JSES - 2026-07-03 - Journal Article
Prediction of Acromial and Scapular Spine Fractures After Reverse Total Shoulder Arthroplasty using Machine Learning: A Retrospective Cohort Study.
Schneller T, Cina A, Maggini E, Klimov A, Braun M, Pfender A, Lazaridou A, Scheibel M
Topics
Key Takeaway
A logistic regression ML model predicts acromial and scapular spine fractures after rTSA with AUROC 0.71, identifying medialized-distalized implant design, cuff tear arthropathy, older age, and osteoporosis as the strongest preoperative risk factors in a cohort with 4.1% overall ASSF incidence.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study used a registry of 2,256 primary rTSA patients (2006–2025) to determine ASSF incidence and develop a logistic regression ML model using 10 preoperative features including implant design. Overall ASSF incidence was 4.1%, with Levy type 2 fractures predominating (51.6%); medialized-distalized designs were associated with a higher proportion of Levy type 3 fractures. The final model achieved sensitivity 0.71, specificity 0.61, and AUROC 0.71 across all implant design subgroups.
Key Limitation
Model performance is only fair (AUROC 0.71) and was developed and validated within a single institutional registry, limiting generalizability across implant systems, surgical techniques, and patient populations not represented in this cohort.
Original Abstract
BACKGROUND
Acromial and scapular spine fractures (ASSF) are an uncommon but significant complication following reverse total shoulder arthroplasty (rTSA). The study aims were to (1) determine the incidence and type of ASSF captured within a local shoulder arthroplasty registry, (2) develop a novel, interpretable machine learning model to predict individual risk of fracture after rTSA and (3) assess model performance variation based on implant design.
METHODS
We conducted a retrospective cohort study including 2,256 registry-documented patients who underwent primary rTSA between 2006 and 2025. Ten preoperative features, including implant design (medialized-distalized (MD), lateralized (L), and lateralized-distalized (LD)) were used as input to a logistic regression machine learning model aimed at predicting ASSF risk. Model performance was evaluated in the overall cohort and different implant design subgroups based on sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC).
RESULTS
Overall ASSF incidence was 4.1%, whereby 15% were Levy type 1, 51.6% type 2 and 33.4% type 3 fractures. A greater incidence of Levy type 2 fractures was observed for L implant designs, whereas a higher proportion of Levy type 3 fractures occurred with MD designs. Design MD, cuff tear arthropathy, higher age and osteoporosis emerged as the strongest predictors. In the test set, the final model had a sensitivity of 0.71, specificity of 0.61 and AUROC of 0.71. All implant design subgroups showed fair predictive performance.
CONCLUSION
Our model demonstrated promising capability for preoperative identification of patients at risk of ASSF following rTSA. Implant design MD, cuff tear arthropathy, higher age and osteoporosis were the strongest predictors of ASSF. Overall, these findings support the integration of personalized risk stratification to guide implant selection and postoperative surveillance.
LEVEL OF EVIDENCE
Basic Science Study; Development of Computer Prediction Model with Machine Learning.