Utilization of Machine Learning Methods for Predicting Orthodontic Treatment Length

Abstract

Treatment duration is one of the most important factors that patients consider when deciding whether to undergo orthodontic treatment or not. This study aimed to build and compare machine learning (ML) models for the prediction of orthodontic treatment length and to identify factors affecting the duration of orthodontic treatment using the ML approach. Records of 518 patients who had successfully finished orthodontic treatment were used in this study. Seventy percent of the patient data were used for training ML models, and thirty percent of the data were used for testing these models. We applied and compared nine machine-learning algorithms: simple linear regression, modified simple linear regression, polynomial linear regression, K nearest neighbor, simple decision tree, bagging regressor, random forest, gradient boosting regression, and adaboost regression. We then calculated the importance of patient data features for the ML models with the highest performance. The best overall performance was obtained through the bagging regressor and adaboost regression ML methods. The most important features in predicting treatment length were age, crowding, artificial intelligence case difficulty score, overjet, and overbite. Without patient information, several ML algorithms showed comparable performance for predicting treatment length. Bagging and adaboost showed the best performance when patient information, including age, malocclusion, and crowding, was provided.

A Novel Machine Learning Model for Predicting Orthodontic Treatment Duration

“…This advantage stems from the stable and reliable estimations of variable relationships that linear models can provide. Elnagar et al yielded comparable results, highlighting advanced decision tree regression models as the most effective among ML models [38]. Their study also quantified accuracy through mean square error, yielding a value of 54.08.…”

“…Notably, unlike our research, Elnagar et al's investigation did not encompass cephalometric measurements. However, their five most important features were identified as patient age, upper/lower crowding, overjet, and AI score (treatment difficulty estimated by AI) [38].…”

“…These technologies have been harnessed to analyze radiographic images [23,[26][27][28][29][30][31][32][33], predict growth [24,34,35], optimize orthodontic treatment decision-making processes [13][14][15][16][17][19][20][21][22]25,36]. Regrettably, a limited number of studies have employed AI and ML methodologies in forecasting orthodontic treatment duration [37,38]. Within this subset, Dharmasena et al conducted a notable investigation utilizing two distinct ML algorithms, namely Naïve Bayes and Random Forest [37].…”

“…Perhaps the most relevant study to ours examined the implementation of a ML algorithm to predict orthodontic treatment duration [38]. This study evaluated the accuracy and comparison of ML algorithms in predicting orthodontic treatment duration.…”

“…The study concluded that decision tree models outperformed other methods (mean square error of 54.08) and revealed that age, malocclusion, and crowding were the most influential predictors. However, this study and other existing studies on this topic have relied on a limited number of independent variables, such as DI or specific questionnaires [6,7,38]. In contrast, our study is unique in its incorporation of cephalometric data for the purpose of predicting orthodontic treatment duration and represents a novel contribution to the field.…”