Classification of skeletal discrepancies by machine learning based on three-dimensional facial scans

The aim of this study was to use machine learning (ML) to classify sagittal and vertical skeletal discrepancies in three-dimensional (3D) facial scans, as well as to evaluate shape variability. 3D facial scans from 435 pre-orthodontic patients were subjected to cephalometric analysis and 3D facial landmark identification. Three ML models were used for the discrimination of skeletal discrepancy: random forest, AdaBoost, and multi-layer perceptron. Each model was evaluated by receiver operating characteristic curve and calculating the area under the curve (AUC). Principal component analysis was conducted to evaluate shape variability. The AUCs for Class II and III patients ranged from 0.91 to 0.95. Random forest achieved the highest accuracy for sagittal classification (88.5% for Class II, 95.5% for Class III). Multi-layer perceptron exhibited the best performance for vertical classification (accuracy of 78.8% for hypodivergent, 86.2% for hyperdivergent). Six principal components explained 94.0% of facial morphology variation. ML methods show promise for assisting in the discrimination of sagittal and vertical skeletal discrepancies based on 3D facial scans. 3D facial soft tissue features appear to be suitable for the discrimination of skeletal discrepancies in most cases.