Applications of artificial intelligence and machine learning in orthodontics

INTRODUCTION

Artificial intelligence (AI) is a subfield of computer science concerned with developing computers and programs that have the ability to perceive information, reason, and ultimately convert that information into intelligent actions.^[1-3] AI as a science is very broad and encompasses various fields, including reasoning, natural language processing, planning, and machine learning (ML).^[4] At present, ML is the most commonly used AI application in the medical and dental fields.

Work in AI started back in 1943,^[5] but it was not until 1956 that the term “artificial intelligence” was first used during a conference held at Dartmouth College.^[6] A few years later, the term “machine learning” was officially applied to a checkers-playing program, considered one of the first successful self-learning tools.^[7] Drawing from other fields such as statistics, mathematics, physics, biology, neuroscience, and psychology,^[8-11] AI and ML progressed quickly.

One of the most important aspects of any intelligent system is learning. Learning is the process of improving performance or behavior by practice and experience.^[12] Similarly, ML is concerned with making machines and computers capable of learning from previous experiences, data, or examples. By utilizing a mixture of statistical and probabilistic tools, machines can learn from previous examples and improve their actions when new data are introduced. This could be in the form of predictions, identifying new patterns or classifying new data.^[8] It is important to note that ML is not intended to mimic human behavior. Instead, it supplements human intelligence by erforming tasks that are beyond human capabilities.^[13] This is what makes ML superior to the rule-based expert systems (ESs) that were used in the past.

ESs are considered among the earliest applications of AI. As the name implies, the knowledge of a specific field is transferred from humans to computers, allowing people to consult the computer.^[14] In other words, ESs act as consultants that can process input information and provide solutions based on if-then rules. ESs have been used widely for diagnosis and treatment planning in medicine,^[15] dentistry,^[16] and orthodontics.^[17] ESs also facilitate the transfer of knowledge to different people in different places. However, rule-based ESs are limited to information available at the time that the system is developed. Continuous updates are required to ensure that the information is correct and current. Due to the availability of more advanced technologies, such as ML, it is now possible to overcome the limitations associated with rule-based ES.

Most algorithms used in ML are also being used in data mining. The difference lies in the algorithm’s goal. If the goal is to optimize decisions, then the algorithms are applied to large historical data sets to look for new patterns or relationships.^[18,19] This process is called data mining. For example, data mining can help clinical practitioners find valuable information within existing patient records. Using this new information, practitioners can optimize future decisions, improve their daily practice, and increase the quality of care. On the other hand, if the goal is to make predictions, then ML should be applied. The clinical practitioner uses available data about a certain disease to train the machine to make predictions about the diagnosis or prognosis of patients that have never been seen before. Importantly, ML predictive models have proven to be more accurate than statistical models.^[20] The aim of the present narrative review was two- fold: (1) To introduce the various types of ML and (2) show orthodontists how ML has been and is currently being applied. The literature was systematically searched using the MEDLINE (through PubMed) and ProQuest databases, covering both the published and unpublished literature reported in English. The studies covered are comprehensive with respect to orthodontic applications only.

TYPES OF ML

ML algorithms are divided into three main categories^[10] [Figure 1] based on the nature of learning and the desired outcome of the algorithm:

Close

Figure 1:

Types of machine learning.

Export to PPT

MAJOR ML ALGORITHMS AND DENTISTRY

There are several ML algorithms that have been used in the dental fields. Depending on the goal, the type, and amount of data, different algorithms can be used. For example, if a practitioner wants to distinguish between patients who need treatment and those who do not, he/she probably would need to use a classification algorithm (e.g., support vector machine, naïve Bayes, etc.). However, if there are many variables and a large amount of data, an algorithm like neural networks is better suited because it can handle noisy data and perform predictions even if the relationships between variables are non-linear.

Interestingly, almost all ML algorithms applied in orthodontics have used the supervised learning method [Table 1]. Most applications have sought to automate clinical procedures that perform or facilitate diagnosis and treatment planning. These applications require training with data that have a known and desired outcome.

