 1 missing feature value were excluded from the primary analysis, and no subjective manual correction was performed. Categorical variables were encoded using one-hot encoding before model training. Class imbalance was addressed by applying synthetic minority over-sampling technique (SMOTE) only to the training set after data splitting. The minority class was oversampled to match the size of the majority class, and a fixed random seed (random_state = 17) was used to ensure reproducibility. Feature scales were harmonized using z-score standardization. For each feature, the mean was subtracted and the result was divided by the standard deviation, with both statistics estimated from the training set and then applied to the training and test data. This procedure standardized each feature to a mean of 0 and a standard deviation of 1, thereby minimizing bias arising from differences in feature scale during model training. Model architecture To predict RP, this study evaluated a variety of machine learning and deep learning–based binary classifiers. The models and algorithms considered for evaluation include Logistic Regression, Convolutional Neural Network, k-Nearest Neighbor (k-NN), Support Vector Machine, and Multi-Layer Perceptron (MLP).6,35 A comparative evaluation of the different AI models is presented in Table II. Among the evaluated models, the MLP-based binary classifier showed the highest retear prediction accuracy and F1-score.17 The MLP comprised four fully connected layers: three hidden layers with 256, 512, and 128 units and a single-unit output layer (Fig. 3), taking 79 input features. Each hidden layer was followed by batch normalization, a LeakyReLU activation, and dropout (P = .5) to mitigate overfitting; a sigmoid was applied to the output logit at inference to produce a probability for binary classification. The network was trained with the Adam optimizer (β1 = 0.90, β2 = 0.999, ε = 1 × 10−8, weight decay = 0.01). Table IIComparison of performance metrics across five classification models: Logistic Regression, k-NN, SVM, CNN, and MLP.ModelAccuracyPrecisionRecallF1-scoreClass 0Class 1Class 0Class 1Class 0Class 1Logistic regression0.710.980.150.700.820.820.25CNN0.850.960.190.870.470.920.27k-NN0.610.600.760.980.110.740.19SVM0.750.980.150.750.710.850.25MLP0.980.980.920.990.650.990.76CNN, Convolutional Neural Network; k-NN, k-Nearest Neighbor; SVM, Support Vector Machine; MLP, Multi-layer Perceptron.The evaluation metrics used include accuracy, precision, recall, and F1-score. The highest values for each metric are highlighted in bold. Comparison of performance metrics across five classification models: Logistic Regression, k-NN, SVM, CNN, and MLP. CNN, Convolutional Neural Network; k-NN, k-Nearest Neighbor; SVM, Support Vector Machine; MLP, Multi-layer Perceptron. The evaluation metrics used include accuracy, precision, recall, and F1-score. The highest values for each metric are highlighted in bold. Structure of the MLP model used in this study. MLP, Multi-Layer Perceptron; POD, postoperative. Process of training The RP prediction AI model receives individual patient data from the dataset as input and predicts the POD 6M RP. Fig. 4 illustrates the training process of the proposed AI model. Schematic diagram of the training process for the AI model predicting POD 6M retear probability (RP). The model takes preprocessed patient data x as input and outputs a predicted RP value fθ(X), which is compared against the ground truth retear label y using a binary cross-entropy loss function. Gradients Gθ are computed via backpropagation and used to update model parameters θ using Adam optimizer to minimize the prediction error. AI, artificial intelligence. The preprocessed feature vector is fed into the model, which outputs the predicted POD 6M RP, f_θ(x) ∈ [0,1], where values closer to 1 indicate higher retear risk. The model is trained by minimizing the binary cross-entropy (BCE) loss between f_θ(x) and the observed retear label y (Equation 1). Model parameters were optimized with the Adam optimizer (β1 = 0.90, β2 = 0.999, ε = 1 × 10−8, weight decay = 0.01).43 The model was trained with a learning rate of 5 × 10−5 and a batch size of 128, for up to 50,000 epochs with early stopping based on validation loss (patience = 200), implemented in PyTorch 2.4.0 on a single GPU (NVIDIA GeForce RTX 3060Ti). Hyperparameter tuning was performed using only the training and validation data. The held-out test set was not used for hyperparameter tuning or model selection and was reserved exclusively