IADR Abstract Archives
Long-Term Predictive Modeling of Craniofacial Complex Using Machine Learning
Objectives: Accurately predicting future growth-related changes in skeletal and dental relationships within the craniofacial complex is of utmost importance for the success and stability of the growth modification therapies. Therefore, the objective of this study was to predict long-term growth-related changes in skeletal and dental relationships within the craniofacial complex by using machine learning.
Methods: Lateral cephalometric radiographs from 301 patients were obtained at two time points (T1: age 11, T2: age 18) from American Association of Orthodontists Foundation (AAOF) Legacy Collection. Each radiograph underwent manual tracing using 15 anatomical landmarks. Three distinct ML models were trained on a subset of 250 subjects, with an additional 61 subjects reserved for testing model accuracy. The mean absolute error (MAE), intraclass correlation coefficients (ICCs) and t-test were employed to assess model accuracy.
Results: MAEs for the skeletal measurements ranged between 1.36° (SNA) and 4.10 mm (mandibular length), while they ranged between 1.26 mm (L1 – NB) and 5.48° (U1 - NA) for the dental measurements. The agreement between the actual and the predicted measurements for each method were good to excellent as measured by the ICC. None of the paired t-tests used to test for a consistent difference (bias) between the predicted and actual measurements were statistically significant (P>.05), except L1 - MP (°) (P<.05). The highest percentages of cases accurately predicted within 2 mm/° clinical threshold were obtained for SNA (°), L1-NB (mm), and ANB (°) (80%, 75% and 70%, respectively).
Conclusions: Within the craniofacial complex, long-term growth-related changes in skeletal relationships were predicted more accurately than changes in dental relationships. Although each algorithm selected several features for its predictions, pre-pubertal values consistently emerged as the most important factor for predicting post-pubertal outcomes. No significant differences were observed between ML methods regarding the absolute differences between predicted and actual measurements.
Methods: Lateral cephalometric radiographs from 301 patients were obtained at two time points (T1: age 11, T2: age 18) from American Association of Orthodontists Foundation (AAOF) Legacy Collection. Each radiograph underwent manual tracing using 15 anatomical landmarks. Three distinct ML models were trained on a subset of 250 subjects, with an additional 61 subjects reserved for testing model accuracy. The mean absolute error (MAE), intraclass correlation coefficients (ICCs) and t-test were employed to assess model accuracy.
Results: MAEs for the skeletal measurements ranged between 1.36° (SNA) and 4.10 mm (mandibular length), while they ranged between 1.26 mm (L1 – NB) and 5.48° (U1 - NA) for the dental measurements. The agreement between the actual and the predicted measurements for each method were good to excellent as measured by the ICC. None of the paired t-tests used to test for a consistent difference (bias) between the predicted and actual measurements were statistically significant (P>.05), except L1 - MP (°) (P<.05). The highest percentages of cases accurately predicted within 2 mm/° clinical threshold were obtained for SNA (°), L1-NB (mm), and ANB (°) (80%, 75% and 70%, respectively).
Conclusions: Within the craniofacial complex, long-term growth-related changes in skeletal relationships were predicted more accurately than changes in dental relationships. Although each algorithm selected several features for its predictions, pre-pubertal values consistently emerged as the most important factor for predicting post-pubertal outcomes. No significant differences were observed between ML methods regarding the absolute differences between predicted and actual measurements.