Artificial intelligence-powered prediction of orthodontic treatment duration: Identifying short and prolonged cases from a dataset of more than 10,000 patients
Abstract
BACKGROUND: Treatment duration is a major factor influencing patient motivation, acceptance, and clinical workflow in orthodontics. Accurately estimating treatment length and identifying cases likely to finish very early or require extended time can significantly enhance treatment planning and patient communication. This study aimed to develop and compare machine learning (ML) models for predicting continuous orthodontic treatment duration and for classifying short-duration (≤12 months) and long-duration (≥30 months) treatments using a large, diverse clinical dataset. METHODS: A total of 10,553 completed orthodontic cases were extracted from 120,000 records in the FMBO database. Seventy percent of the data was used for training and 30% for testing. Seven regression algorithms (Linear Regression, OLS, Bagging, Random Forest, Gradient Boost, AdaBoost, XGBoost) were evaluated. Classification models (Logistic Regression, Random Forest, Gradient Boost, XGBoost, AdaBoost, Bagging) were developed to identify short and long treatments. Class imbalance was addressed using class weighting and optimized probability thresholds (0.24 for short, 0.16 for long). Model performance was assessed using MSE, R2, precision, recall, F1-score, confusion matrices, and AUC. Explainability tools (feature importance, permutation importance, LIME) were applied to identify key predictors. RESULTS: The Gradient Boost Regression model demonstrated the strongest performance for continuous duration prediction (R2 = 0.387; MSE = 23.0). The most influential predictors across regression models were age at treatment start, extraction plan, facial convexity, skeletal pattern, overjet, and overbite. For short-duration classification, the Gradient Boost Classifier achieved an AUC of 0.867 (test set) with a recall of 0.630, a precision of 0.463, and F1 = 0.534 at the optimized threshold (0.24). Age at T1, upper incisor position (A1toAPoP), and molar position (A6toPTV) were the strongest contributors. For long-duration classification, the Random Forest Classifier achieved a test-set AUC of 0.793 with precision 0.437, recall 0.438, and F1 = 0.438 at the optimized threshold (0.16). Permutation importance identified lower incisor AP position (B1toAPoP), upper incisor AP position (A1toAPoP), overbite (IOB), and age at T1 as the most influential predictors. CONCLUSIONS: ML models demonstrated effective performance in predicting orthodontic treatment duration and in identifying fast- and slow-finishing cases. Key demographic, dental, and skeletal features—particularly age, extraction plan, incisor AP position, and vertical/incisal characteristics—were consistently strong predictors. The integration of big data and ML provides a scalable approach to enhance orthodontic diagnosis, optimize scheduling, improve communication of expectations, and support individualized treatment planning. Future studies incorporating compliance, biomechanics, and 3D imaging may further strengthen prediction accuracy.
Related
- Harnessing Artificial Intelligence to Predict Dental Crowding in Modern Orthodontic Practice: A Comprehensive Review
- WITHDRAWN: Relevance Based Prediction: A Transparent, Non-Artificial Intelligence, Mathematical Solution to Personalized Opioid Treatment
- Artificial intelligence in hepatitis: Advances in diagnosis, treatment optimization, and prognostic prediction
- Artificial Intelligence in Eye Disease: A Systematic Review of Screening, Diagnosis, Prognosis, Treatment-Response Prediction, and Clinical Decision Support
- Artificial intelligence empowering the diagnosis and treatment of rotator cuff tears: A systematic review from image optimization to prognosis prediction.
BibTeX
@article{elnagar2026artificial,
title = {{Artificial intelligence-powered prediction of orthodontic treatment duration: Identifying short and prolonged cases from a dataset of more than 10,000 patients}},
author = {Mohammed H. Elnagar and Allen Y. Pan and Aryo Handono and Flavio Sanchez and Sameh Talaat and Christoph Bourauel and Ahmed Kaboudan and Budi Kusnoto},
journal = {Journal of Orthodontic Science},
year = {2026},
month = sep,
doi = {10.4103/jos.jos_186_25},
url = {https://doi.org/10.4103/jos.jos_186_25},
note = {CC BY-NC-SA},
}
RIS
TY - JOUR
TI - Artificial intelligence-powered prediction of orthodontic treatment duration: Identifying short and prolonged cases from a dataset of more than 10,000 patients
AU - Mohammed H. Elnagar
AU - Allen Y. Pan
AU - Aryo Handono
AU - Flavio Sanchez
AU - Sameh Talaat
AU - Christoph Bourauel
AU - Ahmed Kaboudan
AU - Budi Kusnoto
PY - 2026
DA - 2026/09/01
JO - Journal of Orthodontic Science
DO - 10.4103/jos.jos_186_25
UR - https://doi.org/10.4103/jos.jos_186_25
AB - BACKGROUND: Treatment duration is a major factor influencing patient motivation, acceptance, and clinical workflow in orthodontics. Accurately estimating treatment length and identifying cases likely to finish very early or require extended time can significantly enhance treatment planning and patient communication. This study aimed to develop and compare machine learning (ML) models for predicting continuous orthodontic treatment duration and for classifying short-duration (≤12 months) and long-duration (≥30 months) treatments using a large, diverse clinical dataset. METHODS: A total of 10,553 completed orthodontic cases were extracted from 120,000 records in the FMBO database. Seventy percent of the data was used for training and 30% for testing. Seven regression algorithms (Linear Regression, OLS, Bagging, Random Forest, Gradient Boost, AdaBoost, XGBoost) were evaluated. Classification models (Logistic Regression, Random Forest, Gradient Boost, XGBoost, AdaBoost, Bagging) were developed to identify short and long treatments. Class imbalance was addressed using class weighting and optimized probability thresholds (0.24 for short, 0.16 for long). Model performance was assessed using MSE, R2, precision, recall, F1-score, confusion matrices, and AUC. Explainability tools (feature importance, permutation importance, LIME) were applied to identify key predictors. RESULTS: The Gradient Boost Regression model demonstrated the strongest performance for continuous duration prediction (R2 = 0.387; MSE = 23.0). The most influential predictors across regression models were age at treatment start, extraction plan, facial convexity, skeletal pattern, overjet, and overbite. For short-duration classification, the Gradient Boost Classifier achieved an AUC of 0.867 (test set) with a recall of 0.630, a precision of 0.463, and F1 = 0.534 at the optimized threshold (0.24). Age at T1, upper incisor position (A1toAPoP), and molar position (A6toPTV) were the strongest contributors. For long-duration classification, the Random Forest Classifier achieved a test-set AUC of 0.793 with precision 0.437, recall 0.438, and F1 = 0.438 at the optimized threshold (0.16). Permutation importance identified lower incisor AP position (B1toAPoP), upper incisor AP position (A1toAPoP), overbite (IOB), and age at T1 as the most influential predictors. CONCLUSIONS: ML models demonstrated effective performance in predicting orthodontic treatment duration and in identifying fast- and slow-finishing cases. Key demographic, dental, and skeletal features—particularly age, extraction plan, incisor AP position, and vertical/incisal characteristics—were consistently strong predictors. The integration of big data and ML provides a scalable approach to enhance orthodontic diagnosis, optimize scheduling, improve communication of expectations, and support individualized treatment planning. Future studies incorporating compliance, biomechanics, and 3D imaging may further strengthen prediction accuracy.
SN - 2278-1897
SN - 2278-0203
ER -