Purpose or Case Report: To assess the epidemiological distribution, anatomical patterns, and classification trends of pediatric wrist fractures over 15 years, with emphasis on age- and sex-specific variations, laterality, and fracture co-occurrences and to demonstrate the utility of large language models (LLMs) for large-scale, automated extraction and analysis of radiologic data.
Methods & Materials: This retrospective study analyzed pediatric wrist radiographs from 24,592 patients (2010–2025) within our institutional archive (mean age 10.62±3.86 years, 61.8% male). Fracture characteristics were extracted from radiology reports using an open-source medial large language model (LLM), Google MedGemma 27B, prompted with wrist fracture terminology. Extracted features included anatomical region, fracture type, and Salter–Harris (SH) classification along with age and sex. Patients were stratified into six age groups (0–2, 3–5, 6–8, 9–11, 12–14, and 15–18 years) Chi-square tests assessed associations between demographic and radiologic features (p<0.05 considered significant).
Results: Fractures were present in 77.4% of patients, averaging 1.76 fractures per patient. The distal radius (64.5%) was the most common site, followed by the distal ulna (21.4%), ulnar styloid (10.3%), and scaphoid (3.8%). Salter–Harris fractures comprised 22.4% of distal radius and 15.6% of distal ulna fractures. Buckle fractures predominated in younger children, particularly females, while transverse and Salter–Harris fractures were more frequent in older groups Fracture incidence peaked in the 12–14-year cohort (14.9%, p< 0.001). Bilateral fractures were uncommon overall but significantly more frequent among 0–2 years (p < 0.001). Co-occurrence analysis showed the strongest associations between distal radius and distal ulna fractures, followed by radius and ulnar styloid combinations.
Conclusions: This large-scale study delineates the spectrum and epidemiological trends of pediatric wrist fractures across age groups. The distal radius remains the most frequently involved site, with buckle fractures characterizing early childhood and Salter–Harris injuries peaking in adolescence, consistent with the literature. Leveraging large language models enabled automated, reliable analysis of over 24,000 radiographs, demonstrating their potential for scalable epidemiologic research and enhanced data-driven insights in pediatric radiology.