BACKGROUND: Multivariable prediction models are used in oral health care to identify individuals with an increased likelihood of caries increment. The outcomes of the models should help to manage individualized interventions and to determine the periodicity of service. The objective was to review and critically appraise studies of multivariable prediction models of caries increment.

METHODS: Longitudinal studies that developed or validated prediction models of caries and expressed caries increment as a function of at least three predictors were included. PubMed, Cochrane Library, and Web of Science supplemented with reference lists of included studies were searched. Two reviewers independently extracted data using CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies) and assessed risk of bias and concern regarding applicability using PROBAST (Prediction model Risk Of Bias ASessment Tool). Predictors were analysed and model performance was recalculated as estimated positive (LR +) and negative likelihood ratios (LR -) based on sensitivity and specificity presented in the studies included.

RESULTS: Among the 765 reports identified, 21 studies providing 66 prediction models fulfilled the inclusion criteria. Over 150 candidate predictors were considered, and 31 predictors remained in studies of final developmental models: caries experience, mutans streptococci in saliva, fluoride supplements, and visible dental plaque being the most common predictors. Predictive performances varied, providing LR + and LR - ranges of 0.78-10.3 and 0.0-1.1, respectively. Only four models of coronal caries and one root caries model scored LR + values of at least 5. All studies were assessed as having high risk of bias, generally due to insufficient number of outcomes in relation to candidate predictors and considerable uncertainty regarding predictor thresholds and measurements. Concern regarding applicability was low overall.

CONCLUSIONS: The review calls attention to several methodological deficiencies and the significant heterogeneity observed across the studies ruled out meta-analyses. Flawed or distorted study estimates lead to uncertainty about the prediction, which limits the models' usefulness in clinical decision-making. The modest performance of most models implies that alternative predictors should be considered, such as bacteria with acid tolerant properties.

TRIAL REGISTRATION: PROSPERO CRD#152,467 April 28, 2020.

Supragingival dental plaque samples were collected from 40 Swedish adolescents, including 20 with caries lesions (CAR) and 20 caries-free (CF). Fresh plaque samples were subjected to an ex vivo acid tolerance (AT) test where the proportion of bacteria resistant to an acid shock was evaluated through confocal microscopy and live/dead staining, and the metabolites produced were quantified by 1H Nuclear Magnetic Resonance (1H NMR). In addition, DNA was extracted and the 16S rRNA gene was sequenced by Illumina sequencing, in order to characterize bacterial composition in the same samples. There were no significant differences in AT scores between CAR and CF individuals. However, 7 out of the 10 individuals with highest AT scores belonged to the CAR group. Regarding bacterial composition, Abiotrophia, Prevotella and Veillonella were found at significantly higher levels in CAR individuals (p=0.0085, 0.026 and 0.04 respectively) and Rothia and Corynebacterium at significantly higher levels in CF individuals (p=0.026 and 0.003). The caries pathogen Streptococcus mutans was found at low frequencies and was absent in 60% of CAR individuals. Random-forest predictive models indicate that at least 4 bacterial species or 9 genera are needed to distinguish CAR from CF adolescents. The metabolomic profile obtained by NMR showed a significant clustering of organic acids with specific bacteria in CAR and/or high AT individuals, being Scardovia wiggsiae the species with strongest associations. A significant clustering of ethanol and isopropanol with health-associated bacteria such as Rothia or Corynebacterium was also found. Accordingly, several relationships involving these compounds like the Ethanol : Lactate or Succinate : Lactate ratios were significantly associated to acid tolerance and could be of predictive value for caries risk. We therefore propose that future caries risk studies would benefit from considering not only the use of multiple organisms as potential microbial biomarkers, but also their functional adaptation and metabolic output.

INTRODUCTION: The identification of salivary molecules that can be associated to dental caries could provide insights about caries risk and offer valuable information to develop caries prediction models. However, the search for a universal caries biomarker has proven elusive due to the multifactorial nature of this oral disease. We have therefore performed a systematic effort to identify caries-associated metabolites and proteins in saliva samples from adolescents that had a caries experience and those that were caries-free.

METHODS: Quantification of approximately 100 molecules was performed by the use of a wide range of techniques, ranging from nuclear magnetic resonance metabolomics to ELISA, Luminex or colorimetric assays, as well as clinical features like plaque accumulation and gingival index. In addition, simplified dietary and oral hygiene habits questionnaires were also obtained.

RESULTS: The caries-free group had significantly lower consumption of sweetened beverages and higher tooth brushing frequency. Surprisingly, very few compounds were found to individually provide discriminatory power between caries-experienced and caries-free individuals. The data analysis revealed several potential reasons that could underly this lack of association value with caries, including differences in metabolite concentrations throughout the day, a lack of correlation between metabolite concentrations in plaque and saliva, or sex-related differences, among others. However, when multiple compounds were combined by multivariate analysis and random forest modeling, a combination of 3-5 compounds were found to provide good prediction models for morning (with an AUC accuracy of 0.87) and especially afternoon samples (AUC = 0.93).

CONCLUSION: While few salivary biomarkers could differentiate between caries-free and caries-experienced adolescents, a combination of markers proved effective, particularly in afternoon samples. To predict caries risk, these biomarkers should be validated in larger cohorts and longitudinal settings, considering factors such as gender differences, and variations in oral hygiene and diet.