Decision Support Systems in Temporomandibular Joint Osteoarthritis: A review of Data Science and Artificial Intelligence Applications

Abstract

With the exponential growth of computational systems and increased patient data acquisition, dental research faces new challenges to manage a large quantity of information. For this reason, data science approaches are needed for the integrative diagnosis of multifactorial diseases, such as Temporomandibular joint (TMJ) Osteoarthritis (OA). The Data science spectrum includes data capture/acquisition, data processing with optimized web-based storage and management, data analytics involving in-depth statistical analysis, machine learning (ML) approaches, and data communication. Artificial intelligence (AI) plays a crucial role in this process. It consists of developing computational systems that can perform human intelligence tasks, such as disease diagnosis, using many features to help in the decision-making support. Patient's clinical parameters, imaging exams, and molecular data are used as the input in cross-validation tasks, and human annotation/diagnosis is also used as the gold standard to train computational learning models and automatic disease classifiers. This paper aims to review and describe AI and ML techniques to diagnose TMJ OA and data science approaches for imaging processing. We used a web-based system for multi-center data communication, algorithms integration, statistics deployment, and process the computational machine learning models. We successfully show AI and data-science applications using patients' data to improve the TMJ OA diagnosis decision-making towards personalized medicine.

INTRODUCTION

A fundamental question in medical and dental research is how to improve the diagnosis of complex diseases. In this context, decision-support systems have been developed to help clinicians to do a correct diagnosis and treatment plan, allowing an improved decision-making process. Decision-support systems involve different approaches, including universal guidelines for specific diseases, such as the DC/TMD¹ for the temporomandibular disorders, software containing different algorithms, such as image-guided applications, and image exams such as the cone-beam computed tomography for visualization and imaging processing². More recently, due to the large volume of information, the decision-support systems are being integrated into different statistical approaches where machines can now learn with human annotations, predict diseases and provide additional diagnostic tools for a personalized treatment^3-7.

For this reason, data science approaches, including artificial intelligence (AI) with machine learning (ML), are needed to improve statistical models and deployment of the support-system in web-systems, available to the public^8,9. In dentistry, temporomandibular joint osteoarthritis is an example of a disease with multifactorial etiology that requires a large number of exams such as clinical assessment of disease condition, image exams such as cone-beam computed tomography (CBCT), and magnetic resonance (MRI), and biological markers^10-13. All those exams generate information and features for each patient, and ML approaches can integrate and combine those creating prediction models with better accuracy to diagnose the disease rather than individual biomarker¹⁴.

Artificial intelligence was first defined by a group of computer scientists back in 1956 at the Dartmouth Conferences¹⁵. It has been implemented in machines (computers) to solve human problems through different learning algorithms. There are two main categories of machine learning methods called: supervised and unsupervised. The first is based on the prediction of a specific condition, where the machine knows the output in advance, i.e., the presence or absence of TMJ-OA in the data. In the second, there are no outputs to predict; however, the learning is based on trying to find data patterns that can be grouped and separate (classify) the data based on those parameters^16,17. For instance, in the TMJ-OA, the mandibular condyle shape is an important feature to determine if the disease is present or absent, Tubau et al.¹⁸, tested an unsupervised machine learning approach and compared it with nine supervised methods to classify different clinical phenotypes of TMJ-OA using statistical shape modeling properties

Artificial intelligence can also be applied in many fields such as image processing, data and features extraction, text and picture recognition, automated tasks, and others^15,19,20. There are many algorithms and methods to perform different tasks using machine learning, but we will focus on Orthodontics applications in this manuscript. A question that remains unclear is which ML algorithm should be used in a specific project, i.e., for image processing data, standardization is essential to provide accurate results. Image segmentation is the first step towards this goal. To perform reliable segmentation on CBCT images, the researcher can use automated software that lacks precision and perform semi-automated segmentation based on human interaction²¹ that adds more precision; however, it demands a large amount of time. For this goal, machine learning can help automatize using training sets of images, pre-segmented (training data) by a user to perform automatic segmentation on the testing data^22-24.

To assess the statistical models' performance in ML approaches, it is necessary to use methods such as cross-validation. They are applied to the dataset to test the prediction models' effectiveness, feature selection, and other tasks. To perform the CV, is it necessary to have divided the sample into two main categories: 1) Training Set (used for training the learning algorithm) and 2) Testing Set - used for testing and validating the different ML models^25-31.

In general, this century is being changed by this revolution on health care and personalized medicine, requiring holistic models based on the patients' characteristics and needs^32-35. For this reason, in this paper, we aimed to describe and review general aspects of artificial intelligence and machine learning applied in Orthodontics, specifically towards the decision-support system for a precise and personalized diagnosis and treatment planning in the Temporomandibular Joint region.

