Are Artificial Intelligence Models Reliable for Clinical Application in Pediatric Fracture Detection on Radiographs?...

Area of Science:

• Radiology
• Pediatric Orthopedics
• Artificial Intelligence

Background:

• AI shows promise for pediatric fracture diagnosis from radiographs.
• Existing studies have limitations like small sample sizes and inconsistent external validation.
• A meta-analysis is needed to provide robust estimates for clinical application.

Purpose of the Study:

• To determine the pooled diagnostic performance (sensitivity, specificity, AUC) of AI models for pediatric fractures.
• To assess the clinical applicability of AI models using external validation data.
• To investigate the impact of anatomic coverage on AI model performance.

Main Methods:

• A systematic review and meta-analysis following PRISMA 2020 guidelines.
• Inclusion of 16 diagnostic accuracy studies with 10,203 pediatric patients.
• Risk of bias assessed with QUADAS-2 and certainty of evidence with GRADE.

Main Results:

• Pooled sensitivity: 93%, specificity: 91%, AUC: 0.96.
• AI models maintained high accuracy on external datasets (sensitivity: 93%, specificity: 88%, AUC: 0.95).
• Broader anatomic coverage improved sensitivity, while single-region focus improved specificity.

Conclusions:

• AI models accurately detect pediatric fractures on radiographs, validated by external testing.
• Validated AI algorithms can be considered by clinicians to improve diagnostic accuracy in acute care.
• Future research should focus on subgroup performance and prioritize external validation for AI generalization.