Predicting Longitudinal Caries Trajectories from Childhood to Early Adulthood

Abstract

Prior studies have used traditional trajectory analyses to classify caries progression; however, none have applied machine learning (ML) to predict caries trajectories from childhood to early adulthood. The aims of our study are 1) to use unsupervised ML to perform trajectory analysis by clustering the longitudinal caries data into distinct trajectory groups and 2) to utilize supervised ML to predict trajectory group membership from behavioral/dietary, fluoride, and sociodemographic variables. This study was conducted using longitudinal data from 560 Iowa Fluoride Study participants. Trajectory analysis was first done via K-means for longitudinal data on caries data (D²⁺MFS counts) obtained at ages 9 y (n = 523), 13 y (n = 549), 17 y (n = 464), and 23 y (n = 342). The optimal number of trajectory groups was based on the Caliński-Harabasz criterion and clinical relevance. Supervised ML was then performed with trajectory group membership as the outcome variable against 11 predictor variables. The performance of 5 models was compared by Brier score and accuracy: 1) ordered multinomial logistic regression, 2) least absolute shrinkage and selection operation, 3) gradient boosting machine, 4) extreme gradient boosting, and 5) neural network. Of the 560 participants included in this study, 3 caries trajectory groups were identified: low (70.5%), medium (21.1%), and high (8.4%), characterized by minimal, moderate, and severe and progressive disease, respectively. Extreme gradient boosting outperformed the other 4 models, with 85.9% accuracy and a Brier score of 0.21. Top predictors included sex, socioeconomic status, home water fluoride concentration, fluoride intake from other sources, sugar-sweetened beverages, and 100% juice. This is the first study to combine ML models to predict caries trajectories from childhood to adulthood with high accuracy. Additional work is needed for validation using diverse clinical data. Predicting caries trajectories via ML could enable early identification of individuals at high risk and inform targeted, age-appropriate preventive interventions.