Orthodontic diagnosis, as the cornerstone of successful treatment, involves identifying and understanding malocclusions to create individualized treatment plans aimed at achieving optimal oral health, function and aesthetics. A careful analysis of the diagnostic database obtained through the patient’s medical and dental history, clinical examination as well as diagnostic records generates a problem list, which offers insights into the malocclusion's underlying causes and guides the formulation of treatment objectives and plan. (1)

An important aspect of orthodontic diagnosis is the accurate assessment of the anterior vertical relationship, as it plays a key role in evaluating occlusal function, esthetics, and treatment complexity. A normal overbite is around 2–4 mm of overlap or one-third coverage of the lower incisor, while greater or lesser overlap is classified as deep bite or open bite, respectively. (1) Several established indices such as the Peer Assessment Rating (PAR) Index and the Index of Orthodontic Treatment Need (IOTN) include overbite as a key component in measuring malocclusion severity and guide treatment planning. (2),(3) The Index of Complexity, Outcome, and Need (ICON) integrates overbite into its multifactorial scoring system, which considers both aesthetic and functional aspects of malocclusion. (4) It provides a well-established framework for classifying vertical relationships into standardized categories. These classifications range from normal bite to varying degrees of open and deep bites, allowing for a systematic approach to diagnosis and treatment evaluation. (4)

The anterior vertical relationship is traditionally evaluated during clinical examination and through analysis of pre-treatment records including study models, photographs and radiographs. However, such methods can vary significantly between practitioners due to differences in training, experience, and interpretation.

Artificial Intelligence (AI) is a specialized area of study in computer science and refers to the capability of machines to exhibit intelligence by acquiring and applying knowledge from data to address diverse challenges with minimal dependance on human interaction. (5) Integrating Artificial intelligence with dentists' clinical expertise would significantly improve the efficiency and functionality of dental office operations. Machine learning, a subfield of AI, utilizes algorithms to analyze datasets and make predictions, enabling machines to learn and independently solve problems without human intervention. Neural networks, including artificial neural networks (ANNs) and convolutional neural networks (CNNs), function by processing signals through interconnected artificial neurons, mimicking the human brain's operations. (5) Deep learning, a subset of machine learning, involves multi-layered computational networks that detect patterns within input data, enhancing feature recognition processes. CNNs, in particular, have proven effective in analyzing complex inputs like text and images. (6)

The application of artificial intelligence (AI) in the field of orthodontics has demonstrated a wide range of applications, including determination of cervical maturation, growth prediction, automated cephalometric tracing and landmark identification. (7), (8), (9) Furthermore, several articles have addressed the use of AI in diagnosing and planning orthognathic surgeries as well as developing automated skeletal classification systems. (10) In addition, previous work has shown the potential of AI algorithms to simplify therapeutic decision-making in orthodontic diagnostics, such as determining whether tooth extraction should be part of a treatment plan. (11) Talaat et al. validated an AI system using the Dental Health Component of the Index of Orthodontic Treatment Need (IOTN) to assess malocclusions from intraoral photographs. (12) Employing a YOLO-based convolutional neural network, the system accurately identified conditions such as crowding, spacing, crossbites, deep bites, open bites, and increased overjet, achieving near-perfect performance. Building on this, Bardideh et al. developed a multistage neural network trained on 7,500 images from 700 patients to classify occlusal relationships. (13) The AI demonstrated high accuracy in molar (93.1%) and canine (89.1%) classifications, closely matching expert orthodontists. However, its performance in quantifying overjet and overbite was less consistent, with mean absolute errors of 1.98 mm and 1.28 mm, highlighting the need for further refinement. Stetzel et al. focused on dental aesthetics, introducing a deep learning model to predict the Aesthetic Component of the IOTN. (14) Using 1,009 frontal intraoral photographs, the binary classification approach (scores 1–5 vs. 6–10) produced robust outcomes, with 77% sensitivity, 88% specificity, and 82% accuracy. Together, these studies demonstrate AI’s potential to support orthodontic assessment by automating diagnosis, improving objectivity, and reducing clinician workload, while also identifying areas where further optimization is required.

Up to the time of this writing, AI applications have been investigated in various aspects of orthodontic diagnosis and treatment planning, however, none have specifically addressed the detection and categorization of open bite and deep bite using AI. This study thus aimed to develop a fully automated AI model capable of quantifying and categorizing openbite and deepbite, and to validate the model's results by comparing them with those of human observers. By integrating AI technology with the ICON framework, it is possible to create a system that not only automates the measurement and classification of anterior vertical relationships but also aligns with established clinical guidelines. Such a system could serve as a valuable tool for both clinicians and researchers, streamlining diagnostic workflows and enhancing the quality of care.

