NUMERO: #20260423

Título: Predicting Malignant Transformation in Oral Epithelial Dysplasia: AI-Based Models vs Pathologist Assessment

Nome do Apresentador: Diele Carine Barreto ARANTES

Categoria do Trabalho: Pôster - Investigação Original (IO)

Área Temática: Patologia Oral

Resumo: Objective:
To systematically review and meta-analyze studies applying artificial intelligence (AI)-based risk prediction models (RPMs) compared with conventional human microscopy for predicting malignant transformation (MT) of oral epithelial dysplasia (OED).
Study design:
Systematic review and meta-analysis conducted in accordance with PRISMA-DTA and Cochrane guidelines. Studies evaluating AI-based approaches versus conventional H&E microscopy for OED, including leukoplakia, leukoerythroplakia, and erythroplakia-associated lesions, were eligible. The protocol was registered in PROSPERO (CRD42024608333).
Results:
Thirty-two studies (1998-2026) were included: 9 AI-based and 23 human-based, comprising 7,546 cases (AI: 1,813; human: 5,733), predominantly from Europe. Meta-analysis showed strong discriminative performance for AI models (pooled OR = 12.64; 95% CI). Algorithms and input data were heterogeneous, with multilayer perceptron and random forest being the most frequently used models. Human evaluation confirmed the association between dysplasia and MT across both two-tier (OR = 3.33; 95% CI) and three-tier systems (OR = 3.55; 95% CI), with no significant differences between subgroups (p = 0.905). However, substantial heterogeneity was observed (I² = 85%).
Conclusion:
AI-based RPMs demonstrate promising and consistent performance; however, clinical implementation requires prospective, multicenter, and geographically diverse studies with harmonized workflows and rigorous external validation.
Funding: FAPESP 2025/08148-4.

Autor 1:  Diele Carine Barreto ARANTES

E-mail 1:  [email protected]

Autor 2:   Lucas Monteiro Delgado

E-mail 2:  [email protected]

Autor 3 :  Marília de Marco Pinto Paiva dos Santos

E-mail 3:  [email protected]

Autor 4:  Saman Warnakulasuriya

E-mail 4:  [email protected]

Autor 5:  Marcelo Sperandio

E-mail 5:  [email protected]