References

Vincent C, Taylor-Adams S, Stanhope N. Framework for analysing risk and safety in clinical medicine. BMJ. 1998; 316:1154-1157
Bouloux GF, Steed MB, Perciaccante VJ. Complications of third molar surgery. Oral Maxillofac Surg Clin N Am. 2007; 19:117-128
Gerry S, Bonnici T, Birks J Early warning scores for detecting deterioration in adult hospital patients: systematic review and critical appraisal of methodology. BMJ. 2020; 369
Vranckx M. Third Molar Management: Eruption, Removal and Associated Risks.Leuven: KU Leuven; 2020
Mercier P, Precious D. Risks and benefits of removal of impacted third molars. A critical review of the literature. Int J Oral Maxillofac Surg. 1992; 21:17-27
Bruce RA, Frederickson GC, Small GS. Age of patients and morbidity associated with mandibular third molar surgery. J Am Dent Assoc. 1980; 101:240-245
Phillips C, White RP, Shugars DA, Zhou X. Risk factors associated with prolonged recovery and delayed healing after third molar surgery. J Oral Maxillofac Surg. 2003; 61:1436-1448
Renton T, Smeeton N, McGurk M. Factors predictive of difficulty of mandibular third molar surgery. Br Dent J. 2001; 190:607-610
Gbotolorun OM, Arotiba GT, Ladeinde AL. Assessment of factors associated with surgical difficulty in impacted mandibular third molar extraction. J Oral Maxillofac Surg. 2007; 65:1977-1983
Pell G, Gregory G. Impacted mandibular third molars: classifications and modified tech-nique for removal. Dent Digest. 1933; 2:23-54
Rood JP, Shehab BA. The radiological prediction of inferior alveolar nerve injury during third molar surgery. Br J Oral Maxillofac Surg. 1990; 28:20-25
Vranckx M, Fieuws S, Jacobs R, Politis C. Prophylactic vs. symptomatic third molar re-moval: effects on patient postoperative morbidity. J Evid Based Dent Pract. 2021; 21
Van der Cruyssen F, Peeters F, Gill T Signs and symptoms, quality of life and psychosocial data in 1331 post-traumatic trigeminal neuropathy patients seen in two tertiary referral centres in two countries. J Oral Rehabil. 2020; 47:1212-1221
Robbins J, Smalley KR, Ray P, Ali K. Does the addition of cone-beam CT to panoral imaging reduce inferior dental nerve injuries resulting from third molar surgery? A systematic review. BMC Oral Health. 2022; 22
Korkmaz YT, Kayıpmaz S, Senel FC Does additional cone beam computed tomography decrease the risk of inferior alveolar nerve injury in high-risk cases undergoing third molar surgery? Does CBCT decrease the risk of IAN injury?. Int J Oral Maxillofac Surg. 2017; 46:628-635
Ghaeminia H, Burger M, Galjart M Richtlijn derde molaar. 2018;
Royal College of Surgeons Faculty of Dental Surgery. Parameters of care for patients undergoing mandibular third molar surgery. 2020. http://www.rcseng.ac.uk/-/media/files/rcs/fds/guidelines/3rd-molar-guidelines--april-2021.pdf (accessed December 2023)
Feragalli B, Rampado O, Abate C Cone beam computed tomography for dental and maxillofacial imaging: technique improvement and low-dose protocols. Radiol Med. 2017; 122:581-588
Jacobs R. Dental cone beam CT and its justified use in oral health care. JBR-BTR. 2011; 94:254-265
Jaju PP, Jaju SP. Cone-beam computed tomography: time to move from ALARA to ALADA. Imaging Sci Dent. 2015; 45
Erickson BJ. Basic artificial intelligence techniques: machine learning and deep learning. Radiol Clin North Am. 2021; 59:933-940
Vranckx M, Fieuws S, Jacobs R, Politis C. Prophylactic vs. symptomatic third molar re-moval: effects on patient postoperative morbidity. J Evid Based Dent Pract. 2021; 21
Hur SH, Lee EY, Kim MK Machine learning to predict distal caries in mandibular second molars associated with impacted third molars. Sci Rep. 2021; 11
Liu MQ, Xu ZN, Mao WY Deep learning-based evaluation of the relationship between mandibular third molar and mandibular canal on CBCT. Clin Oral Invest. 2022; 26:981-991
Ariji Y, Mori M, Fukuda M Automatic visualization of the mandibular canal in relation to an impacted mandibular third molar on panoramic radiographs using deep learning segmentation and transfer learning techniques. Oral Surg Oral Med Oral Pathol Oral Radiol. 2022;
Vranckx M, Ockerman A, Coucke W Radiographic prediction of mandibular third molar eruption and mandibular canal involvement based on angulation. Orthod Craniofac Res. 2019; 22:118-123
Vranckx M, Van Gerven A, Willems H Artificial intelligence (AI)-driven molar angulation measurements to predict third molar eruption on panoramic radiographs. Int J Environ Res Public Health. 2020; 17
Lee J, Park J, Moon SY, Lee K. Automated prediction of extraction difficulty and inferior alveolar nerve injury for mandibular third molar using a deep neural network. Appl Sci. 2022; 12
Yoo JH, Yeom HG, Shin W Deep learning based prediction of extraction difficulty for mandibular third molars. Sci Rep. 2021; 11
Nogueira-Reis F, Morgan N, Nomidis S Three-dimensional maxillary virtual patient creation by convolutional neural network-based segmentation on cone-beam computed tomography images. Clin Oral Investig. 2023; 27:1133-1141
Hu R, Zhang XY, Liu J Clinical application of cinematic rendering in maxillofacial fractures. Int J Oral Maxillofac Surg. 2022; 51:1562-1569
Recht MP, Dewey M, Dreyer K Integrating artificial intelligence into the clinical practice of radiology: challenges and recommendations. Eur Radiol. 2020; 30:3576-3584

