AI algorithms (predominantly CNN/deep learning) as index test vs Dental expert / visual-tactile clinical assessment (reference standard): Log diagnostic odds ratio
AI algorithms show strong pooled discriminative accuracy for dental caries (log DOR 4.37; DOR 70.9); ungraded.
What each review found
Include earlier review versions| Finding | Effect | Studies | GRADE certainty |
|---|---|---|---|
AI algorithms (predominantly CNN/deep learning) as index test vs Dental expert / visual-tactile clinical assessment (reference standard) Log diagnostic odds ratio · Patients undergoing dental caries detection on dental images (teeth/surfaces across imaging modalities) · Cross-sectional diagnostic accuracy (index test vs reference standard), Dental imaging: intraoral photographs, periapical, bitewing, panoramic radiographs, CBCT, near-infrared transillumination, Overall (all imaging modalities pooled) AI algorithms show strong pooled discriminative accuracy for dental caries (log DOR 4.37; DOR 70.9); ungraded. Examining the diagnostic accuracy of artificial intelligence for detecting dental caries across a range of… (2025) All outcomes | Log DOR 4.37 [3.16, 6.27] | 20 studies · 29423 participants | GRADE not reported |
Findings are shown side by side, never pooled across reviews.
The reviews
| Review | Searched up to | Included studies | AMSTAR 2 |
|---|---|---|---|
| Examining the diagnostic accuracy of artificial intelligence for detecting dental caries across a range of imaging modalities: An umbrella… 2025 All outcomes | October 15, 2024 | 20 | Not applicable |
Studies behind the findings
None of these reviews ties its included studies to this outcome in the data here. Cochrane reviews and reviews with headline outcomes only do not list their studies yet.
How reviews are matched to this question
Reviews word the same question differently. Four decision models (TypeSafe's Jev, Liquid AI's d1-ultra, SubQ Snap and Baseten's Elephant) judged each wording of the population, intervention, comparator and outcome against the others; two wordings count as the same when at least three of them agree and none clearly disagrees, and findings count as this question only when all four elements match. Timing, setting and subgroup never split a question: they are shown on each finding.