AI AND ORTHODONTICS

Dentistry in general and orthodontic specifically has applied AI to solve many different problems. Early attempts to use AI in dentistry and orthodontics were in the form of knowledge- based ES. These systems were mainly aimed at helping non-specialist dentists develop diagnoses and treatment plans.^[22-25] These ESs were useful in countries like England, where hospital-based orthodontists had long waiting lists and were seeing more patients than their counterparts in Europe and the US. Due to the decline in the incidence of caries that occurred at that time, dentists treated the more straightforward cases identified by the ES and referred the more complex cases to orthodontists. However, these systems were limited because they only had been introduced to simple cases (i.e., they could not function well with new cases not already stored in the system). At present, general dentists have more advanced ML systems available to them that can diagnose a broader range of orthodontic cases and determine treatment needs.^[26] Several advanced systems have been developed to help orthodontists diagnose and treatment plan and evaluate treatment outcomes and growth.

ML FOR DIAGNOSIS AND ORTHODONTIC TREATMENT PLANNING

One of the dilemmas during treatment planning is deciding whether or not to extract, with substantial variability between orthodontists’ decisions.^[27] This has led to the development of several decision support systems that reduce the subjectivity of making decisions. Artificial neural networks (ANNs)^[28-30] have been used to develop such systems, and they were shown to be successful at predicting the extraction decision 80%^[28] of the time in one study and 93%^[29,30] of the time in two other studies. Prediction of the detailed extraction patterns (i.e., which teeth needed to be extracted) was also shown to be possible 84%^[29] of the time in one study and 83%^[30] of the time in another study. Recently, a paper used ANN to identify anchorage requirements in cases that were determined by the system to need extractions and it was accurate 83% of the time.^[30]

X-ray analysis, an integral part of diagnosis and treatment planning, has also benefited from ML. One of the most important applications of ML in orthodontics was the automation of landmark detections. A recent systematic review reported 5–15% better accuracy of landmark detection with ML than traditional methods.^[31] ML was also used to automate diagnostics directly from cephalograms, including the sagittal relationships between the maxilla and mandible, as well as normal and abnormal posterior-anterior facial heights ratios, overbite, and overjet.^[32]

Automation of X-rays analysis has also been extended to hand and wrist radiographs for estimating skeletal age. Determining the growth status of patients is essential for deciding whether or not to utilize growth during treatment.^[33] A ML system applied to a sample of 360 images showed an average difference of 0.39 years between its estimate and skeletal age estimated by two expert radiologists.^[34] Another study, using a larger sample of 1100 images, reported an average difference of 0.60 years when compared to two experienced radiologists.^[35] One study comparing the performance of different algorithms to estimate skeletal age reported a root-mean-square error (RMSE) of 0.24 years with ANN and 0.25 years with a genetic algorithm when compared to traditional estimates of skeletal age.^[36]

Taking panoramic radiographs make orthodontists legally liable if they overlook diagnosing a lesion or a tumor. This has led to the development of an automated neural network system that can correctly diagnose ameloblastomas and keratocystic odontogenic tumors from panoramic radiographs 83.0% of the time.^[37] Five oral and maxillofacial surgeons who examined the same radiographs correctly diagnosed the problems 82.9% of the time. The difference lies in the time needed for diagnosis. The ML system required an average of 38 s, while the surgeons needed 23.1 min for each diagnosis. Another system was developed that successfully predicted odontogenic cysts, dentigerous cysts, osteomyelitis, periapical cysts, and ameloblastomas 90.6%, 90.9%, 99.4%, 89.6%, and 100% of the time, respectively.^[38] At present, more and more orthodontists are using cone-beam computed tomography, which has led to the development of an automated system using the support vector machine to correctly diagnose periapical cysts and keratocystic odontogenic tumors 100% of the time.^[39] Neural networks were used to estimate patients’ dental ages from panoramic radiographs.^[40] Its RMSE was 0.9 for girls and 1.1 for boys, while traditional regression had an RMSE of 1.3 and 1.4 for girls and boys, respectively.^[40]

Panoramic and lateral cephalometric X-rays have also been used to predict maxillary canine impactions based on angular and linear measures.^[41] The highest prediction accuracy was obtained with a random forest algorithm, which correctly predicted the actual eruption status of canines 88.3% of the time.

One of the challenges for less experienced orthodontists is the selection of the appropriate treatment modality and appliance, including headgears. To address this, a system was developed to help orthodontists select the headgears that should be used.^[42] Compared to the selections made by eight expert orthodontists, the system correctly identified the appropriate headgears 95.6% of the time. Recently, decision support systems were developed to determine the geometry of orthodontic springs used to close extraction spaces^[43] and to determine the forces needed to align teeth,^[44] but neither system has been applied clinically.

