Machine learning to predict the occurrence of complications after total shoulder arthroplasty for B2-B3 glenoids.

Background: Total shoulder arthroplasty (TSA) for primary glenohumeral osteoarthritis with B2-B3 glenoids is challenging due to the relatively high rate of postoperative complications, such as glenoid implant loosening. Machine learning (ML) is a promising method for predicting outcomes in shoulder arthroplasty. However, no studies have included preoperative radiological data to predict surgical complications using ML. The present study evaluated the potential of ML in predicting the occurrence of complications after TSA in patients treated for glenohumeral osteoarthritis with B2-B3 glenoids by integrating various prognostic factors, such as radiological features. We hypothesized that ML would accurately predict postoperative complications and identify the variables that are most strongly associated with these complications. Materials and methods: This retrospective study included 60 patients with primary osteoarthritis and type B2-B3 glenoids from our institutional TSA database. Prognostic factors, including patient characteristics, clinical scores, radiological features, and surgical techniques, were recorded. Outcomes at a minimum of 2 years of follow-up were characterized by the Aldinger complication scale (scored 0-III). Of the 60 patients, 13 (21.7%) experienced complications, with 8 (13.3%) classified as Aldinger I and 5 (8.3%) as Aldinger III. These data were used to train and test four ML methods: logistic regression (LR), gradient boosting classifier (GBC), support vector machine (SVM), and multilayer perceptron classifier (MLPC). We considered a binary outcome: no complication vs. Aldinger I-III. The data were split into a training set (75%) and a testing set (25%). Results: Among the four ML models evaluated, LR and GBC correctly identified all complication cases (3/12), whereas SVM and MLPC missed one complication. The number of false positives was lower with GBC (2/12) and LR (3/12). Younger age, glenoid version and inclination were the main variables associated with complications. Using a posteriorly augmented glenoid implant was associated with lower complication rates. Conclusion: ML can efficiently predict TSA complications, even with a limited dataset. Glenoid retroversion was identified as a critical radiological feature associated with outcomes, as supported by the literature. In addition, younger age is associated with increased complication risks, likely due to increased functional demand. Thus, ML is potentially a valuable tool for forecasting complications in the surgical decision-making process.