Artificial Intelligence Predicts Failure of Nonoperative Treatment in Degenerative Posterosuperior Rotator Cuff Tears

To develop an artificial intelligence-based algorithm to predict success of nonoperative treatment in degenerative posterosuperior tears and to identify the most influential features.

Patients with degenerative rotator cuff tears initially treated nonoperatively were enrolled prospectively from January 2022 to November 2024. Exclusion criteria included age >75 years; <24 months of follow-up; massive, traumatic, and partial tears; arthritic changes (Hamada >2); and severe fatty degeneration (Goutallier >2). All patients followed a standardized physiotherapy protocol and could choose surgery if symptoms persisted; failure was defined as opting for surgery. Collected variables included demographics, comorbidities (diabetes, dyslipidemia, hypertension, hypothyroidism, and smoking), magnetic resonance imaging tear characteristics (size; International Society of Arthroscopy, Knee Surgery, and Orthopaedic Sports Medicine classification; associated cuff involvement; and biceps pathology), corticosteroid injection history, and functional scores (American Shoulder and Elbow Surgeons score, Subjective Shoulder Value, and the visual analog scale). First, a standard statistical analysis between groups was performed. Later, the dataset was split into a train (80%) and test (20%), and a random forest algorithm model was developed. Finally, SHapley Additive exPlanations values were computed to assess feature importance.

Of the 412 patients enrolled, 347 patients met the inclusion criteria. After 24 months, 66 (19%) patients opted for surgery. Mean follow-up was 30.4 months (range, 24-32). Compared with the successful nonoperative treatment group, patients who failed treatment were younger (58.5 vs 64.6 years, P = .001) and had higher body mass index (28.8 vs 27.2, P = .016), greater subscapularis involvement (13.5% vs 3%, P = .042), higher visual analog scale scores (7.7 vs 7.2, P = .018), lower Subjective Shoulder Value (30.9 vs 38.9, P = .003), and lower American Shoulder and Elbow Surgeons scores (40.6 vs 47.1, P = .001). The model achieved an area under the curve of 84%, predicting success rate of the nonoperative treatment with 94% sensitivity, 62% specificity, 89% positive predictive value, and 77% negative predictive value. The most influential variables for the model were age, tear size, and pretreatment functional scores.

The artificial intelligence-based algorithm accurately predicted nonoperative treatment outcomes in degenerative posterosuperior rotator cuff tears (area under the curve 0.84, sensitivity 94%). Age, tear size, and function were the main predictors.

Level IV, prognostic retrospective case series.