MATERIAL AND METHODS

This section contains the approaches applied to data science and artificial intelligence.

Data acquisition and standardization.

An optimized data-science approach that utilizes AI and ML usually require a large number of data. One of the challenges is to have those data in a standardized manner. Samples that lack standardization of the training information can lead to a not balanced learning model due to the inherent features heterogeneity. For instance, there are several methodologies in the TMJ research related to image acquisition, such as medical computed tomography, CBCT, magnetic resonance image, panoramic image, etc^36-38. However, suppose a particular exam is chosen, such as the CBCT images. In that case, there is heterogeneity also among the different types of equipment (diverse fabricants) and spatial resolution configuration (voxel size, field of view, and others). To be able to create learning models for disease classification, such as the TMJ OA, those "small differences" can cause bias in the analysis. This happens mainly due to the feature heterogeneity present within the images, previously to any image processing technique for extracting the information from that gray-level based scans^39-41. This scenario is commonly seen in the clinician routine. Still, the human mind is trained to perform diagnosis with fewer details than an automatic system, which extracts thousands of data^42,43 from a single scan, requiring standardization within the groups. To contour that challenge, image techniques may be applied to normalize the images, and among those, the filtering algorithms and CBCT grayscale normalization are widely used^44,45. Fig. 1 shows an example of image-pre processing in a TMJ condyle after applying three different Gaussian filters (sigma = 0.1 and 0.3) using the software 3D-Slicer². The baseline population characteristics need to be drawn for each project before the data collection, so the features to be evaluated do not influence the ML model.

Fig. 1 -.

Data capture and standardization, an example of a CBCT with a voxel size of 0.08 mm³ and different filters applied. A) TMJ with the original resolution and filters. B) The same CBCT with a Gaussian filter application (0.1 sigma value). C) The same CBCT with a Gaussian filter application (0.3 sigma value).

Data extraction and assessment.

This step of the data science approach and support systems focuses on extracting information from different sources to be used in the ML models as a set of features. In the TMJ OA, there are three main categories of variables used for helping in the diagnosis of the disease: clinical, imaging, and molecular markers⁴⁶. The clinical sources are mainly dependent on the human operator, and they include various signals and symptoms¹, for the molecular markers are less used in the clinical routine, and the imaging markers are essential for the AI models. The main reason is that using algorithms for data extraction such as radiomics extraction algorithms and methods⁴⁷, there is an increase in the information obtained from the sample, and that information was before not visible to the naked eye, but machine learning models can group those markers into different categories to classify in other groups of patients based on radiomics similarity. Also, the mandibular condyle shape characteristics is another feature that can be extracted from the CBCT images, being a good indicator of the disease status based on the mesh arrangement^48,49.

Data management and storage: Web-systems and computational processing.

It is well known that state of the art for machine learning algorithms performance is mostly driven by the number of samples used for training. Ideally, standardized protocols for data collection in each clinical site should be implemented to increase sample size without compromising data homogeneity. For having an efficient AI system, there is a need to deploy the data in an efficient computation system, capable of communication with different centers and users and at the same time capable of having ML algorithm and statistical methods implementation. This study presents a web-system developed by our group, called: "Data storage computation and integration – DSCI"- to store and share deidentified data⁵⁰ (Fig. 2). It aims to collaborate across medical centers, providing data transfer, sharing, and processing. Additional functionalities include ML algorithms implementations.

Fig. 2 -.

Web-system for data management, storage, and processing. A) CBCT image with standard resolution and large field of view (FOV). B) CBCT image of a TMJ condyle with high resolution and small FOV. C) Digital dental model. D) Illustration of biomolecular information storage. E) Illustration of clinical information storage.

Data processing: image processing, classification, and 3D segmentation (labeling).

Data processing can be included in a web-system and/or in a computer using traditional software for data and image processing. This step's main objective is to process the information extracted from the sample and execute additional necessary tasks. In the TMJ – OA, the CBCT exam requires a manual segmentation of the mandibular condyle to better visualization and 3D surface diagnosis. For this, the use of open-source software such as the ITK-SNAP⁵¹ and 3D Slicer⁵² can be used. This step is essential for the AI because it created the baseline sample and the training set (gold standard) for machine learning validations. Also, a common term, known as data mining refers to discovering patterns in data sets involving machine learning methods, statistics, and database systems for features extraction⁵³ and can also be deployed in the AI to provide a larger number of features; however, caution should be taken due to the possibility of overfitting⁹ the ML models due to possible correlations among the features (Fig.3).