for final model evaluation. Process of inference At inference, the trained model predicts POD 6M RP for new patients and evaluates the model's performance (Fig. 5). The predicted RP fθ(x) was binarized at a threshold T of 0.5 to classify patients as high (yˆ = 1) or low (yˆ = 0) risk (Equation 2). In addition, to assess the robustness of the primary complete-case analysis, a post hoc multiple-imputation–based sensitivity analysis was performed in patients who had originally been excluded because of incomplete data. Missing predictors were imputed using an iterative multivariable imputation approach, and the fixed MLP model from the primary analysis was applied to the imputed datasets. Schematic diagram of the inference process using the trained AI model to predict POD 6M retear probability (RP). The model takes preprocessed patient data x as input and outputs a predicted probability fθ(X) between 0 and 1. This value is then converted into a binary classification label yˆ based on a threshold T (eg, 0.5), indicating high or low risk of retear. AI, artificial intelligence; POD, postoperative. Results Patient demographics Baseline characteristics of patients with and without retear at 6 months are summarized in Table III. Anteroposterior and mediolateral tear sizes were significantly larger in the retear group, and hypertension and dyslipidemia were more frequent; sex, body mass index, follow-up imaging timing, and other comorbidities did not differ significantly between groups. Table IIIBaseline demographic and clinical characteristics of patients with and without retear at 6 mo.FeaturesNo retear (class 0, n = 1,342)Retear (class 1, n = 86)P valueFemale (%)800 (59.6%)52 (60.5%).91BMI (male)25.6 ± 5.125.8 ± 2.9.82BMI (female)24.8 ± 3.224.6 ± 3.2.77Age66.1 ± 9.567.4 ± 8.1.17Tear size (AP, cm)1.4 ± 0.81.8 ± 1.0<.001Tear size (ML)1.5 ± 0.91.9 ± 1.0<.001DM133 (9.9%)8 (9.3%)>.99HTN208 (15.5%)26 (30.2%)<.001Dyslipidemia120 (8.9%)22 (25.6%)<.001Hyperthyroidism6 (0.4%)1 (1.2%).35Hypothyroidism19 (1.4%)0 (0.0%).62Osteoporosis14 (1.0%)3 (3.5%).08Cancer15 (1.1%)1 (1.2%)>.99Mean ± standard deviation of each parameter.BMI, body mass index; AP, anteroposterior; ML, mediolateral; DM, diabetes mellitus; HTN, hypertension. Baseline demographic and clinical characteristics of patients with and without retear at 6 mo. Mean ± standard deviation of each parameter. BMI, body mass index; AP, anteroposterior; ML, mediolateral; DM, diabetes mellitus; HTN, hypertension. Dataset description Of 1,428 included patients, 1,142 were allocated to the training set and 286 to the test set. Within the training set, 69 patients experienced retear and 1,073 did not (Table IV); the test set comprised 17 retear and 269 no-retear cases. Table IVClass distribution of the training dataset before and after applying SMOTE.Sampling statusClass 0Class 1Before SMOTE1,07369After SMOTE1,0731,073SMOTE, synthetic minority over-sampling technique.The number of minority class (class 1: retear cases) samples was increased through synthetic over-sampling to address class imbalance and improve model performance. Class distribution of the training dataset before and after applying SMOTE. SMOTE, synthetic minority over-sampling technique. The number of minority class (class 1: retear cases) samples was increased through synthetic over-sampling to address class imbalance and improve model performance. The prediction target, POD 6M retear occurrence, is formulated as a binary classification task. Class 1 indicates cases where a retear occurred at 6 months post-operative (POD 6M), and class 0 indicates cases without retear. Among the total dataset, there were 86 cases in class 1 and 1,342 cases in class 0, revealing a severe class imbalance. This imbalance can degrade the performance of typical classification models and is a major factor in reducing the sensitivity to the minority class.38 To address this issue, the SMOTE was applied to the training set to increase the proportion of retear cases.8 SMOTE generates synthetic minority-class samples by interpolating between existing minority-class cases. In our setting, the number of retear training samples was increased from 69 to 1,073, matching the 1,073 no-retear samples and resulting in 2,146 training instances after SMOTE (Table IV). The test set remained unchanged to maintain natural class distribution for model evaluation. Evaluation of retear risk prediction model To evaluate the predictive performance of various machine learning algorithms for retear risk classification, we tested