Methodological quality of machine learning-based caries prognostic models: a scoping review

Objective: To evaluate the methodological quality of published machine learning (ML) models for caries prognosis, with focus on the reporting of key performance metrics: discrimination, calibration, and clinical utility.

Methods: We conducted a scoping review following the PRISMA-ScR guidelines. We searched PubMed for studies published between January 2000 and April 2026 that developed multivariable ML models for caries risk prediction (PROGRESS 3 type). We assessed each study against an eight-step framework covering problem selection, data quality, study design, model development, internal and external validation, performance metrics, reporting transparency, and deployment readiness.



Results: Twenty-six studies met the inclusion criteria. Most studies (92%) reported some measure of discrimination, but only 65% reported AUC values. Calibration was reported in 4 of 26 studies (15%). Clinical utility through decision curve analysis was reported in 1 study (4%). External validation was performed in 3 studies (12%). Only 2 studies (8%) reported a sample size calculation. Pre-registration was documented in 2 studies (8%). No study reported a model update or data drift plan. Only 1 study (Zhang et al., 2025) addressed most framework steps, including external validation, calibration, clinical utility, TRIPOD adherence, and a deployed web-based calculator.



Conclusions: The current ML literature on caries prognosis suffers from critical reporting gaps, particularly in calibration and clinical utility. Without these metrics, the added value of ML models over existing caries risk assessment tools (Cariogram, CAMBRA) remains undemonstrated. Adherence to TRIPOD+AI guidelines, external validation, and assessment of clinical impact should be standard requirements for future studies.