Abstract

Objectives: To evaluate the efficacy of Chat Generative Pre-Trained Transformer (ChatGPT) in generating antimicrobial management recommendations for pediatric patients with contaminated traumatic wounds, particularly in the absence of infectious disease (ID) consultation.

Methods: Three pediatric cases involving severely contaminated traumatic injuries were retrospectively presented to ChatGPT-4, including clinical data such as injury mechanism, cultures, susceptibilities, and presence of osteomyelitis. The model’s antibiotic recommendations were compared with those made by the pediatric ID team. A pediatric ID specialist graded each recommendation based on appropriateness, pathogen coverage, and clinical judgment.

Results: ChatGPT provided 29 antibiotic recommendations: 3 for initial antibiotic selection, 8 for pathogen management, and 18 for follow-up changes. All of the initial antibiotic recommendations were deemed appropriate or identical to ID guidance. Of the pathogen-specific recommendations, 87.5% were appropriate and 12.5% were unnecessary. Among the follow-up changes, 11.1% were appropriate, 77.8% were unnecessary, and 11.1% were incorrect. In total, 41.4% of recommendations were appropriate or identical, 51.7% were unnecessary, and 6.9% were incorrect. The majority of unnecessary changes involved redundant adjustments to existing effective therapies.

Conclusions: ChatGPT performed well in initial antibiotic selection and pathogen-directed management but frequently proposed unnecessary follow-up modifications. Although promising as a supplemental tool in resource-limited settings, its current limitations highlight the need for further refinement before broader clinical application.