Study-Level Data and Reproducible R Code for Artificial Intelligence in Dental Caries Detection across Clinical Imaging Modalities: A Systematic Review and Descriptive Synthesis

Abstract

This repository contains the study-level data, executable R code, methodological audit files, supplementary documentation, and derived descriptive outputs supporting the systematic review entitled “Diagnostic Accuracy of Artificial Intelligence for Dental Caries Detection across Clinical Imaging Modalities: A Systematic Review and Descriptive Synthesis.” The literature search was updated through 10 June 2026. The review included 29 reports representing 28 unique studies. Twelve standalone-AI reports supplied exact, internally coherent 2 × 2 data for clinically interpretable observational units. These reports are presented as a descriptive availability subset rather than pooled across non-exchangeable imaging modalities, lesion thresholds, observational units, reference standards, and validation designs. No cross-modality pooled operating point, bivariate meta-analysis, HSROC curve, prediction region, pooled likelihood ratio, diagnostic odds ratio, or prevalence-dependent predictive-value analysis is produced. Controlled clinician-plus-AI evidence from Devlin et al. and Mertens et al. is summarized separately. Risk of bias was assessed using QUADAS-3 at the selected-estimate level, and certainty was evaluated using a structured GRADE-DTA framework. Overall risk of bias was high for all 28 assessed estimates, and certainty was rated very low for both standalone-AI diagnostic accuracy and incremental clinician performance with AI assistance. The repository includes study characteristics, estimate-selection decisions, exact contingency data, study-level sensitivity and specificity, controlled reader evidence, PRISMA accounting, PRISMA-DTA reporting data, protocol amendments, QUADAS-3 assessments, GRADE-DTA judgments, author-reported limitation statements retained as an ancillary transparency corpus, descriptive figures and tables, consistency checks, and R session information. The search workflow identified 137,274 raw database records. Of these, 130,245 were marked ineligible through Rayyan-assisted deterministic preprocessing before duplicate human screening. The retained materials do not contain the complete bulk-excluded record set, rule-specific counts, or a human-screened validation sample. Consequently, the false-negative rate of this preprocessing step cannot be estimated or retrospectively reconstructed. All files contain secondary study-level information extracted or derived from published reports. No individual participant data, identifiable clinical information, dental images, or copyrighted full-text articles are included. The review protocol was prospectively registered in PROSPERO (CRD420251232014).