Machine learning to predict distal caries in mandibular second molars associated with impacted third molars

Abstract

Impacted mandibular third molars (M3M) are associated with the occurrence of distal caries on the adjacent mandibular second molars (DCM2M). In this study, we aimed to develop and validate five machine learning (ML) models designed to predict the occurrence of DCM2Ms due to the proximity with M3Ms and determine the relative importance of predictive variables for DCM2Ms that are important for clinical decision making. A total of 2642 mandibular second molars adjacent to M3Ms were analyzed and DCM2Ms were identified in 322 cases (12.2%). The models were trained using logistic regression, random forest, support vector machine, artificial neural network, and extreme gradient boosting ML methods and were subsequently validated using testing datasets. The performance of the ML models was significantly superior to that of single predictors. The area under the receiver operating characteristic curve of the machine learning models ranged from 0.88 to 0.89. Six features (sex, age, contact point at the cementoenamel junction, angulation of M3Ms, Winter's classification, and Pell and Gregory classification) were identified as relevant predictors. These prediction models could be used to detect patients at a high risk of developing DCM2M and ultimately contribute to caries prevention and treatment decision-making for impacted M3Ms.

Figure 1

Polar histogram presenting the prevalence of DCM2Ms. DCM2M distal caries in mandibular second molars, M2Ms mandibular second molars, CEJ cementoenamel junction.

Figure 2

Receiver operating characteristic curves plotted from testing dataset. The optimal threshold is plotted as the point closest to the top-left part of the plot. AUROC area under the ROC curve, CI confidence interval.

Figure 3

Relative feature importance computed using the Boruta algorithm. The blue violin plots correspond to the minimal, average, and maximum Z scores of a shadow attribute. The red and green violin plots represent the Z scores of the rejected and confirmed attributes, respectively. The black dots and horizontal lines within each violin plot represent the mean and median values, respectively. All features that received a lower relative feature importance than that of the shadow feature were defined as irrelevant for prediction. Laterality was considered as an irrelevant feature (marked in red).