# Open Bite Classification Using Machine Learning: A Cephalometric Analysis

## TL;DR

This study uses machine learning to accurately classify and understand different types of anterior open bite malocclusion in orthodontic patients.

## Contribution

The study introduces an interpretable machine learning model for diagnosing and phenotyping anterior open bite using cephalometric data.

## Key findings

- The decision tree classifier achieved 96.2% accuracy in distinguishing open bite from healthy subjects.
- ML-NSL was the most influential feature in classification, followed by facial axis and PFH/AFH.
- Unsupervised clustering identified ten craniofacial clusters, including mixed groups with intermediate skeletal patterns.

## Abstract

Background: Anterior open bite (AOB) is a complex malocclusion characterized by different vertical craniofacial growth and heterogeneous skeletal patterns, making objective diagnosis challenging using conventional cephalometric assessment alone. Recent advances in machine learning offer new opportunities to improve phenotypic characterization and diagnostic accuracy in orthodontics. Methods: This retrospective study analyzed lateral cephalometric records from 1056 orthodontic patients, comprising 621 patients with an anterior open bite and 435 healthy controls, all of whom were from the Arab population in Israel. Five clinically relevant cephalometric parameters related to vertical skeletal relationships were evaluated: the mandibular plane angle (ML-NSL), palatal plane angle (NL-NSL), posterior to anterior facial height ratio (PFH/AFH), gonial angle, and the facial axis. Statistical comparisons were made between the open bite and healthy subgroups, and these analyses were conducted in an exploratory framework to support hypothesis generation. A decision tree classifier was developed to distinguish AOB from healthy subjects using these features, and model performance was evaluated on a hold-out test set. Additionally, agglomerative hierarchical clustering was applied to explore latent craniofacial phenotypes. Results: Significant differences in vertical skeletal parameters were observed between open-bite and healthy subjects across various subgroups. The decision tree classifier achieved a test accuracy of 96.2%, with a precision, recall, and F1-score of approximately 0.97. ML-NSL emerged as the most influential feature, followed by facial axis and PFH/AFH. Unsupervised clustering identified ten distinct craniofacial clusters, including pure open bite and pure healthy phenotypes, as well as mixed clusters representing borderline or intermediate skeletal patterns. Clusters dominated by open bite cases exhibited steep mandibular planes, reduced PFH/AFH ratios, increased gonial angles, and decreased facial axis values, consistent with known vertical dysplasia patterns. Conclusions: Machine learning applied to cephalometric data enables accurate classification and meaningful phenotypic stratification of anterior open bite malocclusion. Beyond binary diagnosis, clustering analysis reveals clinically relevant subgroups that reflect varying degrees and types of vertical skeletal imbalance. These findings support the potential role of interpretable machine learning models as decision-support tools in orthodontic diagnosis and personalized treatment planning.