Fig. 3 –

Biomarkers selection to avoid overfitting of the ML models.

Artificial Intelligence: Statistical analysis, cross-validation, and machine learning models.

The AI is applied as an integrative part of a decision support system. It incorporates data science approaches using different machine learning algorithms and statistical analysis to perform specific tasks, such as classifying a patient as having or not having TMJ OA¹⁴, segmenting CBCT images automatically²⁴, and features selection^19,54. To perform those tasks, the applications use algorithms based on data quality enhancement, active deep learning, active contour modeling, and typical algorithms for ML are random forest, support vector machines, light GBM, XGboost, UNet, ResNet, and many others^55-59. The decision-making process aims to use AI and ML to test a large number of machine learning algorithms to see which one has better performance in a determined task. To determine how accurately a predictive model is, cross-validation (CV) is necessary. They are mainly divided into two subtypes: K-fold cross-validation and Leave-p-out cross-validation; the CV also can be used as well to features selection^25,60.

RESULTS AND DISCUSSION

This section results from data science approaches and AI presented, including image processing steps, data standardization, statistical analysis, and 3D visualization for Temporomandibular joint applications and TMJ OA. For this goal, our team²⁴ has developed an automatized image processing algorithm called: TMJSeg in a web-system to segment in 3D the CBCT images and convert them to a surface volume (Fig. 4). This provides a mandibular condyle model that can be used to assess the shape variability in a population.

Fig. 4 –

Automatic TMJ condyle 3D segmentation - Image processing steps involved in the algorithm.

We also applied the AI using ML to detect the variability in condyles' surface with TMJ OA and health. In the study published by Ribera et al.¹⁸, our group demonstrated the ShapeVariationAnalyzer application and accuracy of a deep machine learning model to classify the condyles in different categories, according to their shape condition. Experient clinicians labeled the learning set using the degree of skeletal degeneration at the mandibular condyles' surface. The ML algorithms extracted mesh features such as Normal vector, distances, curvatures, shape index, curvedness, position, and heat kernel signatures to classify the mandibular condyles based on the shape. As a result, we obtained a maximum accuracy during the training of 92% for our ML deployed. Fig. 5 shows an illustration of the different degrees of degenerations among the sample. Other authors have evaluated mandibular condyles' shape using traditional imaging methods, such as the presence or absence of radiographic findings using the CBCT slices⁶¹. However, in images with high resolution, this subjective interpretation may be biased by the lack of a gold standard for comparisons of the thin line between health and disease in the aging process of the TMJ joints. For this reason, AI methods and computational approaches using different software can be used for better assessment and more quantitative feature extraction. A recent paper from our group⁶² showed that it is possible to extract radiomics features from the trabecular bone using CBCT with high resolution of mandibular condyles and differentiate between control and disease in patients with initial stages of TMJ OA.

Fig. 5 –

TMJ condyles used in the previous study for classification of the clinicians' training set. Degeneration 0 means that the condyle has little change to its shape than the control, and degeneration 5 is the highest disproportion compared to the healthy condyle.

For the AI approaches and ML implementation, feature extraction is needed to perform the models' cross-validation and training. The TMJ - OA is a multifactorial disease, with multisource biomarkers such as clinical variables, proteins assessment, and imaging markers that may be responsible for disease progression. Figures 3 and 5 show a summary of key features that can be extracted from mandibular condyles prior to the AI implementation or concomitantly when the ML models are used to select features and weight those features.

Also, in a recent study, our group showed the importance of integrating those variables to detect the TMJ OA in early stages¹⁴ using artificial intelligence. In Fig. 7, we illustrate the diagnosis of TMJ OA based on single features and outline the limitations of the current procedures. We have demonstrated that the integration of multisource biomarkers using ML models is able to predict the TMJ OA in its initial stages with an accuracy of 0.823. We found that not only the single features are essential to the AI's decision process, but also the interaction between them has a higher value for our tested prediction machine learning models.

Fig. 7 -.

Spectrum of decision support systems in temporomandibular joint osteoarthritis.

CONCLUSION

This paper presented an overview of data science, artificial intelligence, and machine learning approaches used in decision support systems in the temporomandibular joint research area. In conclusion, clinicians can benefit from the large amount of data obtained in the last years, using trained ML models to integrate those features, helping in the clinical decision making, especially in complex diseases with multifactorial etiology, such as the TMJ OA. This will support the clinician to do an early diagnosis and personalized treatment with higher predictability.

Fig. 6 –

Multisource Features selection and extraction to be applied in AI and ML models training.