Artificial Intelligence in the Diagnosis of Oral Potentially Malignant Disorders: A Critical Narrative Review of Diagnostic Validity, Prognostic Value and Translational Readiness

Abstract

Oral potentially malignant disorders carry a variable but clinically meaningful risk of progression to oral squamous cell carcinoma, and their recognition depends on visual examination and on histological grading of epithelial dysplasia, both of which are subjective and unevenly available. Artificial intelligence has been proposed as a means of standardising detection, triaging lesions in settings without specialists, and estimating the risk of malignant transformation. This critical narrative review examines whether the published evidence supports these expectations. Peer-reviewed literature published between January 2000 and 27 July 2026 was identified through structured searches of biomedical and multidisciplinary scholarly indexes, supplemented by citation searching, and was appraised for study design, reference standard, validation strategy, calibration, reporting quality and clinical applicability. The synthesis is organised around five task families: photograph-based detection and classification, optical and cytological adjuncts augmented by machine learning, computational histopathology of oral epithelial dysplasia, multimodal risk prediction, and emerging large language and vision–language models. High discrimination is consistently reported for distinguishing lesions from normal mucosa, yet performance declines for clinically difficult distinctions, notably among white lesions, and falls further when images are acquired by untrained operators or originate from new centres. Evidence is dominated by retrospective, single-centre studies with curated images, histologically confirmed case mixes that differ from screening populations, and reference standards that inherit the poor reproducibility of dysplasia grading. The most methodologically mature work concerns whole-slide image analysis for predicting transformation, where externally validated models achieve discrimination comparable to, but not clearly better than, established grading systems. Pooled estimates from several meta-analyses are high but rest on heterogeneous and frequently optimistic primary data, and certainty is generally low. Prospective field validation, calibration reporting, decision-curve analysis and assessment of patient-relevant outcomes remain scarce. Artificial intelligence is best regarded at present as a promising decision-support adjunct for triage and research rather than a validated diagnostic replacement for clinical examination and biopsy. Progress will depend on multicentre prospective studies, harmonised reference standards, adherence to reporting guidelines for artificial intelligence and explicit attention to equity in the populations most affected.