Referral Tendencies of ChatGPT in Oral and Maxillofacial Lesions: A Comparative Evaluation

Abstract

Objectives: This study aimed to evaluate the specialty referral patterns suggested by ChatGPT-4 in response to clinical scenarios involving oral and maxillofacial lesions. The primary objective was to assess the distribution of referrals among ear nose and throat (ENT), dermatology, plastic surgery, oral and maxillofacial surgery, oral diseases and radiology. Methods: A qualitative content analysis and expert review were conducted using the GPT-4 model. Twenty standardized clinical vignettes were created to represent a wide range of oral and maxillofacial pathologies. Each vignette was input into GPT-4 using a uniform triage-style prompt, asking the model to suggest the most appropriate medical specialty for referral (ENT, dermatology, plastic surgery, oral diagnosis and radiology, oral and maxillofacial surgery, or other) along with a brief justification. GPT-4’s referral outputs were documented and categorized. Results: Out of 20 clinical vignettes involving oral and maxillofacial lesions, ChatGPT-4’s referral decisions were rated by an oral and maxillofacial radiology expert. A total of 47 out of 60 points were awarded based on appropriateness, reflecting a 78.3% overall accuracy. The model most frequently referred cases to oral and maxillofacial surgery (60%), followed by dermatology (30%), oral diseases and radiology (25%), ENT (20%), and plastic surgery (10%). While GPT-4 performed well in common benign, cystic, and salivary lesions, limitations were noted in syndromic, infectious, and metastatic cases where co-referral or interdisciplinary awareness was lacking.Conclusions: This study investigates ChatGPT-4’s potential utility and limitations as a triage assistant in identifying appropriate specialty referrals for oral and maxillofacial lesions. While the model shows promise in pattern recognition, differences between AI-generated referrals and expert consensus indicate the need for interdisciplinary oversight and further refinement before clinical implementation. Integrating AI models into referral systems may support clinical workflows but should be approached with caution to ensure patient safety and specialty-appropriate care pathways.

Ähnliche Arbeiten