five representative models, including Logistic Regression, k-NN, Support Vector Machine, Convolutional Neural Network, and MLP. For a fair comparison, all models were trained and tested using the same preprocessed training and test sets. The evaluation metrics included Accuracy, Precision, Recall, and F1-score.12 Considering the class imbalance present in the dataset, model performance was primarily assessed based on the metrics for Class 1 (retear cases). While accuracy represents the proportion of correct predictions, in this study—where most patients belong to class 0—high accuracy alone is not a sufficient indicator of model effectiveness under class imbalance; a model predicting all samples as class 0 could still achieve high accuracy while failing to detect the clinically critical class 1 cases. For class 1, precision refers to the proportion of patients predicted as retear who actually experienced retear, and recall to the proportion of actual retear cases correctly identified. We therefore evaluated performance primarily using precision, recall, and the F1-score. As shown in Table II, the MLP showed the most balanced performance, with class 1 precision of 0.92 and F1-score of 0.76; logistic regression showed high class 1 recall but lower overall balance, whereas k-NN had the weakest minority-class performance. The MLP also had the largest area under the receiver operating characteristic curve (AUROC; Fig. 6) and was retained over deeper architectures because given the complete-case sample size and the limited number of retear events, added complexity was likely to increase overfitting without clear predictive benefit.46 Given the inherent class imbalance and low retear incidence, these findings support its potential as an adjunctive tool for early post-operative risk stratification of MRI-defined 6-month retear. Receiver operating characteristic curves generated for all trained models. AUROC, area under the receiving operator characteristic curve; MLP, multi-layer perception; CNN, convolution Neural Network; k-NN, k-Nearest Neighbor; SVM, Support Vector Machine. Because the test set contained only 17 retear events, we quantified uncertainty by bootstrap resampling (1,000 iterations). The 95% confidence intervals were wide, particularly for recall (AUROC 0.82 [0.68-0.94]; precision 0.92 [0.70-1.00]; recall 0.65 [0.38-0.86]; F1 0.76 [0.53-0.90]), indicating that the point estimates should be read with caution. The confusion matrix, calibration (Brier score 0.03), and decision-curve analysis are shown in Fig. 7; net benefit was positive across threshold probabilities of about 0.05-0.50, peaking near 0.04 at a threshold of 0.20. Decision curve. MLP, multi-layer perception; SMOTE, synthetic minoirty over-sampling technique. To further explore the robustness of the primary complete-case analysis, we performed a post hoc multiple-imputation–based sensitivity analysis in 1,662 patients who had originally been excluded because of incomplete data. This analysis was performed after the primary train–test split and model training had been completed, so no information from the excluded cohort influenced model fitting. Missing predictor values were imputed five times to generate five imputed datasets; only predictor variables were imputed, and the outcome (POD 6M retear occurrence) was not. Records without an observed outcome were excluded from this analysis, and the observed outcome was used only as the ground-truth label. Across the five imputed datasets, the fixed MLP model achieved a mean AUROC of 0.82 and an overall accuracy of 0.93. Class-wise, the mean precision, recall, and F1 were 0.93, 0.99, and 0.96 for the non-retear class and 0.73, 0.71, and 0.72 for the retear class. We evaluated the model at two levels. First, a center-exclusion sensitivity analysis, removing each hospital from the test set only, showed that no single center drove the pooled result, with retear-class performance broadly comparable across subsets (Table V). Second, we performed temporal validation: the model was trained on patients treated in 2011-2022 (n = 1,117; 62 retears) and tested on those treated in 2023-2024 (n = 311; 24 retears) (Table VI). The retear-class F1-score was 0.30 under temporal validation, compared with 0.76 under the random 80:20 split. Table VPerformance of the final fixed MLP model in the center-exclusion sensitivity analysis on the held-out test set after sequential removal of each treating hospital.Excluded hospitalNumber of class 0Number of class 1AccuracyPrecisionRecallF1-scoreClass 