Arthroscopy - 2026-10-24 - Journal Article
Magnetic Resonance Imaging Findings and Demographics Predict Surgical Versus Nonoperative Management of Rotator Cuff Injuries: A Natural Language Processing Analysis.
Keeley TS, Halvorson RT, Niknam K, Ziegeler K, Feeley BT, Link T, Majumdar S, Lansdown DA
Topics
Key Takeaway
Full-thickness supraspinatus tears carry an OR of 3.25 for surgical management, and female sex independently reduces surgical odds by 24% (OR 0.76) across 3,555 rotator cuff patients analyzed via GPT-4 NLP of MRI reports.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study asked which MRI findings and demographic factors predict surgical versus nonoperative management of rotator cuff injuries using GPT-4 NLP analysis of MRI reports from 1996–2023. NLP achieved 93.9% sensitivity and 99.3% specificity against human validation on 50 reports. Full-thickness supraspinatus tears (OR 3.25) and subscapularis tears (OR 1.74) predicted surgery, while female sex reduced surgical odds (OR 0.76); race and insurance type showed no significant association.
Key Limitation
Validation of NLP accuracy was performed on only 50 of 3,555 reports, leaving the error profile across the full dataset—particularly for complex multi-pathology reports—incompletely characterized.
Original Abstract
PURPOSE
To identify pathologic and demographic predictors of treatment selection and timing in patients with rotator cuff injuries using natural language processing of magnetic resonance imaging (MRI) reports.
METHODS
Data were retrospectively reviewed in our institution's Information Commons database. Adult patients who underwent shoulder MRI within 30 days of rotator cuff injury diagnosis (1996-2023) were identified using International Classification of Diseases codes and categorized by treatment. A Generative Pre-Transformer-4 model analyzed MRI reports using a reference list of shoulder pathologies, with validation against a human reference for 50 reports. χ 2 tests evaluated differences in shoulder pathology across treatment groups. Logistic regressions identified predictors of surgery-overall, as initial treatment, and after an initial trial of conservative management-while controlling for age, sex, insurance, race, and other imaging findings.
RESULTS
A total of 3555 patients were identified, of whom 1139 underwent surgery. Natural language processing achieved 93.9% sensitivity and 99.3% specificity. Full-thickness supraspinatus tears predicted surgery overall (odds ratio [OR] 3.25; 95% confidence interval [CI], 2.54-4.17, P < .001) and after an initial trial of conservative management (OR 3.14; 95% CI, 2.23-4.41, P < .001). Full-thickness subscapularis tears also predicted surgery overall (OR 1.74; 95% CI, 1.14-2.66, P = .011). Female sex decreased the odds of surgery (OR 0.76; 95% CI 0.64-0.90, P = .001). No differences in surgical likelihood were observed by race (P = .272-.986) or between Medicare and private insurance (P = .735).
CONCLUSIONS
MRI findings and demographics were associated with treatment selection and timing for patients with rotator cuff injuries. Natural language processing analysis of MRI reports enabled the identification of pathologic treatment predictors.
LEVEL OF EVIDENCE
Level III, retrospective prognostic cohort study.