The Use of Artificial Intelligence in Third Molar Surgery Risk Assessment

From Volume 51, Issue 1, January 2024 | Pages 28-33

Authors

Fréderic Van der Cruyssen

MD, DDS, PhD, Resident

Department of Oral and Maxillofacial Surgery, University Hospitals Leuven, Leuven, Belgium

Articles by Fréderic Van der Cruyssen

Email Fréderic Van der Cruyssen

Pieter-Jan Verhelst

MD, DDS, Resident

Department of Imaging and Pathology, OMFS-IMPATH Research Group, Faculty of Medicine, University Leuven, Leuven, Belgium

Articles by Pieter-Jan Verhelst

Reinhilde Jacobs

DDS, PhD, Professor

Department of Oral Health Sciences, KU Leuven and Department of Dentistry, University Hospitals Leuven, Leuven, Belgium. Department of Dental Medicine, Karolinska Institutet, Stockholm, Sweden

Articles by Reinhilde Jacobs

Abstract

Third molar removal complication rates can be as high as 30%. Risk assessment tools may lower these rates. Artificial intelligence (AI) driven prediction models are a promising approach to predict possible unfavourable outcomes and cone beam computed tomography imaging may play an important role. AI prediction models are showing excellent results in research settings. To be implemented in clinical practice they will need to overcome some robustness, security, liability, and practical issues. If they do, AI prediction models can be integrated in electronic patient record systems by alerting clinicians in case of an imminent unfavourable outcome so it can be avoided.

CPD/Clinical Relevance: Artificial intelligence-driven risk assessment tools will lower complications in third molar surgery.

Article

Safety is of paramount importance in our rapidly evolving healthcare system. Numerous medical risks occur every day, and to mitigate them, risk assessment strategies are increasingly being used.1 Several risks may arise during third molar removal, one of the most commonly performed dento-alveolar surgical procedures.2 These can be found in the areas before, during, and after surgery and may involve trigeminal nerve injuries, prolonged post-operative pain, infection, haemorrhage, oro-antral communications and fractures. To mitigate these risks and avoid unfavourable outcomes, a probability assessment of the likelihood of occurrence is crucial. Advanced statistical models are used for this, and with the rise of artificial intelligence (AI), powerful prediction tools are becoming available for clinical care. Ideally, risk assessment is performed automatically based on electronically registered patient data and is integrated in a warning system that alerts the clinician when a high probability of an unfavourable outcome is present. This allows the clinician to act appropriately.