0Class 1Class 0Class 1Class 0Class 1Yeouido St. Mary's Hospital14730.9930.9931.0000.9971.0000.6670.800Eunpyeong St. Mary's Hospital235170.9720.9750.9960.9850.9170.6470.759St. Vincent's Hospital264170.9750.9780.9960.9870.9170.6470.759Seoul St. Mary's Hospital161140.9660.9700.9940.9820.9000.6430.750MLP, Multi-Layer Perceptron. Performance of the final fixed MLP model in the center-exclusion sensitivity analysis on the held-out test set after sequential removal of each treating hospital. MLP, Multi-Layer Perceptron. Table VITemporal validation of the MLP model.ClassPrecisionRecallF1Non-retear (class 0)0.940.990.96Retear (class 1)0.560.210.30MLP, Multi-Layer Perceptron. Temporal validation of the MLP model. MLP, Multi-Layer Perceptron. Explainability analysis of retear risk prediction model To examine which clinical variables most strongly informed the predictions of the MLP model, we performed complementary explainability analyses using Integrated Gradients, SHAP estimated using Monte Carlo Shapley-value approximation, and permutation importance. Fig. 8 illustrates the global explainability results of the MLP model based on the variables included in the final prediction model. Across methods, post-operative ROM- and MMT-related variables were consistently among the highest-ranked predictors, although the exact ranking order varied somewhat by method. Global explainability analysis of the retear risk prediction model using SHAP values, Integrated Gradients, and permutation importance. SHAP, SHapley Additive exPlanation. The single most influential feature by mean absolute SHAP value was POD 4M MMT grade. Several post-operative ROM measures were also repeatedly highlighted, including POD 2M FF, POD 2M IR, POD 4M IR, POD 4M ER-side, POD 2M ER-abd, and POD 2M add. In the permutation analysis, permuting POD 2M FF, POD 4M ER-side, POD 4M MMT grade, and POD 3M FF produced marked decreases in the test-set F1-score, indicating that model performance was particularly sensitive to these variables. When features were grouped by clinical category, post-operative ROM showed the largest cumulative contribution, followed by other post-operative clinical measures. Focused analysis of ROM and MMT variables confirmed that POD 4M MMT grade and several early-to-mid post-operative ROM variables contributed substantially to prediction (Fig. 9). Focused explainability analysis of post-operative ROM and MMT variables in the retear risk prediction model. Orange bars indicate ROM variables and green bars indicate MMT variables. MMT, manual muscle test; ROM, range of motion. Discussion Using a comprehensive dataset of 1,428 fully documented cases derived from an initial multicenter registry of 3,090 ARCR cases, we identified a MLP model as the optimal classifier, achieving an F1-score of 0.76 for retear prediction under a random 80:20 split, despite significant class imbalance. Among the algorithms evaluated, the MLP showed the highest retear-class F1-score and AUROC. Compared to previously published AI-based studies in shoulder surgery reviewed by Gupta et al (2023)—many of which reported moderate AUROC values (typically 0.65-0.79) and lacked integration of explainability—our model showed a comparable-to-higher AUROC (0.83) under random-split evaluation, although this estimate should be regarded as an upper bound. Although structural integrity at 6 months does not necessarily align with concurrent patient-reported outcomes, the clinical impact of retear may become more evident over time. Prior longitudinal work has shown that patients with retear may have worse pain and functional outcomes at later follow-up despite limited early differences.28 In this context, early post-operative stratification of MRI-defined structural retear risk may still be clinically meaningful by identifying patients who may benefit from closer post-operative monitoring and more individualized follow-up. These findings align with emerging evidence supporting machine learning applications in orthopedic surgery. Recent systematic reviews have demonstrated enhanced predictive accuracy across multiple subspecialties, including joint arthroplasty, spinal surgery, and sports medicine.14,18,39,45 Studies in ARCR have consistently shown superior performance of machine learning models compared to traditional clinical scoring systems.1,16,41 Similarly, investigations in joint arthroplasty have reported comparable improvements in predicting complications and functional outcomes.19,34 However, these studies vary in design and validation, and