Assessing the clinical deployability of AI tools for caries detection: A structured narrative synthesis and proposed...

To synthesise how studies evaluating AI-based caries detection on bitewing radiographs report evidence relevant to real-world clinical use beyond discrimination metrics (e.g., sensitivity, specificity, and area under the curve [AUC]), and to propose a structured appraisal framework for assessing clinical deployability.

A structured narrative synthesis of peer-reviewed studies evaluating AI-based caries detection on bitewing radiographs was conducted. Reporting was mapped to seven deployability domains (population fit/prevalence, calibration/thresholding, predictive values, validation integrity, reference standards, robustness, and workflows/governance), and studies were grouped into descriptive evidence profiles (Types A-C).

Thirty-nine studies met the inclusion criteria. Most were retrospective single-dataset evaluations. Type A evidence profiles predominated (33 studies), characterised by reporting of discrimination but limited information across several clinical-use domains. Type B profiles were less common (6 studies) and showed stronger validation practices (including patient-level separation and/or external validation) but persistent gaps in calibration and prevalence-sensitive reporting. No study met the criteria for Type C (documented regulatory clearance with deployment-level validation), limiting independent appraisal of commercial readiness from published sources.

The current evidence base emphasises discrimination metrics but frequently omits key elements required for safe, interpretable, and governable clinical use. Appraisal should incorporate prevalence context, calibration and thresholds, validation integrity, reference standards, robustness and workflow/governance. The seven-pillar framework provides a pragmatic lens for clinicians, procurement officers, and regulators appraising marketed AI tools.