A great example is the automatic early warning scores for hospitalized patients, which allows doctors to detect deterioration of their patients in real time.3 Based on certain patient parameters that are registered in the hospital, such as blood pressure, respiratory rate and heart rate, the electronic health record system calculates an early warning score and prompts the nurse to contact the treating physician when a high score is registered on the system. These types of systems are being developed in different branches of healthcare and are a clear example of the integration of data science in clinical practice. The current article discusses the importance of risk assessment in third molar management, the role of AI systems and their integration into clinical practice.

Importance of risk assessment in third molar surgery

The complication rate after wisdom tooth extractions varies, but can be as high as 30%, according to some, resulting in a high return of patients to the practitioner's office after surgery.2 Vranckx et al reported a revisitation rate of 20% in a group of 6010 patients who had one or more wisdom teeth removed.4

Many authors have already tried to determine the degree of difficulty of a third molar extraction and therefore the risk associated with this procedure.4,5,6,7,8,9 The primary factors contributing to this are determined pre-operatively, such as age, ethnicity, body mass index, gag reflex, and presence of anxiety. Dental factors such as crown and root morphology, horizontal and vertical position, bone density and proximity of the inferior alveolar nerve (IAN) also play an important role. The Pell and Gregory classification on depth of impaction and the Rood and Shehab classification on predicting IAN injury are the best known risk classification systems for injury to this nerve in relation to third molar surgery.10,11

The occurrence of IAN injury is one of the most feared complications. This occurs in about 2% of all third molar extractions, but can vary greatly depending on the location of the wisdom tooth and increases with age.12 Once nerve damage has occurred, it can be temporary or permanent, with a significant impact on the patient's quality of life.13 To prevent this nerve damage, a pre-operative panoramic radiograph is classically used. Depending on the position of the wisdom tooth with respect to the mandibular canal, an additional cone beam computed tomograph (CBCT) can be taken. Until now, it could not be demonstrated that CBCT reduces the risk of IAN damage.14 However, we must realize that many studies had too little power and methodological bias to be able to draw this conclusion. Others have shown the relevance of CBCT in specific cases because it might change surgical management.15 Some guidelines on wisdom tooth extraction recommend CBCT in cases of overlap of the mandibular canal on panoramic imaging.16,17 Moreover, low radiation dose protocols of CBCT are continually improving.18,19 Thus, in light of the as low as diagnostically acceptable (ALADA) principle, a new discussion is possible about the use of three-dimensional CBCT, which obviously contains more diagnostic information compared to a two-dimensional orthopantomogram (OPT).20 These developments are driving a resurgence in risk prediction research in which artificial intelligence (AI) and machine learning (ML) developments are playing an important role.

Artificial intelligence and machine learning

AI, ML and deep learning (DL) research is growing abundantly. These terms although closely related are not interchangeable. AI is an umbrella term for automating intellectual tasks normally performed by humans. ML and DL are some of the tools to achieve this goal. Machine learning tries to represent a dataset by means of an algorithm. In conventional statistics, this algorithm can easily be calculated, whereas in ML, the algorithm is generated by training a subset of the data. The variables that are being used to train the model or algorithm are called features. The most commonly used AI terminology is explained in Table 1.


Table 1. Commonly used machine learning and artificial intelligence terminology.
Machine learning term Epidemiology term
Attribute, feature, predictor or field Independent variable
Input and output Independent (exposure) and dependent (outcome) variables
Classifier estimator Model
Learner Model fitting algorithm
Dimensionality Number of explaining variables
Label Value of dependent variables, outcomes
Imbalanced data Data set in which some cases or risk categories occur much less frequently than others
Loss function Error measure
Data munging or data wrangling Data preparation

To develop the best ML algorithm, we can use different teaching methods, such as guiding it with examples (supervised learning), letting it figure things out on its own (unsupervised learning), using a mix of both (semi-supervised learning), or rewarding it for good decisions (reinforced method). In supervised models, the training data is labelled, meaning the features and outcome (target) are known. In unsupervised models the algorithm tries to detect patterns in the dataset without being informed about the target. Semi-supervised learning combines both methods because not all features have an associated target. This can be useful for very large datasets where labelling the whole dataset would be too time consuming. Finally, reinforcement learning focuses on improving a task by trial and error. Artificial neural networks and deep learning are ML methods that mimic biological neural networks. Several layers of nodes (resembling cell bodies) are connected (resembling dendrites and synapses). The nodes are then weighted against each other depending on their ability to provide the desired outcome.21

