Novel Machine Learning Algorithms for Prediction and Treatment Decision in Patients with Class III

The decision to treat adult patients with Class III malocclusion is complicated by limited options: orthodontic camouflage or orthognathic surgery. The decision-making process is often guided by the clinician's expertise in managing similar presentations of the malocclusion; and influenced by a myriad of phenotypic and psychosocial factors unique to each patient. Various radiographic and photographic features extracted from the pretreatment records are typically analyzed visually, and perhaps by discriminate analysis formulas, to make the selection approach more objective. Despite the high predictive accuracy percentages reported by those methods, they were limited by computational power and lacked account for the multitude of nonquantifiable variables thought to impact the treatment decision. Aims: 1) Identify morphological characteristics (key demographic, radiographic, and clinical parameters) that affect the treatment decision for non-growing Class III patients. 2) Establish a comprehensive data set for training and testing of the Machine Learning (ML) model/Artificial Neural Networks (ANNs). 3) Conduct internal data mining and use descriptive statistics to investigate the extent of difference in parameter values between the two groups and determine their statistical significance. 4) Build, test, and validate various ML models and compare their predictive accuracy. 5) Calculate the relative contribution of the features in each network model. Methods: Pretreatment records of 182 patients (118 surgical:65 camouflage) who received treatment post their pubertal growth spurts were analyzed and 40 demographic, radiographic, and clinical parameters were collected for each case. Data mining steps were applied to the parameters to identify statistical difference between the two groups. The cases were also divided into a testing and validation set for eight different ML models: Support Vector Machine (SVM), Random Forest, k-Nearest Neighbor (kNN), Logistic Regression, Multi-Layer Preceptron (MLP), Convolutional Neural Network (CNN), Extreme Gradient Boosting (XGBoost), and pruned XGBoost + Selected Features. The performance metrics of the various models were calculated and compared. Results: retreatment parameters vary in their contribution to the therapy approach in patients of Class III malocclusion. Different machine learning models can produce the desired binary output with an accuracy percentage ranging from 78% for the kNN algorithm to 93% for the pruned XGBoost model. The three cephalometric variables of Wit’s appraisal, mx/md ratio, and overjet showed both statistical significance with independent t-test comparison between the camouflage and orthognathic groups, and high weight values across the ML algorithms tested. Conclusion: a highly predictive artificial intelligence model can be developed that is more accurate than all existing RBES and CBES statistical models.