Automated artificial intelligence modeling for clinical decision-making of extraction versus non-extraction cases: Unleashing the potential for all

INTRODUCTION

With continued global research efforts, technology has advanced significantly. The past 25 years have seen a fourfold expansion in the field of information technology, and its clinical use in dentistry has had a considerable impact on both general and orthodontic dentistry.^[1] Artificial intelligence (AI), machine learning (ML), robotics, and other technologies have exceedingly decreased the amount of time needed, required human prowess, etc. Researchers and physicians in all domains can benefit from its sophisticated pattern identification and prediction algorithms.^[2-4] Originally, attempts were made to create handcrafted AI, which is the process of trying to fit every possible scenario and its related solution into a computer program. In some situations, it functions well. However, because AI is incapable of “learning” anything on its own, it is by no means capable of surpassing the human brain. Regarding contemporary AI, which aims to give machines the capacity to learn, it has the capacity to acquire capabilities that surpass those of its developer. ML is how we attempt to achieve AI.^[5]

In general, ML is the process of deriving a function from input data. Later, the function can convert fresh incoming data – such as sounds and images – into useful output data, such as facial and speech recognition. One may categorize ML into two processes: Testing and training. During the training process, a collection of functions, referred to as a model, is assessed using computer algorithms and training data to determine the optimal function. After the testing procedure, the selected function is evaluated to decide whether or not it is feasible by reevaluating the testing data.^[5]

Many areas of orthodontics can benefit from the application of AI (artificial learning), ML, and deep learning, given the quickly expanding and developing disciplines of technological advancements.^[6] When used in orthodontics, AI and ML systems offer useful instruments that can enhance clinical procedures. Orthodontists practicing with these clinical decision support tools can work more productively, with less subjectivity and variability.^[7,8] Digital data processing, automated cephalometric tracing, diagnosis, and treatment planning, predicting the need for orthodontic extractions, facial image recognition, a framework for automated diagnosis and computer-assisted planning in plastic and reconstructive surgeries, identifying the cervical vertebrae maturity stages for growth prediction, and more are among the applications of AI in the orthodontic specialty.^[9] Rather than focusing solely on the oral and maxillofacial region, orthodontists must make a holistic evaluation of their patients by examining various facial structures from various angles.

To increase evaluation speed and accuracy, diagnostic imaging must be automated.^[1] A proper diagnosis and carefully thought-out treatment procedures are essential for the effectiveness of orthodontic therapy. Making a treatment plan is the most crucial step in orthodontic therapy. Choosing which teeth to take and whether to perform an extraction is a crucial step in treatment planning because extractions cannot be reversed.^[1] Based on their expertise and experience, orthodontists typically use information from radiography, dental models, photos, and clinical evaluations to determine the best course of action. In the process of treatment planning, this frequently results in intra-clinician and inter-clinician variability. Treatment plans may differ depending on which data were utilized for the diagnosis. Decisions on extraction that differ could be significant.^[10]

Automated systems based on AI can be very helpful in clinical decision-making and serve as a backup for physicians, particularly those with limited clinical expertise^.[11] This study was conducted to prepare AI decision-making models of extraction versus non-extraction cases and to compare the performance of three different ML classifier algorithms, that is, random forest classifier, decision tree classifier, and Extreme gradient boost (XGBoost) classifier, to provide researchers and clinicians with powerful pattern-finding algorithms.

MATERIAL AND METHODS

RESULTS

To understand the true utility of these models in a clinical setting, the ability of each model to correctly predict the treatment modality, referred to as model “accuracy,” must be determined. The confusion matrix [Figure 1] was used to calculate the accuracy of the performance of each model in diagnosing extraction and non-extraction cases for all data sets. Accuracy was computed from the confusion matrix for each test set, and the average accuracy across all test sets was reported for each model.