our own temporal-validation results caution against assuming broad applicability without external validation. Nevertheless, the clinical value of retear risk stratification extends beyond technical performance metrics. Studies demonstrate that surgical and conservative treatments can yield comparable outcomes in certain degenerative rotator cuff tear populations, highlighting the critical need for models that identify patients most likely to benefit from surgical intervention while minimizing unnecessary procedures.32 In principle, if adequately validated, such predictive models could help optimize resource allocation and support treatment decision-making in ARCR.15 Our center-exclusion and temporal-validation analyses bear directly on how the model's performance should be interpreted. The center-exclusion analysis establishes robustness to single-center influence but does not test transfer to a new institution or a future period. The random-split estimate should be read as an optimistic upper bound because randomly partitioning a cohort collected over 13 years allows patients from the same period to fall on both sides of the split, whereas the temporal result—markedly lower—represents a more realistic lower bound. Together with the small number of retear events in the test set, this indicates that the model does not yet generalize reliably over time and should be regarded as a hospital-network-specific tool requiring local recalibration and external validation before clinical use. Consistent with this caution, not all studies report uniformly positive results. Jacobien and colleagues (2024)22 reported that machine learning failed to outperform conventional competing risk modeling for revision arthroplasty prediction, suggesting context-dependent effectiveness. Discrepancies across studies may reflect differences in data preprocessing, feature selection, and algorithm choice, as well as variations in patient populations and surgical techniques.33 Some investigations have further emphasized that generic machine learning approaches require careful customization and validation to achieve optimal performance across diverse orthopedic applications,20,21 underscoring the importance of rigorous methodology and appropriate model selection for clinical implementation.16 A major strength of this investigation is the use of a multicenter complete-case analytical cohort of 1,428 patients derived from an initial registry of 3,090 ARCR cases across four institutions within the Catholic Medical Center network. Regular interinstitutional conferences are held within this network to harmonize documentation protocols and support consistent clinical recording across sites. In addition, the dataset included a broad range of clinical variables, tear-related imaging features, post-operative functional measures, and patient-reported outcomes, allowing assessment of the multifactorial determinants of rotator cuff repair healing.42 While previous machine learning studies in ARCR have largely focused on isolated outcome measures, our framework lays the groundwork for future multi-dimensional modeling that may include recovery trajectory analysis and rehabilitation planning.1,4 Existing clinical scoring systems remain non-standardized and often fail to account for the complex interplay between patient-specific factors, tear characteristics, and post-operative variables that jointly influence tendon healing.13 By leveraging our large multicenter dataset and a class-imbalance appro handling ach, the present model supports enhanced prognostic counseling, early identification of high-risk patients, and more individualized collaboration with rehabilitation specialists. Among the oversampling and cost-sensitive strategies we compared, no single approach was uniformly superior: SMOTE yielded the highest recall for the retear class without collapsing precision, whereas a class-weighted loss achieved higher overall discrimination (Table VII). We prioritized minority-class recall, given that a missed retear is clinically more consequential than a false alarm. Table VIIComparative performance analysis of training strategies.Training strategyAUROCF1-scorePrecisionRecallOur method (SMOTE)0.820.760.920.65Class-weighted BCE loss0.860.550.560.53Focal loss (gamma = 2.0)0.850.520.700.41BCE loss (no SMOTE)0.760.570.730.47Penalized logistic regression0.800.220.130.59AUROC, area under the receiver operating characteristic curve; BCE, binary cross-entropy; SMOTE, synthetic minority over-sampling technique; SMOTE, synthetic