Validation of Machine Learning Models for Craniofacial Growth Prediction

Area of Science:

• Orthodontics
• Craniofacial Biology
• Biomedical Engineering

Background:

• Predicting longitudinal craniofacial growth is crucial for effective orthodontic treatment planning.
• Accurate growth prediction aids in personalized treatment strategies and improved outcomes.
• Previous models have limitations in predicting individual growth trajectories.

Purpose of the Study:

• To identify the most accurate predictive model for longitudinal craniofacial growth in a Japanese pediatric population.
• To compare the performance of various statistical and machine learning models in growth prediction.
• To evaluate the utility of the best-performing model for future orthodontic applications.

Main Methods:

• Longitudinal lateral cephalometric radiographs from 59 Japanese children (ages 6-12) were analyzed.
• Models included multiple regression, least absolute shrinkage and selection operator (LASSO), radial basis function network, multilayer perceptron, and gradient-boosted decision tree.
• Independent variables comprised skeletal landmarks, linear, and angular parameters; dependent variables were future cephalometric values at age 13.
• Root-mean-square error was used to assess prediction accuracy.

Main Results:

• The LASSO model demonstrated the smallest average error across skeletal landmarks, linear, and angular parameters.
• Highest prediction accuracies achieved were 97.87% for skeletal linear parameters and 94.45% for skeletal angular parameters at age 13.
• The LASSO model effectively integrated multiple independent variables for robust prediction.

Conclusions:

• The LASSO model is the most accurate for predicting longitudinal craniofacial growth in the studied Japanese population.
• This predictive capability is highly valuable for personalized orthodontic treatment planning.
• The model's accuracy supports its application in anticipating individual growth patterns for clinical use.