Trajectories of Dental Caries From Childhood to Young Adulthood: Unsupervised Machine Learning Approach

OBJECTIVE: To determine the dental caries trajectories over the life course (from age 9 to 23) using an unsupervised machine learning approach. METHODS: This is a longitudinal study of caries trajectories over a life course using data from 1,382 individuals from the Iowa Fluoride Study birth cohort. The trajectory analysis of caries in the permanent dentition at ages 9, 13, 17 and 23 was performed using the unsupervised machine learning algorithm known as K-means for Longitudinal Data (KmL), a k-means based clustering algorithm implemented in R specifically designed for analyzing longitudinal data. The trajectory grouping was performed by assessing the distances of the individual trajectories from the centroid and the prediction of the “best” partition was performed based on the Calinsky & Harabatz criterion. The number of cluster partitions assessed was 2 to 6. The number of re-runs with different starting conditions for each number of clusters was 20. RESULTS: The trajectory analysis identified three trajectory groups with 70.5%, 21.1%, and 8.4% of participants in the low, medium, and high caries trajectory groups, respectively. The mean D(2+)MFS counts of the low caries trajectory groups at ages 9, 13, 17, and 23 were 0.23, 0.37, 1.10, and 1.56, respectively. The mean D(2+)MFS counts of the medium caries trajectory groups at ages 9, 13, 17, and 23 were 0.92, 2.09, 6.24, and 9.55, respectively. The mean D(2+)MFS counts of the high caries trajectory groups at ages 9, 13, 17, and 23 were 1.49, 4.80, 12.91, and 22.52, respectively. There were steeper increases in the D(2+)MFS scores of the three trajectory groups between age 13 and 17, with less steep but also strongly positive slopes from age 17 to 23, suggesting that the period from age 13 to 17 is the highest risk period. CONCLUSION: There was an increase in the trajectory slopes after age 13 which might be due to changes in risk factors. The next step in this study will be to identify those factors that predict trajectory group membership by modeling their relationships using supervised machine learning techniques.