Clinical implementation of AI systems in orthodontics

Drs. Aaron Glick, John Garza, Nisha Manila, and Jennifer Chang discuss the pros and cons of AI in orthodontic and general dental practice

Introduction

Although artificial intelligence (AI) has been integrated in daily life, most are not aware that it has become an integral part of many aspects of modern life over the past decades given the increased computing power, digitization of data, advances in algorithm, and improvements in accuracy. In fact, when a group of Americans were surveyed, over half were not able to identify that AI powered wearable fitness trackers, automated chatbots, product recommendations, security cameras that identify unrecognized people, music playlist recommendations, and email services identifying spam messages.¹ In dentistry, some of these same AI systems, such as chatbots, have been integrated for general office support and software automations. Other AI systems that are more dental-specific have also been used for approximately a decade. These include systems that automatically identify dental landmarks in CBCT,² maxillary/mandibular segmentation from radiographic images,³ and digital reconstruction of tooth surfaces.⁴

Some examples of more current AI systems in dentistry are decision support systems to aid clinicians in the diagnosis of caries and periodontal conditions, AI-enabled robotic surgeries, and creation of AI-generated dental restorations/appliances.⁵ Since AI is a tool, it can be used in multiple applications within the dental field; as the algorithms advance in accessibility and accuracy, the amount of products with embedded AI will likely increase (Table 1). Therefore, it is important as a dental practitioner to understand the tools used, particularly if they can have a direct or indirect effect on patients.

How AI works

AI is a broad term that generally refers to the ability for computers to perform tasks that typically require human intelligence (Figure 1); for example, viewing a dental X-ray and identifying interproximal caries. Machine learning (ML) refers to the mathematical approach to accomplish this perceived human intelligence. These algorithms are programmed to take information and “learn” from these datasets. Since there are alternative methods for artificial intelligence, machine learning is considered a subset of the approaches to achieve artificial intelligence. Multiple types of algorithms can be programmed by the user. In one approach, a convolutional neural network (CNN) can feed a large dataset to identify the features that are most salient in a classification task. For instance, with multiple pictures of apples or cats, the neural network would be a mathematical model used to make the binary decision on if the currently presented image should be classified as an apple or cat. The CNN approach is typically selected for image operations since this algorithm is optimized to process images based on the extraction of spatial features.

Depending on the task to be performed, there are multiple types of models that are specific to optimizing for the type of data or output intended. For instance, chatbots require a text-based approach with grammatical rules built into sequences of words. Therefore, language-specific models would be used for text-based tasks, typically called natural language processing (NLP).

Further into the black box

While most dentists today don’t consider building custom AI products, there are products that will democratize the use of AI through no-coding technologies.⁶ Whether coding the algorithms or not, understanding the basics of how the data is used to derive a clinical decision/outcome is important as dentists are increasingly interacting with AI-enabled systems and machines. When software manufacturers train an AI model, they require a large amount of data to process. In general, the full dataset is split into 3 distinct groups: 1) training set, 2) validation set, and 3) testing set (Figure 1). In a supervised model, a team of clinicians will label the full dataset. For instance, if the classification task is to automatically identify caries in bitewing radiographs, the clinician would identify areas of the radiographs with caries. These labeled radiographs serve as the “ground truth” for the algorithm. The bulk of the data is used in the training set to optimize the weights of the neural network identifying salient features of the categorized dataset. The model is then further optimized by using a validation dataset, where additional processes can improve the accuracy. The final step is to understand how accurate the fully trained AI model can perform on the similar dataset. Therefore, the testing dataset is then categorized by the trained model to provide an accuracy typically reported as a percentage of accurately categorized responses and Receiver Operating Characteristics (ROC) curve/area under the curve (AUC). The ROC curve and AUC offers additional details similar to the true positive rate and false positive rate. Ultimately, the large dataset is trained, and the same isolated dataset is used to assess the accuracy in a non-clinical setting.

When building these algorithms, there are some additional key considerations as to how the algorithm uses data. In an oversimplification, an algorithm can be delineated as 1) supervised/unsupervised or 2) static/continuous learning. Supervised learning is based on a clinician providing the algorithm with a ground truth for labelled data, similar to the previous example. Unsupervised learning is where the model learns without explicitly provided information to classify and can be accomplished through patterns within the data. Static learning algorithms are the traditional type of AI. However, some current algorithms are designed to update and improve themselves based on additional data acquired during use.

