Table 5.

Exploratory analysis of model-generated diagnostic error patterns.

Case info	Ground truth	AI diagnosis	Observed diagnostic error pattern
Case 11: 47 years old, female, right humerus, no trauma	Simple bone cyst with hemorrhage (benign)	All Groups: bone metastasis, osteosarcoma	Demographic Anchoring Bias: The AI overrelied on the statistical prior that “middle-aged patients with bone lesions have a high probability of metastasis” and did not appropriately account for the relatively benign morphological margins visible in the images.
Case 27: 15 years old, male, right femur, trauma (1 day)	Osteosarcoma (malignant)	Base/RAGa-Radiologist: Salter-Harris II fracture, epiphyseal separation Base/RAG-Ortho Oncologist: correctly identified osteosarcoma with pathological fracture	Premature Closure via Trauma Masking: The radiologist persona prematurely anchored its diagnosis on the mechanical trauma history, misinterpreting the malignant lesion as a benign sports injury. Conversely, the orthopedic persona successfully resisted this text bias, demonstrating the clinical resilience of specialized prompts.
Case 3: 6 years old, female, left tibia, no trauma	Aneurysmal bone cyst	Base Model: bone cyst, non-ossifying fibroma (benign cluster) RAG Model: Ewing sarcoma, acute osteomyelitis	RAG-Associated Diagnostic Shift: While the base model correctly confined its differential to benign entities based on visual features, the RAG integration retrieved aggressive pediatric textbook priors (Ewing sarcoma), overriding visual intuition and coinciding with a marked shift toward malignant diagnoses.

^aRAG: retrieval-augmented generation.