Machine Learning-Based Model for Predicting Short- and Long-Term Growth in Untreated Class III Malocclusion.

OBJECTIVE: To develop a Machine Learning-based model to predict growth of Caucasian subjects with untreated Class III malocclusion in the short- and long-term. MATERIALS AND METHODS: A longitudinal sample of 144 Caucasian subjects with untreated Class III malocclusion was selected (80% training data; 20% test data). Cephalograms of the subjects of the test group were divided into short- and long-term observations. Sixteen cephalometric landmarks were digitised in an X-Y Cartesian coordinate system. The trained model was a Graph Neural Network. A one-sample t-test and the Euclidean distances between predicted and observed values were calculated. RESULTS: In the short-term prediction, 16 subjects were examined. On the X-Y axis, the following cephalometric points were statistically significant: SX, PgY, BY, PNSY, NY and SY. Mean Euclidean distance between predicted and actual values revealed high values for the mandibular points Go (2.6 mm), Me (1.9 mm), Gn (1.9 mm), Pg (2.0 mm), and B (2.0 mm). In the long-term prediction, 13 subjects were examined. On the X-Y axis, the following cephalometric points were statistically significant: MeX, GnX, PgX, BX, BY and PtY. Mean Euclidean distance between predicted and actual values revealed high values for the mandibular points Go (3.1 mm), Me (4.3 mm), Gn (4.1 mm), Pg (4.5 mm), and B Point (4.1 mm). CONCLUSIONS: The ML-based prediction model was accurate for the majority of the landmarks. Cephalometric mandibular landmarks (Go, Me, Gn, Pg, and B Points) showed the highest mean Euclidean distances between predicted and observed values, indicating lower prediction accuracy.