Artificial Intelligence in Prosthodontic, Restorative and Implant Dentistry: A Reporting-Quality Scoping Review

Abstract

To map the reporting completeness and risk of bias of primary artificial intelligence (AI) studies in prosthodontics, restorative dentistry and implant dentistry, and to identify recurring methodological deficiencies. A scoping review was conducted following the Arksey and O’Malley framework with Levac and Joanna Briggs Institute refinements, reported in accordance with PRISMA-ScR. Primary studies of AI models for prosthodontic, restorative or implant tasks were charted for dataset provenance, external validation, code and data availability, reference-standard definition, sample-size justification and handling of class imbalance. Each study was routed to a design-appropriate instrument: reporting completeness was scored against a condensed twelve-domain adaptation of the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) for imaging and generative studies and against TRIPOD+AI for prediction models, while risk of bias was judged with QUADAS-2 or PROBAST+AI as appropriate.

Nineteen studies (2016–2024) were charted: six restorative, eight prosthodontic, five implant. Fifteen used convolutional neural networks, three generative adversarial networks and one a knowledge-based ontology. All nineteen used private datasets and none provided a code- or data-availability statement. Eighteen of nineteen (95%) were single-centre, reported no external validation, offered no sample-size justification and declared no adherence to any reporting guideline; 17/19 (89%) did not address class imbalance; and 15/19 (79%) defined the reference standard by expert annotation without reported inter-examiner reliability. Under QUADAS-2, 5/13 studies were at high and 7/13 at unclear overall risk of bias; under PROBAST+AI, 2/3 were at high risk, driven by the analysis domain.