AI in orthodontics

AI is changing our daily lives and approaches to dentistry in general. In orthodontics, AI is being used for the following:

1. More rapid diagnosis through facial, radiographic, photographic, and intraoral scanning
2. Treatment planning for surgery/outcome prediction
3. Treatment progression through remote monitoring and treatment progression

AI is being used clinically as assistive technology that can assess various forms of diagnostic data to provide identification and decision-making support for orthodontics. For instance, automatic cephalometric analysis to identify anatomical landmarks in CBCT radiographs has been shown in one study to be up to 92% accurate.⁷ Additionally, automatic tooth position detection and tooth identification have been shown to have accuracies above 88%.⁸ Identification and automatic digital labelling can allow for a more rapid assessment and potential diagnosis. Further aiding in the decision-making process is determination of dental age based on radiographic indicators that has shown to be highly accurate at 94%.⁹ Other examples of a decision support system that could be used in orthodontics is an early-stage detection of osteoarthritis AI system trained using multiple patient-specific factors such as CBCT, biomolecular, and clinical features.¹⁰ Currently, AI analyses have been used to detect facial midline and asymmetries, yet has not shown to improve accuracies over traditional methods.¹¹

Treatment planning in orthodontics can be aided by AI through decisions to extract, recommendations for orthognathic surgery, dentofacial complex predictions of outcome, and patient satisfaction in aligner therapy. In all cases, a well-functioning system can improve confidence for early-stage orthodontists, reduce overall treatment time, improve esthetic outcomes, and provide an improved patient experience. For instance, one study used patient, radiographic, and clinical data to construct an AI model to predict recommendations for extraction patterns and anchorage patterns.¹² Predictive algorithms have also been developed to predict the posttreatment facial changes¹³ and patient experience with clear aligner therapy with indices of pain and anxiety.¹⁴ Other applications include prediction of C. difficile infections for major surgeries, selection of headgear, selection of fixed appliances, and optimizing the selection of patients for removable appliances.¹⁵

AI is being used during and posttreatment in orthodontics. Remote patient monitoring can utilize AI to track 3D tooth movement during treatment with enough fidelity to reconstruct full 3D digital models.¹⁶ Use of remote patient monitoring can be done with the patient using their smartphone and can lead to increased patient compliance.¹⁷ In finalizing successful treatment, outcomes can be assessed using facial attractiveness models.¹⁸

A light in the black box

AI presents an exciting tool to simulate human intelligence to help dentists achieve a higher quality of life, better decision-making capability, and fewer postoperative complications.¹⁹ However, despite these advances, AI in dentistry is in its infancy. On the horizon are more autonomous systems that do not require human input at the clinical stage within the AI algorithm. While a good understanding of the basics of AI is useful, the actual processes of most algorithms and datasets are not transparent to clinicians. Accuracy, trust, data stewardship, and equity/sustainability are concerns that clinicians and AI software manufacturers need to address to ensure patient safety, clinical success, and safeguarding from legal complications.

Accuracy

The accuracy of an AI system is based on a predefined gold standard. Most clinical decision support tools are supervised algorithms and labeled by clinical dentists without histologic confirmation. Since the amount of data that is required to train the models is so large, it is impractical to build highly accurate datasets that are verified with high-quality standards. In fact, the sensitivity for dentists to detect interproximal caries in bitewings in dentin is low (.36),²⁰ and provider agreement also is low (.4 kappa).²¹ Similar findings have been seen in other fields such as in periodontal conditions yet are task specific.²² While the quality of the dataset will likely not be transparent, the process for labeling the data as an important factor in the accuracy of the system could be stated by the manufacturer. Another issue that can affect AI is the imbalance of certain conditions like rare pathology that limit the number of instances for the algorithm to learn.

What is the clinical accuracy of these AI models? Is the trained dataset similar to your patient population or data in your EHR? Most report accuracies on the original dataset that was used, yet the true accuracies with different datasets or after clinicians have interacted with the AI system remain limited, and independent testing remains even more limited. In fact, some studies have shown that when AI presents incorrect information, clinicians are more likely to be misled by this recommendation which could result in poor patient outcomes.^23,24 Should incorrect guidance be provided by the AI system and patient harm occurs, the current legal climate would hold the clinician liable.²⁵ Additionally, when the AI system is accurate, there are concerns of other biases and trust.

In all cases, a well-functioning system can improve confidence for early-stage orthodontists, reduce overall treatment time, improve esthetic outcomes, and provide an improved patient experience.

