Large-Scale Large Language Model (LLM)-Assisted Report Mining for Elbow Tendon Co-occurrence Epidemiology: Prevalence, Association, and Validation.

introductionDistal biceps tendon pathology and common extensor (lateral epicondylar) tendon pathology are frequent causes of chronic elbow pain and may arise from overlapping degenerative or overuse mechanisms. However, the frequency with which these two entities coexist on elbow MRI has not been well characterized in a large imaging cohort. This study aimed to determine the prevalence of and association between distal biceps pathology and common extensor pathology on elbow MRI using a large language model (LLM)-assisted radiology report extraction pipeline and to evaluate the pipeline's reliability against independent human review.

methodsIn this retrospective cross-sectional observational study, 51,169 upper-extremity MRI reports were queried from an institutional dataset from 2025 to 2026. A broad, non-pathology-filtered elbow MRI cohort was identified using procedure metadata, yielding 10,222 examinations. Reports were classified for distal biceps pathology and common extensor/lateral epicondylar pathology using a locally deployed LLM (Qwen2.5 7B Instruct via Ollama; Alibaba Cloud (Qwen Team, Alibaba Group), Hangzhou, China) with a structured, negation-aware prompt and constrained JavaScript Object Notation (JSON) output. Prevalence was estimated with Wilson confidence intervals (CIs), the association was assessed with the chi-square (χ²) test and summarized as a prevalence ratio, and effect magnitude was evaluated with Cramér's V (φ). To evaluate extraction reliability, 400 reports were randomly sampled for blinded independent human review, with results compared against LLM classifications to estimate agreement.

resultsOf 10,222 reports, 10,086 (98.7%) were successfully classified by the model. Distal biceps pathology was present in 1,301 examinations (12.7%; 95% CI: 12.1-13.4%), and common extensor pathology was present in 3,818 examinations (37.4%; 95% CI: 36.4-38.3%), with both present concurrently in 696 examinations. Distal biceps pathology was significantly associated with a higher prevalence of concomitant common extensor pathology (prevalence ratio: 1.53; χ² = 165.3, p < 0.001), though the association was modest (φ = 0.127). Human validation of 400 reports demonstrated substantial agreement for both variables (distal biceps: accuracy 84.2%, κ = 0.685; common extensor: accuracy 85.2%, κ = 0.705). Classification errors were balanced for distal biceps pathology, but the model systematically over-classified common extensor pathology (sensitivity 96.1%, specificity 78.5%, McNemar p < 0.001), indicating that the reported common extensor prevalence is likely a modest overestimate.

conclusionDistal biceps pathology on elbow MRI is associated with a modestly increased prevalence of concomitant common extensor pathology. However, because the extraction pipeline systematically over-classified common extensor pathology relative to human review, this association is more likely overestimated than underestimated; a shared degenerative or biomechanical mechanism therefore remains only one plausible explanation. Large-scale LLM-assisted report extraction is a feasible adjunct to manual review for musculoskeletal imaging epidemiology, though bias-adjusted, externally validated estimates are necessary before these findings are considered definitive.