Evaluation of different machine learning algorithms for extraction decision in orthodontic treatment

Area of Science:

• Orthodontics
• Machine Learning in Healthcare
• Dental Treatment Planning

Background:

• The orthodontic extraction decision critically impacts treatment outcomes.
• Objective and standardized methods are needed for consistent extraction decisions.
• Machine learning (ML) offers potential for standardizing clinical decisions.

Purpose of the Study:

• To identify the best-performing ML model for standardizing orthodontic extraction decisions.
• To determine key variables influencing extraction decisions.
• To guide clinicians, especially those with less experience.

Main Methods:

• Trained seven ML models on data from 1000 orthodontic patients (500 extraction, 500 non-extraction).
• Utilized 36 variables including demographics, cephalometric, and model measurements.
• Evaluated models using accuracy and Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve.

Main Results:

• The Stacking Classifier model demonstrated the highest performance for extraction decisions (91.2% AUC).
• Key variables for extraction decisions included arch length discrepancy, Wits Appraisal, and ANS-Me length.
• The Stacking Classifier also performed best for extraction type decisions (76.3% accuracy), with key variables including mandibular arch length discrepancy and cephalometric overbite.

Conclusions:

• The Stacking Classifier model is the top performer for standardizing orthodontic extraction decisions.
• ML models show high efficacy in extraction decision-making but less so in determining extraction type.
• This approach can aid in standardizing treatment planning and supporting clinical judgment.