TP: True Positives

TN: True Negatives

FP: False Positives

FN: False Negatives

Accuracy is calculated as the ratio of true positives and true negatives to the total number of predictions (true positives, true negatives, false positives, and false negatives)^[12]

The obtained confusion matrices for extraction versus non-extraction cases by implementing the decision tree classifier [Figure 1a], random forest [Figure 1b], and XGBoost classifier algorithms [Figure 1c] for the overall dataset suggested that the decision tree classifier correctly predicted 76 non-extraction and 64 extraction instances but misclassified nine non-extraction cases as extraction and 10 extraction cases as non-extraction. This reflected a moderate performance with notable room for improvement in both sensitivity and specificity. The random forest classifier showed a marked improvement, correctly identifying 82 non-extraction and 73 extraction cases with only four misclassifications in total (three false positives and one false negative). This indicated enhanced precision and recall, especially in detecting extraction cases more accurately. The XGBoost classifier demonstrated the best classification accuracy among the three. It correctly predicted all 74 extraction cases and 82 out of 85 non-extraction cases, with just three false positives and zero false negatives. This perfect recall for the extraction class, combined with minimal error, highlighted XGBoost’s superior learning and generalization capabilities, offering significantly higher reliability and classification precision, making it more suitable for critical decision-making applications.

Close

Figure 1:

(a) Confusion matrix for a decision tree classifier showing classification performance. (b) Confusion matrix for a random forest classifier demonstrating classification accuracy. (c) Confusion matrix for an extreme gradient boost classifier showing optimal classification performance.

Export to PPT

The bar plot [Figure 2] illustrated the performance metrics of each model. The XGB classifier had the highest accuracy score of 0.985, followed by the random forest classifier and decision tree classifier with 0.975 and 0.881, respectively. The XGB classifier and random forest classifier had a similar precision score of 0.961, and the decision tree classifier had a 0.877 precision score. The XGB classifier had the highest recall score of 1, followed by the random forest and decision tree classifiers with 0.986 and 0.865, respectively. The F1 score was highest for the XGB classifier, followed by the random forest and decision tree classifier, being 0.98, 0.973, and 0.871, respectively. The above findings suggested that the XGB classifier algorithm outperformed the random forest classifier and decision tree classifier, achieving the highest recall score (1.00) and overall balanced performance.

Close

Figure 2:

Bar plot of performance metrics of the three models.

Export to PPT

[Figure 3] displayed the Pearson correlation coefficients among various cephalometric measurements, revealing the strength and direction of linear relationships. The matrix aided in identifying interdependent features that may influence diagnostic or treatment decisions in orthodontic analysis. Each cell in the heatmap showed a Pearson correlation coefficient ranging from −1 to +1, indicating the direction and strength of correlation between two variables. A value closer to +1 indicates a strong positive linear relationship, while values near 0 indicate weak or no correlation. SNA and SNB showed a strong positive correlation (0.69), as both measure cranial base angles relative to the maxilla and mandible, respectively. ANB and Witts appraisal are highly correlated (0.63), reflecting their shared role in assessing anteroposterior jaw discrepancy. UI-NA (deg) and UI-NA (mm) showed a strong correlation (0.5), indicating that changes in angular measurements often accompany proportional changes in linear distance. LI-NB (deg and mm) also exhibited a high correlation (0.51), consistent with observations in lower incisor analysis. Ashley’s and Nance’s analyses, which assess dental arch relationships, were significantly correlated (0.52), and both were highly correlated with Bolton’s ratio (Ashley’s: 0.53; Nance: 0.66), suggesting their mutual importance in evaluating dental proportionality. IMPA, which reflects incisor mandibular plane angle, showed moderate correlations with both UI-NA (deg: 0.28) and LI-NB (deg: 0.48), suggesting its influence on incisor inclination.

Close

Figure 3:

Heatmap of Pearson correlation matrix for cephalometric and dental variables.