Current state of the art

Risk prediction using ML techniques is possible both in the pre-, peri- and post-operative setting for patients having wisdom tooth surgery (Figure 1). In the pre-surgery phase, our research group conducted a yet unpublished study using the same database by Vranckx et al.22 We were able to predict with 81% accuracy which patients would have severe pain after surgery. We used a ‘random forest’ (a collection of decision-making trees) and the ‘Boruta algorithm’ (a method to select the most important factors) to make these predictions. Also, unexpected significant predictive features were revealed using this method providing new insights in pain prediction. In a similar manner, Hur et al calculated the risk factors for developing distal caries on the second molar associated with third molar impaction. A random forest model outperformed the other ML methods and showed the importance of the examined risk factors in developing these caries, such as age, gender, location of contact point and Pell and Gregory classification.23

 

Figure 1. Possible areas of risk prediction machine learning applications in third molar surgery. A dynamic feedback model is exemplified. Dynamic feedback models in risk prediction machine learning are systems that learn to estimate potential risks or dangers by continuously updating their knowledge based on new information and adjusting their predictions accordingly. This helps them become better at predicting over time. The information that could be used to develop these models can be collected pre-operatively by means of patient-reported outcomes (questionnaires, surveys, history taking), clinical examination features and radiological findings. In the peri-operative setting combining vital parameters could be useful and postoperatively again patient-reported outcomes, pain scores, medication use, etc could be integrated in the model.

In the future, these models could be integrated into the electronic health record to aid the treating physician in consenting and informing the patient in advance, and to select the best management strategy.

In our dental practices, we rely heavily on oral imaging, a field that will see many machine-learning developments in the next years. In Figures 25, we show some examples of radiological developments related to wisdom tooth surgery. In a first step, researchers succeeded in using a convolutional neural network (CNN) to automatically detect and segment the mandibular canal and the third molar on panoramic images.24,25 An important step in the further development of risk prediction models. A convolutional neural network is a type of computer program designed to process and recognize patterns, such as images or sounds. It works similarly to how our brain recognizes things, by breaking down complex information into simpler parts and then combining them to understand the bigger picture. This makes CNNs particularly good at tasks like image recognition and analysis. Vranckx et al used automatic segmentation and subsequent angulation measurements of mandibular molars to predict whether the third molar had a chance for functional eruption. Using a linear regression model, they found that an angle of more than 27° between the longitudinal axes of the third and second molar could accurately predict whether the wisdom tooth would erupt. Moreover, they showed that with increased angulation, the probability of a close relationship between the third molar and IAN increased significantly.26,27 Lee et al trained a similar model to predict extraction difficulty and likelihood of IAN injury on panoramic radiographs with accuracies over 80%.28 Others used a similar approach to predict the Pederson difficulty score for extraction of third molars.29 Their CNN was able to accurately predict the ramus relationship, depth of impaction and angulation, and thus predict the final difficulty index.

Figure 2. Automatic segmentation of molars and subsequent calculation of angulation between the third molar and its neighboring teeth to estimate eruption potential (developed by Vranckx et al in collaboration with Relu, Leuven, Belgium).
Figure 3. Fully automatic 3D model segmentation using CBCT imaging (Relu, Leuven, Belgium).
Figure 4. Virtual patient creator by Relu (Leuven, Belgium). This cloud-based platform automatically segments dentomaxillofacial CBCT imaging datasets into customizable 3D models that can be used for virtual surgical planning and CADCAM applications.
Figure 5. An automatically created 3D model (Virtual patient creator, Relu, Leuven, Belgium) indicating the third molar and mandibular canal relationship.

Subsequent developments by Choi et al, led to the use of CBCT and panoramic radiographs to train a learning algorithm to predict on a panoramic whether the IAN is located buccally or lingually from the mandibular canal and whether there was true contact between the two structures. This is important because a lingual position and close contact bear a greater risk for nerve damage when removing that third molar. Here, the algorithm was superior to the experienced radiologists, with an accuracy to detect true contact of 72%. The question remains whether such accuracy is sufficient to estimate a potentially dangerous, but infrequent risk.

