Hierarchical clustering analysis & machine learning models for diagnosing skeletal classes I and II in German patients

Background: Classification is one of the most common tasks in artificial intelligence (AI) driven fields in dentistry and orthodontics. The AI abilities can significantly improve the orthodontist’s critical mission to diagnose and treat patients precisely, promptly, and efficiently. Therefore, this study aims to develop a machine-learning model to classify German orthodontic patients as skeletal class I or II based on minimal cephalometric parameters. Eventually, clustering analysis was done to understand the differences between clusters within the same or different skeletal classes. Methods: A total of 556 German orthodontic patients were classified into skeletal class I (n = 210) and II (n = 346) using the individualized ANB. Hierarchical clustering analysis used the Euclidean distances between data points and Ward’s minimum variance method. Six machine learning models (random forest (RF), K-nearest neighbor (KNN), support vector machine (SVM), linear discriminant analysis (LDA), classification and regression trees (CART), and General Linear Model (GLM)) were evaluated considering their accuracy, reliability, sensitivity, and specificity in diagnosing skeletal class I and II. Results: The clustering analysis results showed the power of this tool to cluster the results into two–three clusters that interestingly varied significantly in many cephalometric parameters, including NL-ML angle, NL-NSL angle, PFH/AFH ratio, gonial angle, SNB, Go-Me (mm), Wits appraisal, ML-NSL, and part of the dental parameters. The CART model achieved 100% accuracy by considering all cephalometric and demographic variables, while the KNN model performed well with three input parameters (ANB, Wits, SNB) only. Conclusions: The KNN model with three key variables demonstrated sufficient accuracy for classifying skeletal classes I and II, supporting efficient and still personalized orthodontic diagnostics and treatment planning. Further studies with balanced sample sizes are needed for validation.