Export to PPT

The performance error was also calculated for all three algorithm types between training and testing. The learning curves for the decision tree, random forest, and XGBoost classifiers revealed distinct trends in how each model generalizes with increasing training data. The decision tree classifier [Figure 4a] showed consistently zero error on the training set, indicative of overfitting. The test set error remained relatively high and fluctuated with different training set sizes, suggesting poor generalization and high variance. While there was a slight downward trend in test error, the lack of convergence indicates limitations in model complexity and stability. In contrast, the random forest classifier [Figure 4b] demonstrated better generalization. The training error was zero, similar to the decision tree, but the test error was significantly lower and stabilized early as more training data were added. The minimal variance in test error across all training sizes reflects the ensemble model’s robustness and ability to prevent overfitting by aggregating multiple decision trees. The XGBoost classifier [Figure 4c] further improved on this trend, showing the lowest and most stable test error across the entire range of training set sizes. Like the others, it also achieved zero training error, but unlike the decision tree, its test error consistently decreases and levels off with more data. This indicated that XGBoost achieved both high accuracy and excellent generalization due to its use of boosting, regularization, and optimization techniques. Overall, while the decision tree struggles with overfitting and generalization, Random Forest offers more stable performance, and XGBoost clearly outperforms both in terms of predictive consistency and learning efficiency.

Close

Figure 4:

(a) Learning curve showing misclassification error for the decision tree classifier across varying training set sizes. (b) Learning curve showing misclassification error for the random forest classifier across varying training set sizes. (c) Learning curve showing misclassification error for the extreme gradient boost classifier across increasing training set sizes.

Export to PPT

The receiver operating characteristic (ROC) curves illustrated the classification performance of three different ML models: Each curve plots the true-positive rate (sensitivity) against the false-positive rate, and the area under the curve (AUC) value that quantifies overall model performance. The decision tree classifier [Figure 5a] showed a moderately strong performance with an AUC of 0.88, indicating its ability to distinguish between classes, though with a relatively higher false-positive rate. In contrast, the random forest classifier [Figure 5b] achieved a significantly higher AUC of 0.97, reflecting improved sensitivity and specificity due to its ensemble-based architecture. The XGBoost classifier [Figure 5c] outperformed both, with an AUC of 0.98, closely approaching an ideal classifier. This highlights its superior learning capacity, optimization, and regularization techniques, which contribute to minimal misclassifications. Overall, while the decision tree provides a baseline level of accuracy, the random forest and especially the XGBoost classifiers deliver enhanced and more reliable performance in terms of true positive recognition and reduced false alarms.

Close

Figure 5:

(a) The receiver operating characteristic (ROC) curves illustrating the classification performance of the decision tree classifier. (b) The ROC curves illustrate the classification performance of the random forest classifier. (c) The ROC curves illustrate the classification performance of the extreme gradient boost classifier. AUC: Area Under Curve.

Export to PPT

DISCUSSION

ML has been applied to categorization and decision-making difficulties in numerous research.^[13-16] One way to approach the decision to extract is a sort of categorization problem. Over the past several years, the number of scoping reviews about “AI and ML algorithms and their application in orthodontics” that have been published in orthodontic literature has steadily increased.^[17,18] Numerous studies demonstrate how well various types of algorithms may be applied to segment, automatically detect, analyze, and extract characteristics from images to improve orthodontic diagnosis.^[10,19-25]

AI models are increasingly transforming clinical decision-making by providing data-driven support that enhances diagnostic accuracy, optimizes treatment planning, and improves patient outcomes. Their integration into clinical workflows allows for real-time interpretation of complex datasets, leading to more precise and customized treatment plans and acceptable outcomes. AI also plays a pivotal role in risk stratification and prognosis prediction in orthognathic jaw surgery cases. The direct clinical impact of AI lies in its ability to augment human expertise, reduce diagnostic and therapeutic variability, and facilitate more informed, consistent, and efficient decision-making. As these technologies continue to evolve, their responsible implementation holds the promise of advancing precision and elevating standards of care.

