Machine Learning and Clustering Analysis of Class II and III Malocclusions

This prospective observational study aims to accurately classify individuals as skeletal class II/III by applying several machine-learning algorithms. Furthermore, using the k-means clustering analysis deepened our understanding of the characteristics of skeletal class II/III malocclusion (SCIIMO/SCIIIMO) patients.

This multicenter study consisted of 379 German orthodontic patients. Following their distribution to SCIIMO (n = 208, 54.8%) or SCIIIMO (n = 171, 45.1%) by the individualized ANB of Panagiotidis and Witt. Machine-learning analysis was performed using different models (KNN, RF, LDA, SVM, CART) to find the model with the highest accuracy in correctly classifying skeletal class II/III. After determining the ideal number of clusters using the Elbow approach, k-means clustering was finally used.

The KNN machine learning model demonstrated the ability of reduced cephalometric parameters to capture much of the same sagittal discriminatory information as the predefined class construct. The results showed that Wits-appraisal only leads to achieving high accuracy in determining skeletal class (accuracy = 94.72%), while adding SN-Pg angle improved the accuracy slightly to 95.86%. The clustering analysis showed that three or four clusters were optimal for SCIIMO and SCIIIMO patients and revealed interesting characteristics for each cluster and significance between the differences in all parameters except S-N and age.

A reduced parameter set that included only the Wits appraisal and SN-Pg angle enabled a correct classification. In the exploratory clustering analysis, three or four distinct clusters emerged as optimal when analyzing the differences among SCIIMO/SCIIIMO patients, emphasizing the complexity and diversity of malocclusion phenotype.