Explainability, Bias and Generalizability of AI Models in Dentistry: A Systematic Review of Model Interpretability and Equity

Resumo original em inglês

O que aconteceu

AI-based dentistry has advanced significantly in recent years. AI models like deep learning (DL) and machine learning (ML) have paved the way for new approaches to image diagnostics and early risk prediction, making patient treatment plans more personalized.

The objective of this study was to assess the explainability, bias, and generalizability of AI models used in dentistry and evaluate the correlation between AI models.

Four databases were searched to retrieve relevant research records. The protocol was registered with PROSPERO. The data extraction sheet was designed according to PRISMA guidelines, and the data were managed in MS Excel. Also, a correlation analysis was performed to determine the nature of the relationship between the variables using SPSS. All tests were performed at a 95% confidence interval. Additionally, a critical appraisal of the included studies was also performed using the PROBAST tool.

O que encontraram

Eleven studies were included in this review. Overall, the assessment indicated variability in correlation strength between AI model accuracy and attributes of trustworthiness (r = 0.367-0.987). Analysis demonstrated the good performance of DL models (3D U-Net; accuracy = 95.10%) relative to others (73%-98.20%). However, the heterogeneous nature of included studies (n = 11) focused on different dental domains like diagnosis, dental service use, and disease risk prediction, which limits its generalizability.

Por que isso importa

Findings from this review indicated the importance of methodological rigor while using AI models in dentistry. Results suggest that the incorporation of trustworthiness attributes can improve dental treatment planning and early disease diagnosis.

Resumo publicado sob licença CC BY. A tradução para o português ainda não chegou a este estudo.