NUMERO: #20260199

Título: APPLICATION OF GENERAL-PURPOSE ARTIFICIAL INTELLIGENCE SYSTEMS IN STOMATOLOGY: A DIAGNOSTIC ACCURACY PILOT STUDY

Nome do Apresentador: Gabriel Silva Rezende FREITAS

Categoria do Trabalho: Pôster - Investigação Original (IO)

Área Temática: Estomatologia

Resumo: Objective: To evaluate the diagnostic accuracy of general-purpose AI systems (ChatGPT and Gemini) to identify oral lesions considering different levels of clinical description detail, compared to reference diagnoses.

Study design: This observational, cross-sectional diagnostic accuracy study included 50 previously diagnosed clinical cases. Standardized descriptions at two levels of detail (detailed vs. simplified), combined with clinical images, were submitted to ChatGPT and Gemini using independent accounts to eliminate memory bias. AI-generated diagnoses were categorized as primary, differential, or incorrect, and compared against a reference standard (histopathology or expert consensus).

Results: In the detailed scenario, global accuracy (Top-3 accuracy) was 86% for ChatGPT and 94% for Gemini (Kappa = 0.445). Gemini correctly identified the primary diagnosis in 78% of cases, compared to 68% for ChatGPT. With simplified descriptions, global accuracy dropped to 78% for ChatGPT and 88% for Gemini (Kappa = 0.31), with primary diagnosis success falling to 56% and 68%, respectively. Performance was highest for malignant and distinctive lesions, while reactive and cystic cases posed greater diagnostic challenges.

Conclusion: General-purpose AI systems demonstrated high diagnostic accuracy, with Gemini consistently outperforming ChatGPT. Reduced clinical detail significantly impaired performance, highlighting the necessity of high-quality information for reliable AI-assisted clinical decision-making in stomatology.

Autor 1:  Gabriel Silva Rezende FREITAS

E-mail 1:  [email protected]

Autor 2:   Wander Tamura

E-mail 2:  [email protected]

Autor 3 :  Carlos Deyver de Souza Queiroz

E-mail 3:  [email protected]

Autor 4:  Cássio Vinhadelli Ribeiro

E-mail 4:  [email protected]

Autor 5:  William Eduardo Pirola

E-mail 5:  [email protected]