ShRed: a machine learning model developed to predict shoulder redislocation.

Abstract

This feasibility study aimed to develop and evaluate machine learning models to predict shoulder redislocation using joint-specific imaging characteristics, in addition to traditional demographic variables. A prospective dataset from a tertiary referral centre was analysed, including cartilage MRI thickness measurements. Six classification algorithms were compared using 10-fold stratified cross-validation as the primary evaluation method. Preprocessing involved one-hot encoding of categorical variables and median imputation for missing values. The dataset was stratified into training (80%) and testing (20%) subsets. A total of 42 patients (54.8% redislocation rate) were included. Random Forest demonstrated the highest cross-validated accuracy of 79.0% (± 16.1%), precision of 88.3%, recall of 75.0%, and AUC of 0.84, with a 95% confidence interval of 69.0% to 89.0%. Feature importance analysis identified years since first dislocation as the most influential predictor, followed by age at first dislocation and glenoid cartilage thickness. This feasibility study demonstrates that machine learning models can predict shoulder redislocation with moderate accuracy. External validation on larger, multicentre datasets is required.