These models could also be used to enhance pre-operative visualization of important structures. Nogueira-Reis et al created reliable virtual patient models from CBCT images using automatic segmentation in an average time of 1.7 minutes.30 Moreover, the platform allowed selection of distinct anatomical structures, such as the airway, teeth, and maxillary sinuses, which should enable fast and accurate virtual surgical planning in the near future. However, it will remain difficult to demonstrate that these sophisticated visualization tools reduce the risk of complications in wisdom tooth surgery. Randomized studies in which the surgeon is offered only one type of imaging technique pre-operatively to assess risk is often considered unethical. Moreover, large groups are needed to demonstrate a statistically significant difference on outcomes using different imaging modalities because the incidence of nerve injury is low. Finally, it is also difficult to demonstrate a behavioural change in surgeons the moment they are shown the pre-operative imaging study. In other fields, however, it was recently demonstrated that true-to-life 3D cinematic rendered models are superior for detecting maxillofacial fractures compared to 2D CT images or volume rendered models (Figure 6).31 A similar approach could be adopted to study the use of enhanced 3D models in pre-operative surgical planning of wisdom tooth removal.

Figure 6. Comparison of a volume rendered model based on CBCT imaging (left) versus a cinematically rendered model based on the same CBCT dicom dataset (right).

Future perspectives and clinical implementation

There is a clear exponential growth of AI applications in dentistry and medicine. It is evident that these applications will soon change the way we work. However, there are some obstacles that need to be overcome before we can fully integrate these tools into our daily practice (Figure 7).32

Figure 7. Summary of obstacles that should be taken into account for successful implementation of AI and ML risk prediction models.

First of all, we should ensure that these risk prediction tools are robust. Many of the already-available tools are not validated on large populations and thus certain groups may be over or underrepresented causing bias in the algorithm. This may lead to conclusions made by the algorithm that are not applicable to the desired population. To properly evaluate this, we require transparent AI systems that are ethically vetted. These systems emerge on big data and can become even more accurate with feedback from new dynamic real-world data in which the algorithm keeps improving by its daily application. For this, continuous feedback and monitoring systems need to be in place.

This leads us to our second obstacle, which is security and privacy. Data protection and privacy issues are at the forefront of all IT applications in healthcare nowadays. If these systems are used in practice, and especially if a dynamic feedback model is used, patients and healthcare providers should be aware and consent on how this data is used and stored to improve a mostly commercial product. Furthermore, a robust security system is necessary to avoid data leaks in these mostly cloud-based products.

A third area of focus in implementing artificial intelligence in dental practices concerns how clinicians use these tools, and the implications for liability. It is essential to understand that risk prediction models serve as a supportive tool, and like any other tool, they must be used correctly. These AI systems should assist clinicians in making decisions rather than automatically generating responses.

This highlights the importance of clinicians maintaining responsibility for the outcomes of decisions informed by AI systems. To ensure proper use and understanding of AI technologies, healthcare providers interacting with such systems should have at least a basic knowledge of their functions and underlying principles.

To facilitate this understanding, it is crucial to integrate AI technologies into dental and medical education programmes. This will provide healthcare professionals with the necessary foundation to effectively use AI tools while ensuring they remain accountable for their decisions. Additionally, addressing ethical and legal concerns related to AI implementation will promote transparency and help safeguard patient interests.

Finally, after these robustness, security and liability issues are addressed a very practical problem becomes clear. Many of the risk assessment applications available today in third molar management are trained to solve very specific tasks requiring very specific input in a research setting. The transition to a clinical application poses a challenge as clinical reality is a bit messier. Clinicians have multiple questions: should I remove the third molar and if so, how big is the chance of nerve injury or a post-operative infection? This requires a combination of multiple AI models. Furthermore, these should be integrated in electronic patient record systems in a user-friendly way to make sure they are easily accessible.

Conclusion

The complication rates in third molar surgery promote the development and use of risk prediction systems to allow for adequate case selection, safe surgical planning, and individualized post-operative care. AI systems are rapidly propelling these prediction systems forward in very promising way. Owing to technical improvements, declining radiation doses and a high diagnostic yield, CBCT imaging plays a crucial role in this process. Accurate risk-prediction tools are already available in a research setting. The widespread adoption of these tools in clinical practice is on the horizon, but needs to overcome robustness, security, liability and practical issues.