From remote screening to precision prevention: responsible multimodal AI for risk prediction and equitable oral healthcare

Artificial intelligence (AI) applications in oral healthcare have expanded considerably over the past decade. Deep-learning systems now report diagnostic performance exceeding 0.85 sensitivity and 0.90 specificity for several image-based tasks, with the highest pooled estimates reported in AI-assisted clinical photography for oral cancer and oral potentially malignant disorder (OPMD) detection (diagnostic odds ratio 68.4; AUC 0.938). These figures, however, mask three translational gaps. First, many studies often carry high risk of bias, lack external validation, and rest on heterogeneous reference standards. Second, datasets often provide limited demographic reporting, leaving uncertainty about model performance in the populations most likely to benefit from remote screening. Third, teledentistry and mHealth tools have outpaced the regulatory, validation, and fairness-auditing, and clinical-integration frameworks required for safe deployment. Preventive value emerges most clearly when AI is embedded in multimodal systems, such as imaging, sensors, behavioural feedback, clinician support rather than evaluated as an isolated classifier. Progress will be defined less by additional accuracy gains and more by external validation in diverse populations, transparent demographic reporting, equity-focused evaluation, and integration into prevention-oriented care pathways.