Early Childhood Predictors for Dental Caries: A Machine Learning Approach.

Abstract

We aimed to develop and validate caries prognosis models in primary and permanent teeth after 2 and 10 y of follow-up through a machine learning (ML) approach, using predictors collected in early childhood. Data from a 10-y prospective cohort study conducted in southern Brazil were analyzed. Children aged 1 to 5 y were first examined in 2010 and reassessed in 2012 and 2020 regarding caries development. Dental caries was assessed using the Caries Detection and Assessment System (ICDAS) criteria. Demographic, socioeconomic, psychosocial, behavioral, and clinical factors were collected. ML algorithms decision tree, random forest, and extreme gradient boosting (XGBoost) were employed, along with logistic regression. The discrimination and calibration of models were verified in independent sets. From 639 children included at the baseline, we reassessed 467 (73.3%) and 428 (66.9%) children in 2012 and 2020, respectively. For all models, the area under receiver operating characteristic curve (AUC) at training and testing was above 0.70 for predicting caries in primary teeth after 2-y follow-up, with caries severity at the baseline being the strongest predictor. After 10 y, the SHAP algorithm based on XGBoost achieved an AUC higher than 0.70 in the testing set and indicated caries experience, nonuse of fluoridated toothpaste, parent education, higher frequency of sugar consumption, low frequency of visits to the relatives, and poor parents' perception of their children's oral health as top predictors for caries in permanent teeth. In conclusion, the implementation of ML shows potential for determining caries development in both primary and permanent teeth using easy-to-collect predictors in early childhood.