Potential and Pitfalls of Multimodal Large Language Models in Cerebral Palsy Hip Surveillance: A Radiographic...

Area of Science:

• Orthopaedic Surgery
• Artificial Intelligence in Medicine
• Pediatric Radiology

Background:

• Cerebral palsy (CP) hip displacement necessitates ongoing monitoring, burdening caregivers.
• Multimodal Large Language Models (MLLMs) may bridge health literacy gaps but require validation for interpreting pediatric radiographs.

Purpose of the Study:

• To assess the effectiveness and safety of MLLMs in addressing caregiver questions about CP hip management.
• To evaluate MLLM performance in interpreting pediatric pelvic radiographs for hip displacement.

Main Methods:

• Three MLLMs (GPT-4o, Claude 3.5, Gemini 1.5 Pro) analyzed 15 pediatric pelvic radiographs.
• Standardized caregiver prompts simulated clinical queries, generating 95 responses per model.
• Evaluated response length, interactivity, disclaimers, and diagnostic accuracy.

Main Results:

• GPT-4o showed highest safety (96.9% disclaimers); Claude 3.5 had shorter responses and diagnostic errors.
• MLLMs provided generic, not patient-specific, management advice and lacked quantitative analysis.
• All models correctly advised consulting an orthopaedic surgeon.

Conclusions:

• MLLMs can improve health literacy for CP hip displacement by offering summaries and stressing professional consultation.
• Radiographic "hallucinations" and lack of specific guidance prevent MLLM use as standalone diagnostic tools.
• MLLMs currently serve as educational adjuncts needing expert oversight in pediatric orthopaedics.