Approval for the study was obtained by the Research Ethics Committee in the University of …… (approval number (REC-23-10-08-03-PG)). Consent for the use of patient records from the University database was already obtained from an informed consent signed by the patients prior to initiating orthodontic treatment.

The study was structured in three phases: model training, validation, and final testing. For training and validation, a dataset of 750 intraoral clinical images was sourced from orthodontic records and divided into two subsets. Of these, 660 images (88%) were used to train the neural network models, while 90 images (12%) were designated to monitor the model's performance during training and fine-tune hyperparameters. An independent dataset of 300 additional intraoral images and scans formed the testing set.

The dataset consisted of records of orthodontic patients undergoing treatment in the Orthodontic Department, College of Dental Medicine, University of Sharjah between 2022 and 2024. Intraoral images of the patients were exported from the Dolphin Imaging and management solutions software (v.12, Chatsworth, CA, USA). All patients were scanned using the iTero intraoral scanner (Align Technology, San Jose, Ca, USA) and records of their pre-treatment intraoral scans were retrieved from the iTero website. Patients’ records were selected based on the following criteria: patients with a full permanent dentition (excluding third molars), scans taken prior to the initiation of any orthodontic treatment and that included both arches in occlusion. Additionally, cases with any type of malocclusion were eligible for inclusion, allowing the study to evaluate the AI model's performance across a wide range of clinical scenarios. Scans of patients in the primary or mixed dentition and cases where frontal intraoral photos were taken at an angle that could affect overbite detection were excluded.

A validation set of 90 cases was used to validate the generated AI model. Discrepancies were identified during this phase, particularly in cases where teeth were in crossbite, in detecting normal overbite, and in borderline cases between two categories. Retraining of the AI model was thus carried out to address these shortcomings. Once the generated model was validated, the following step was to test the model. Validation and test loss were monitored, with early stopping techniques applied to prevent overfitting, ensuring robust generalization to real-world data.

The sample size required for testing of the AI model by comparing it to human observers was determined using a web-based calculator (https://wnarifin.github.io/ssc/sskappa.html). With a minimum acceptable Kappa of 0.8, expected Kappa of 0.9, outcome proportion of 0.5, significance level of 0.05, and study power of 80%, the sample size was calculated to be a minimum of 283 patients. Accordingly, 300 pre-treatment intraoral scans were selected for the testing dataset. Sample demographics including age, gender, and race were noted from the patients’ records in aXium software (Exan Group, a Henry Schein company, Vancouver, BC, Canada).

The AI model was trained to measure and classify the anterior vertical relationship using the intraoral frontal photographs. The human observer performed the same assessment on the intraoral scans using OrthoCAD software (Cadent, Fairview, NJ, USA). A cross-section was taken passing through the middle of the incisal edge of the maxillary central incisor, and the overbite was measured as the vertical distance from the incisal edge of the maxillary central incisor to the incisal edge of the lower incisor that appears on the cross-section to obtain the overbite measurement in millimeters (Fig. 2). To calculate the percentage of vertical overlap, the obtained measurement was divided by the clinical crown height of the lower incisor. Openbite cases were expressed in millimetres. The higher value for the overbite/openbite was then selected from either the right or left central incisor. These data were further categorized based on the ICON index for incisor overbite/openbite (Table 1).

The corresponding set of frontal intraoral photographs were extracted from the Dolphin Management software. These photographs were then analyzed by the AI model. The AI model's results, including its measurements and categorizations, were compiled into an Excel sheet and compared with those assigned by the human examiners to evaluate the model's accuracy.

All measurements were performed by the primary investigator. To determine intra-examiner reliability, twenty scans of the included sample were randomly selected to be measured again two weeks after the first assessment. To determine inter-examiner reliability, the same scans were measured by another examiner. Error measurement was then assessed using interclass correlation coefficient.

Table 2 shows the distribution and agreement between the anterior vertical relationship categories as determined by the AI and human observers and accuracy of the AI model. The AI model demonstrated excellent agreement with manual assessments, achieving a Kappa coefficient of 0.89 (p < 0.0001) and an overall accuracy of 92%. Discrepancies were minimal, with a maximum difference of 3% between classifications. The chi-square test (χ² = 2.327, df = 7, p = 0.940) confirmed no significant difference between AI and manual outcomes, indicating strong reliability in classifying anterior vertical relationships. Residual analysis showed the largest deviations in normal bite (9 cases; 40.9%), deep bite with 1/3–2/3 coverage (7 cases; 31.8%), open bite < 1 mm (4 cases; 18.2%), and deep bite with full coverage (2 cases; 9.1%). Perfect agreement was achieved in deep bite (2/3–full coverage), open bite 1.1–2 mm, 2.1–4 mm, and > 4 mm categories.

*Significant difference as P ≤ 0.05.