In general, extractions are done to relieve crowding, lessen the prominence of dental arches, and adjust the disparity in the anteroposterior connections between arches. In addition, spaces created by tooth extraction enable the correction of vertical dimension disparities. Moreover, extraction spaces can be used to rectify tooth size discrepancies (congenital agomphosis or Bolton index abnormity), width discrepancies across arches (cross-bite or scissor bite of posterior teeth), and so on.^[19] Malocclusions were considered to have “improved” if the PAR index values were lower after therapy than they were before treatment by more than 30%.^[26] Choosing between extraction and non-extraction therapy can be one of the most difficult daily decisions in a clinical orthodontic practice, particularly in situations that are considered “orderline.”

To create a highly precise simulation performance model and ensure a successful course of therapy, the present study employed more significant parameters. The study utilizes widely recognized ML algorithms; however, its innovation resides in the clinical design architecture that integrates the largest multivariable cephalometric-study model dataset (n = 800) with clinically validated parameters,^[19] stringent noise filtering, the inclusion of solely clinically successful outcomes, and an extensive performance comparison and generalization analysis (ROC, AUC curve, and confusion matrix). All of these things combined make a model for making irreversible extraction decisions that are more clinically dependable and easier to understand. This suggests that the developed population-specific cohort-based model, which makes use of three distinct ML algorithms, can be a useful, evidence-based tool for determining which treatment option is preferable when it comes to extractions versus arch expansion.^[10,27,28]

In borderline orthodontic cases where the decision between extraction and non-extraction is not clear-cut, AI models assist by analyzing a combination of clinical features such as crowding severity, incisor inclination, arch space, and facial profile.^[28] These models are trained on large datasets that include past treatment decisions and outcomes, allowing them to recognize patterns and predict which treatment path is more likely to succeed. Rather than giving a simple yes-or-no answer, the AI often provides a probability or confidence level for each option, helping the clinician weigh the risks and benefits. In some systems, the AI also highlights which factors most influenced the recommendation, offering transparency and supporting the clinician’s final judgment. This helps reduce uncertainty and supports more consistent and informed decision-making in challenging cases.

The strength of this model lies in the fact that it was created from a pool of 800 patients treated by different attending orthodontists using a variety of treatment modalities and has varying treatment philosophies, reducing the risk of institutional bias. Only cases with post-treatment outcomes classified as objectively successful were used for model labeling. The prepared model was successful in judging whether an orthodontic treatment for a particular case needs extraction or not, since this study had considered the parameters in both anteroposterior and vertical dimensions, and both hard and soft-tissue structures along with study models. The demographic variation – such as age, sex, ethnicity, growth status, and craniofacial features – can impact treatment choices; however, the current dataset was limited to people aged 13–35 years with comparable development trends.

Although this study was successful in diagnosing cases requiring extraction or not using three different algorithm classifiers, Random Forest, XGB classifiers, and a decision tree classifier, it reduced subjectivity and variability. With high-performance metrics like an AUC of 0.98 – can greatly aid clinical decision-making, they do not account for the ethical dimensions involved in irreversible procedures such as tooth extractions. Relying too heavily on ML outputs without any critical appraisal, especially in borderline cases without fully considering the broader clinical, psychological, and long-term implications for the patient, could undermine the individualized, patient-centered approach that is foundational to ethical practice. An openly accessible, simplified inference-only version of the trained model, along with anonymized example data for independent benchmarking by researchers and clinicians, is currently under development. This open validation will ensure transparency, reproducibility, and clinical credibility.

CONCLUSION

The following conclusions could be drawn from this research:

1. The constructed AI model in this study was able to decide the orthodontic cases requiring extraction or not with an accuracy of 98%

2. The XGB classifier system had the highest accuracy, recall, and precision scores, outperforming the random forest and decision tree classifier algorithms. However, the random forest classifier demonstrated a predictive performance comparable to the XGBoost classifier, which cannot be underestimated.

3. The use of this model could be one of the best alternatives available to simplify the process of decision-making, as it is less operator-dependent and has accurate predictions.