minority over-sampling technique. Comparative performance analysis of training strategies. AUROC, area under the receiver operating characteristic curve; BCE, binary cross-entropy; SMOTE, synthetic minority over-sampling technique; SMOTE, synthetic minority over-sampling technique. Explainability analyses consistently identified post-operative ROM and MMT variables as the principal predictors across all methods, with POD 4M MMT grade and early-to-mid ROM measures contributing most prominently. These findings should not be read as evidence that early functional deficits cause retear. Reduced post-operative ROM and MMT may instead reflect early functional impairment arising from a repair that has already failed or is in the process of failing, in which case these variables are markers of structural failure rather than its determinants. They therefore carry prognostically relevant information for subsequent structural integrity, but reflect predictive contribution rather than causal effect and should not be interpreted as modifiable targets or as direct guidance for rehabilitation prioritization. Several limitations warrant acknowledgment. First, the retrospective design and complete-case analysis may have introduced selection bias, and unmeasured surgeon- and procedure-related factors—including repair technique, canchor configuration, suture construct, intraoperative decision-making, concomitant procedures, and follow-up practices—may have contributed residual confounding. The post hoc multiple-imputation sensitivity analysis showed that the model behaved reasonably on imputed versions of the excluded cohort. However, this analysis reduces but does not eliminate concern about selection bias: because patients were excluded precisely because their data were incomplete, they may differ systematically from those with complete data in ways that imputation—anchored to the complete-case distribution—cannot fully recover. The sensitivity analysis should therefore be read as a partial robustness check rather than as evidence that selection bias is absent. Second, the study cohort predominantly consisted of small to medium-sized tears, with relatively fewer large or massive tears, which may limit the broader application of the model, particularly in patients with more advanced rotator cuff pathology. Furthermore, structural integrity assessed at 6 months may not fully reflect long-term tendon healing or patient-reported outcomes, as structural and clinical outcomes are not always directly correlated. Additionally, because the current model incorporates post-operative variables up to POD 4M, it should be interpreted as an early post-operative risk-stratification model for MRI-defined 6-month structural retear, rather than as a purely pre-operative prediction model. This limits its immediate utility for pre-operative counseling and highlights the need for future models based on pre-operative variables and prospective longitudinal validation. Future prospective validation in real-world settings is essential, alongside integration into electronic medical records and clinical decision support systems.39 Successful deployment will require sustained collaboration among orthopedic surgeons, data scientists, and health-care informaticians to ensure interpretability, workflow compatibility, and patient safety. Conclusion In this multicenter retrospective cohort, a machine-learning model incorporating pre-operative and early post-operative variables demonstrated potential for stratifying risk of MRI-defined structural retear at 6 months after ARCR. Because the model relies on post-operative predictors, has been evaluated within a single hospital network, and had few positive events in the test set, further external validation is required before clinical implementation. Disclaimers Funding: No funding was disclosed by the authors. Conflicts of interest: Jong-Ho Kim acknowledges financial support from the Catholic Medical Center Research Foundation in the 2023 program year. Any additional authors, their immediate families, and any research foundations with which they are affiliated have not received any financial payments or other benefits from any commercial entity related to the subject of this article. Footnotes This retrospective multicenter cohort study was approved by the institutional review board of Yeouido St. Mary's Hospital, The Catholic University of Korea (IRB No. SC24WISI0061). Contributor Information Tae Hyung Kim, Email: katris@naver.com. Jong-Ho Kim, Email: taehyung@hongik.ac.kr. References