A Systematic Review of Bias, Validation, and Explainability in Artificial Intelligence Algorithms for Dental Diagnostics.
Authors
Affiliations (3)
Affiliations (3)
- Centre of Oral Bioengineering, Queen Mary University of London, London, UK.
- School of Electronic Engineering and Computer Science, Queen Mary University of London, London, UK.
- Centre of Oral Bioengineering, Queen Mary University of London, London, UK. Electronic address: [email protected].
Abstract
To critically evaluate the current role of artificial intelligence (AI) in dental diagnosis across multiple dental disciplines, assessing diagnostic performance, methodological quality, risk of bias, validation, explainability, and clinical applicability. This systematic review was conducted in accordance with PRISMA guidelines and registered in PROSPERO (CRD42025639340). PubMed/MEDLINE, Scopus, Embase, LILACS, Cochrane Central Register of Controlled Trials, Web of Science, and ClinicalTrials.gov were searched up to December-2025. Two reviewers independently performed study selection, data extraction, and risk-of-bias assessment. Risk of bias in diagnostic accuracy studies was assessed using QUADAS-2-tool, evaluating the domains of patient selection, index test, reference standard, flow and timing whilst randomised studies were assessed using Cochrane RoB 2-tool. Reporting quality and transparency of AI-model development and validation were assessed using the selected items from TRIPOD-AI reporting guideline. Methodological quality and risk of bias in AI model development and evaluation were additionally assessed using PROBAST+AI. Ninety-six studies met the inclusion criteria, of which 90.6% were retrospective. Convolutional neural networks (CNN) were the most frequently used architecture (82.3%). Sixteen studies failed to specify the CNN architecture used. Reported diagnostic accuracy ranged from 81% to 99% across dental specialties, with high performance in implantology (92-99%), orthodontics (88-99%), endodontics (87-99%). Panoramic radiographs and cone-beam computed tomography were the most used imaging modalities. QUADAS-2 analysis indicated that high-risk of bias in 55.2% of studies and most lacked external validation, whilst Assessment using selected TRIPOD+AI items identified important gaps in reporting completeness and transparency, particularly in sample-size justification, missing-data handling, protocol availability, study registration, and data or code availability. AI demonstrates substantial technical capability to enhance diagnostic accuracy and efficiency. However, the predominance of retrospective designs, limited external validation, and methodological weaknesses constrain the strength of current evidence. The use of AI is potentially considered as a decision-support tool rather than replacement for clinical judgement.