Prediction of postoperative persistent pain after rotator cuff repair based on psychological measurements via machine learning

Background With the development of artificial intelligence technique, machine learning (ML) algorithms have been widely applied to forecast outcomes of various diseases. Current study aimed to investigate the potential utility of machine learning algorithms for predicting postoperative persistent shoulder pain (PSP) and to investigate the related risk factors following rotator cuff repair surgery. Materials and methods A cohort of 491 patients with rotator cuff tears who had undergone surgical intervention was enrolled and analyzed. Visual Analog Scale (VAS) score of ≥4 at one year follow-up was defined as indicative of ongoing shoulder pain. Consequently, patients were classified into a residual pain group and a pain-free group. Nine commonly used machine learning models were trained to distinguish between patients with residual pain and those without. Furthermore, the three models that exhibited the highest predictive performance were combined into a single ensemble of the nine models to improve the classification of patients with residual pain versus those without. The area under the curve (AUC), specificity, sensitivity and accuracy were calculated for each model to evaluate their predictive capabilities. A prespecified LASSO-based feature-selection strategy was used to identify candidate clinical variables, with additional stability analyses (forward selection, exhaustive search) reported in the Supplementary Materials. Variables consistently selected were considered candidate risk factors for PSP in patients. Results (1) Machine learning models demonstrated their effectiveness in classifying patients with postoperative pain. The predictive performance was notably enhanced using an ensemble learning technique. Among the models, the Support Vector Machine (SVM) exhibited the highest accuracy when serving as the aggregator. (2) The ensemble learning models identified several critical predictive factors, including the Fear of Pain Questionnaire, the Pain Catastrophizing Scale, the Pittsburgh Sleep Quality Index (PSQI)-Perceived Sleep Quality, the Central Sensitization Inventory, and the Pain Surface Body Chart score. Conclusion Our findings suggest that a stacking ensemble of preoperative psychosocial variables, developed and internally validated in this single-center retrospective cohort, shows promise for preliminary risk stratification of clinically significant shoulder pain at one year after rotator cuff repair. These hypothesis-generating results require multicenter external validation before clinical use. Level of evidence Level 4, Diagnosis. Clinical trial number Not applicable. This was a retrospective observational study and was not registered as a clinical trial.