For deep bite detection, sensitivity was 95.9% (95% CI: 91.7–98.3%), specificity 100% (95% CI: 95.3–100%), and accuracy 97.2% (95% CI: 94.2–98.9%), with an AUC of 0.979 (95% CI: 96.2–99.7%). The positive predictive value (PPV) was 100%, while the negative predictive value (NPV) was 91.7%.

For open bite detection, sensitivity reached 96.3% (95% CI: 87.3–99.5%), specificity 100% (95% CI: 95.3–100%), and accuracy 98.4% (95% CI: 94.6–99.8%), with an AUC of 0.98 (95% CI: 95.2–100%). Both PPV and NPV were high at 100% and 97.4%, respectively.

Reliability testing confirmed exceptional consistency. Intraobserver agreement yielded an ICC of 0.999 (95% CI: 0.997–1.000; F = 895.549; p < 0.0001), while interobserver reliability produced an ICC of 0.998 (95% CI: 0.994–0.999; F = 443.813; p < 0.0001), reflecting nearly perfect reproducibility.

This study aimed to develop a fully automated AI system capable of measuring open and deep bites and categorizing them based on the ICON index, with validation achieved through comparison to assessments conducted by human observers. The AI system demonstrated high agreement with human observers, achieving an accuracy of 92% and a Kappa coefficient of 0.89. This highlights AI’s potential to standardize diagnosis, reduce inter-clinician variability, and support clinical decision-making. The AI model was specifically trained using two-dimensional intraoral images rather than three-dimensional scans to enhance generalizability and integration into routine orthodontic workflows, as photos are part of the diagnostic phase of every orthodontic treatment making them readily available.

Bardideh et al. reported 93.1% accuracy using intraoral photographs for occlusion classification. (13) However, their AI system showed reduced accuracy in measuring overjet and overbite, likely due to the limitations of 2D image analysis, which lacks the depth perception available to human clinicians during 3D evaluations. Talaat et al. achieved an accuracy of 99.99% in assessing orthodontic treatment need, although their task was broader and less detailed, potentially explaining the higher performance. (12) Akl et al. developed a software for diagnosing anterior open bite using a standardized checklist and reported 82% agreement with experts, (16) while El-Dawlatly et al. validated their AI-based decision support system for deep bite treatment planning with a precision of 94.4%, (17) closely matching the 92% accuracy in the current study. In contrast, Stetzel et al. reported lower accuracy (82%) for AI-based prediction of the aesthetic component of the IOTN. (14) Factors contributing to this discrepancy included the lack of image segmentation and potential racial variability in gingival features, which affected the AI's ability to isolate dental structures.

One of the key strengths of this study lies in its advanced application of AI beyond basic malocclusion identification. Unlike previous studies, which focused on general diagnostic tasks such as identifying the presence of malocclusion or broadly classifying open and deep bite cases, the current study required the AI to perform automatic tooth segmentation followed by millimetric measurements of open bite cases and categorization of deep bite severity based solely on photographs. Notably, the AI achieved this level of accuracy without being fed additional data, such as the total crown length of the lower incisors. This distinction represents a significant advancement in AI applications for dental diagnostics. This study thus offered a more objective method that can minimize variation among orthodontists, reduce their workload, and optimize the time spent prioritizing and determining the complexity of cases. It also serves as a valuable tool for training inexperienced doctors and general dentists, while paving the way for advancements in automated systems that facilitate orthodontic treatment planning and decision-making.

The AI system demonstrated strong adaptability, managing variations in image quality, scale calibration, and lighting across a diverse patient population. It was trained to identify the most severe vertical discrepancy between right and left segments, successfully categorizing borderline and complex cases—including anterior crossbites—without requiring standardized calibration tools. This real-world adaptability, combined with high accuracy in identifying normal, open, and deep bite cases, underscores the model’s reliability across a broad range of clinical conditions. Preprocessing techniques (e.g., image denoising, contrast enhancement) and advanced post-processing algorithms enabled the model to analyze lower-quality images and refine classification results. Training on images from different cameras and conditions also minimized overfitting, improving generalizability across clinical scenarios. Importantly, this study stands out for its complexity. It challenged the AI to perform measurements using non-standardized clinical photographs and perform comparably to human examiners which affirms its clinical readiness and potential integration into daily orthodontic practice.

Nevertheless, study limitations must be acknowledged. Variations in photo angle may have impacted measurement precision, and all cases were derived from pre-treatment photos of patients with healthy gingiva. The model’s performance in cases with gingival recession or hyperplasia remains unknown. Unlike human clinicians, the AI cannot yet account for changes in crown visibility caused by soft tissue variations, which may affect overbite assessment.

The developed AI model demonstrated high accuracy (92%) and excellent agreement with manual measurements, confirming its potential as a reliable and objective tool for automated quantification of anterior vertical relationships in orthodontic diagnosis.