Another orthodontic challenge during treatment planning is predicting the size of unerupted teeth. To address this, a hybrid system using ANN and genetic algorithms was used to predict canine and premolar sizes.^[45] Its maximum error was 2.4 mm for the mandibular and 1.6 mm for the maxillary teeth. The errors were often half as large as the error produced with linear regression prediction models.

ML AND TREATMENT OUTCOMES

One of the more useful applications of AI in orthodontics is the prediction of soft tissues treatment outcomes. Recently, ANN was used to predict the change in lip curvature after orthodontic treatment with or without extractions.^[46] Its prediction of change and the actual change that occurred differed by 29.6% and 7% for the upper and lower lips, respectively. Both predictions were much better than those based on linear regression.

The topic of beauty is controversial because it is subjective and affected by factors such as age, sex, and ethnic backgrounds. Using ANN, facial attractiveness was quantified on a scale from 0 to 100 (0 extremely unattractive and 100 extremely attractive) before and after orthognathic surgery.^[47] The difference between the pre- and post-surgery scores was shown to be statistically significant, with facial attractiveness improving 74.7%.

Predictions of treatment outcomes in Class II and Class III patients have also been reported. Using ANN, predictive models were developed to predict the post-treatment peer assessment rating (PAR) index in Class II patients based on their pre-treatment PAR index.^[48] The neural network model used in this system was able to correctly predict the final PAR score 94.0% of the time; linear regression was correct only 82.0% of the time. A system has also been developed to predict outcomes in untreated Class III patients.^[49] Unsupervised learning was used to cluster patients as hypermandibular, hyperdivergent, or balanced based on cephalometric variables. The system was then applied to a treated sample, where it showed that all of the unsuccessful cases belonged to either the hypermandibular or the hyperdivergent cluster. Another system was able to correctly predict the prognosis of Class III treatment 97.2% of the time, which was slightly better than 92.1% reported for discriminant analysis.^[50]

ML AND GROWTH PATTERNS

Several methods have been introduced to help orthodontists classify their patients’ growth patterns.^[51-53] In 1998, an ANN was used to classify the growth of 43 untreated children based on size and shape changes.^[54] However, the system was not validated on an external sample. A recent study used cephalometric variables to classify patients’ craniofacial growth as either normal or abnormal.^[55] It showed that support vector machines could correctly classify abnormal growth patterns 99.8% of the time. Another study using support vector machines to classify normal or abnormal skeletal patterns based on craniofacial measures was correct only 74.5% of the time.^[56]

Classification of Class III growth patterns has also been performed. Based on longitudinal data of untreated Class III subjects, who were classified as either good or bad growers based on the changes in their sagittal relationships, a classification tree had a significantly lower rate of misclassification (12.0%) than discriminant analysis (40.7%), both of which were based based on the same 11 cephalometric variables.^[57] When the system was tested on new data, it was able to successfully identify good and bad growth patterns 64.0% of the time.

CONCLUSIONS

AI and ML systems applied in orthodontics provide promising tools that can improve clinical practice. These clinical decision support systems can help orthodontists practice more efficiently, reduce variability, and eliminate subjectivity.^[58] The accuracy of most systems presently available is considered good to excellent ranging from approximately 64% to 97%. The accuracy at the lower end of this range should be expected to improve in the future as sample sizes increase and more information becomes available. Most of the systems were developed using restricted samples that reduce their generalizability. For example, patients were often excluded because they needed surgery or had missing teeth, unusual extraction patterns, or asymmetries. Future studies are needed to build predictive models that include different types of patients. Algorithms should also be expected to improve, making it possible to handle more complex data such as images. Systems based on images require more time, experience, and training data than systems based on discrete or continuous data values. This is especially important in the era of digital dentistry, where all patient’s records such as dental models, X-rays, and facial photos are stored in computers in the form of digital images.

It is important to note that AI models are limited and have drawbacks. They should be used only after careful considerations. Like any statistical model, the ML algorithms are based on assumptions and have limitations. If used incorrectly, they can give misleading information. In addition, the quality of data is very important.^[59] Data with a lot of noise, missing information, and more variables than observations can result in poor models. Moreover, the phenomena called overfitting occurs when a model is trained too many times on too few observations.^[60] Such models perform poorly when introduced to new data. Keeping that in mind, orthodontists should understand that these AI models are meant to assist with the clinical judgment and not to substitute for the knowledge and expertise of humans.