Data sources: The Cochrane CENTRAL and Medline databases and reference lists of identified were searched.
Study selection: Prospective longitudinal cohorts or randomised controlled trials were included.
Data extraction and synthesis: Study assessment and data extraction were carried out independently by at least two reviewers. The quality of studies was assessed using the QUADAS and AMSTAR tools. The overall quality of evidence was assessed using the GRADE approach.
Results: 90 studies were included, seven were of high quality, 35 of moderate quality and the rest poor. The accuracy of multivariate models was higher for pre-school children than for schoolchildren/adolescents. As the models had rarely been tested in independent populations their accuracy is uncertain. The single predictor baseline caries experience had moderate/good accuracy in pre-school children and limited accuracy in schoolchildren/adolescents. In general, the quality of evidence was limited.
Conclusions: Multivariate models and baseline caries prevalence performed better in pre-school children than in school-children/adolescents. Baseline caries prevalence was the most accurate single predictor in all age groups. The heterogeneity of populations, models, outcome criteria, measures and reporting hampered the synthesis of results. There is a great need to standardise study design, outcome measures and reporting of data in studies on caries risk assessment. The accuracy of prediction models should be validated in at least one independent population.