Trust

A human-computer interaction exists between the dentist and AI system that can affect the perception of clinical information and ultimately treatment performance and decision. Trust is a well-studied concept that is affected by many factors with transparency, reliability, and tangibility as some of the commonly cited values that apply most to human-computer interaction.²⁶ Transparency is inherent in trust, and AI systems can be built to be explainable. While most traditional forms of AI suffer from the “black box problem,” some offer detection strategies through computer vision,²⁷ and others provide embeddings through natural language processing.²⁸ Reliability offers a perception that the algorithm is making similar decisions in seemingly similar circumstances. For continuously learning algorithms, there is a potential risk that adaptive learning could introduce new errors and reduce reliable performance.^29,30 Tangibility refers to the physical or virtual presence of the AI system and when it appears more real or anthropomorphic, which improves trust.³¹

Data stewardship

Machine learning sits at the intersection of big data and software algorithms. Given the need for large datasets to train models, the data stewardship in regulated industries, especially dentistry, is important. Dental offices handle private data on patients within their electronic health record system such as financial, demographic, disease state information, and other sensitive clinical information. With the digitization of records and equipment, the dental industry is a prime target for hackers and also lacks adequate knowledge about the prevention of these crimes.³² In fact, Henry Schein, a large distribution and software company with more than 23,000 employees and 1 million customers globally, halted operations due to a cyber security incident.³³ Large, consolidated datasets with patient data are vulnerable to attacks. Additionally, generative AI that ingests data from multiple sources feeding back into the algorithm can pose a risk if patient data is used. For instance, divulging patient data on a generative AI platform to write patient notes more efficiently could put the dentist at risk of breaching HIPAA with patient data being used to continuously train the model. For example, Sarah Silverman is suing openAI and Meta over violations of her copyrighted book that surfaced when these platforms regurgitated substantial elements of her copyrighted materials.³⁴

Equitability/sustainability

The equitable treatment of patients is an ethical value upheld by the American Dental Association where all treatment decisions should be determined on the same basis as all other patients.³⁵ As AI systems are built with human inputs, biases can be encoded into the model.³⁶ In one example of this, Optum, a health service company, created an AI program to spot high-risk patients who should receive medical follow-up care. However, the follow-up care disproportionately identified white patients with 82% white and 18% black patients identified compared to an independent analysis finding that 53% white and 46% black patients should have been identified based on medical need. Before this error could be detected, the program had already been applied to 100 million patients.³⁷ Additionally, the hardware that powers AI requires high energy and a large local ecological footprint.³⁸ In fact, GPT3 took 1.287 gigawatt hours of electricity and 700,000 liters of clean freshwater.³⁹

Implementation of AI systems

As manufacturers continue to release new and update current AI-embedded technologies, dentists should be aware of how these tools operate and can potentially change patient care. Many dentists are likely to already implement AI-embedded technologies without fully realizing it since AI is commonly used as a tool to accomplish a task or optimize tasks, and most backend systems are not transparent to the user.

In order to implement new technologies with AI into a dental office, technology acceptance and continual quality assurance should be a consideration. The primary factors for technology acceptance are perceived usefulness and perceived ease of use.⁴⁰ Therefore, removing barriers to trust within the human-computer interaction can improve usage. Training with technology in a safe environment can build confidence and improve ease of use. Additionally, having a plan for continuous quality assurance with the use of AI can aid in patient safety given that the systems can change with updates or continuously learning models. AI behavior should be periodically monitored, and a set of well-defined metrics are needed to guide regularly scheduled testing.⁴¹

The U.S. government also functions to protect patients and dental offices’ implementation of AI through the FDA clearance process of software as a medical device. FDA has become increasingly more stringent on this type of software that predicts and identifies medical/dental conditions.⁴² As dentists expand their appetite for additional functionality of these software systems, there needs to be a balance between the regulatory burden and the ability to innovate.

Exciting future of AI in dentistry

With the digitization of dental records and radiographs, dentistry is an ideal profession to harness AI. Current applications of AI are already being implemented to change the workflow of dental offices and patient care. Despite risks of inaccuracies, human-computer mistrust, data security concerns, and inequities in applications, it is easy to imagine overcoming these obstacles as the perceived future benefits outweigh the risks.⁴³ Healthcare in general is expected to see an increase in venture investments as the value of AI in health is expected to grow from $4.9B in 2020 to $45.2B in 2026. Dentistry, however, is lacking in the amount and accuracy of data.⁴⁴ To accelerate the advancement of AI, manufacturers need access to large, structured datasets that contain valid and reliable data. In fact, each of our practices holds valuable data that has the potential to improve the field.