2.1 Types of CDSSs

Clinical decision support systems have been categorized according to (a) function of the support: provision of alerts (active) or reaction to patients’ information/providers’ inputs (passive); (b) time of support delivery: before, during or following decision-making and (c) method of development: knowledge or non-knowledge-based.

Knowledge-based CDSSs contain compiled data entered directly by users or extracted from patients’ electronic records, phone apps11 and data about medications12 or clinical protocols and guidelines based on the intent of using a clinical decision system.3, 12 Such systems will deliver support to users, usually following the IF-THEN rule. For instance, in a system for detection of medication interactions, the rule will be IF drug X and IF drug Y are used, THEN alert the clinician.12 Non-knowledge-based CDSSs also require a data source; however, decisions are made with statistical pattern recognition or machine-learning (ML) approaches. Consequently, computers learn from previous experiences, discover patterns within the data and eliminate the need for expert inputs or rules. Artificial neural networks and genetic algorithms compose the main types of non-knowledge-based CDSSs.13, 14

2.2 Data science approaches for the development of CDSSs

Clinical decision support systems development is a highly challenging and multidisciplinary task that integrates clinical knowledge with decision science to adapt clinical practice and workflow with the decision system.15 Here, we provide a simplified overview of steps involved in creating a clinical decision support system (Figure 2A).

Overview of steps involved in developing clinical decision support systems (CDSSs). A, General spectrum of data science approaches in the CDSSs. B, Implementation of robust data management, Dental Storage for Computation and Integration, in the CDSSs

2.2.1 Data collection, storage and management

The advancement of information technology promoted exponential growth of health data and generation of big data15; that is, large-volume data that are produced at high speed and integrates different types of data, such as: (a) clinical data (e.g. orthodontic diagnosis and treatment progress notes, imaging, health records), (b) omics data (e.g. genomics, proteomics, metabolomics), (c) patient-generated data (e.g. wearable devices and scanners, social media) and (d) normative data sets (e.g. data collected in clinical trials or nationwide surveys).2, 16, 17 Current evolution of data capturing, storage and analytical methods will allow us to transform the wealth of knowledge in big data into actionable plans to overcome challenges in clinical decision, deliver personalized care and improve the population's health.18, 19

Collection of patient diagnosis and treatment progress data is considered valuable when it is done in a systematic way following interlinkable and coherent data standards that produce high-quality information.17-20 Clinical data constitutes an essential resource for medical and health research—electronic health records (EHR) are one of the major types of clinical data.21 The use of EHR is highly encouraged to improve clinicians’ compliance with documentation and enable data sharing among different members of the healthcare team.16 In 2009, the Health Information Technology for Economic and Clinical Health (HITECH) Act was created to support the adoption and meaningful use of health information technology, including a provision of financial incentives for using EHR systems.22 By 2015, the adoption of certified EHR programs by hospitals and office-based physicians reached 96% and 78%, respectively.23, 24 Nevertheless, the quality of data within EHR was affected by duplication, missing information, fragmentation and inconsistent organization. The accuracy and reliability of data can influence the development and use of CDSSs, even in the presence of sophisticated advanced technologies. Thus, there is a need to standardize the terminology within the dental and orthodontic fields using well-structured forms and templates to assist in ensuring consistency of the collected data.16 There have been several governmental efforts to standardize the data within different EHR systems; however, at this time no federal or professional association program has been able to produce universally accepted high-quality data. Indeed, data governance requires policies for care providers and auxiliary staff, hands-on training, a culture of responsibility and the right tools to improve and monitor data quality.25

Multicentre data collection is commonly performed to construct clinical prediction models. Although such data structures create additional challenges for data analysis, they cover a broader population and can improve the generalizability of the artificial intelligence models.26 Digital data repositories provide web-based platforms that enable researchers from multiple institutions to access and manage their data securely. Data repositories are designed to store large amounts of data, ranging from thousands of data set repositories, supported by funding and government agencies, to small data sets, supported by a research team for a certain study.27

‘BigMouth’ is an example of an oral health data repository that contains EHR from 11 dental schools across the United States. It provides access to the demographic, dental and medical data of over 3 million patients to facilitate the advancement of research and patient care outcomes.28, 29 Defining and establishing a code of conduct is important for big data collaborations to guide the ethical and meaningful use of shared data. When preparing data for sharing, it is essential to strike a balance between privacy protection (e.g. de-identification & anonymization of patients' data, security measures, controlled data sharing) and maintenance of data utility.16, 30

2.2.2 Data processing

After defining the proper source for data collection, it is essential to verify the quality of data and prepare a final data set for analysis or machine training.31

In 1955, a mathematician named John McCarthy coined the term artificial intelligence (AI) to describe the ability of machines to conduct tasks that lie within the range of intelligent activities. Afterwards, Richard Bellman defined AI as the ability to automate activities with human thinking capabilities; for example, problem-solving, learning and decision-making.32 Machine learning (ML) is a subfield of AI, whereby algorithms are utilized to find structures and patterns within data. Consequently, machines will learn to predict similar patterns on unseen data, and their actions will improve each time new data are introduced without human inputs.9, 33, 34

High-quality data sets are essential for developing effective machine learning models. To build an ML model, three non-overlapping data sets are utilized for training, validation and testing. The training data set is used to develop the ML model. Then, model parameters are adjusted with the validation data set. The last step involves testing the model performance on unseen data, that is, test data set.35 In ML, raw data is not usually suitable for learning; features/variables should be identified and extracted from the raw data via data mining. However, raw data should be cleaned and prepared for data mining; noise and errors within raw data might confuse the data mining process and thus result in faulty detection of patterns.36 The aims of cleaning and enriching data are primarily data normalization, elimination of redundant features, data standardization, removal of duplicates, resolution of inconsistent data, management of missing data and data matching across multiple sources.35, 37 Images normalization is an example of a data preparation procedure required before applying statistical or ML algorithms. Putting several images in a common statistical distribution based on the size and pixel values (spatial normalization) will enable reliable detection of changes between several individuals or within the same individual at different time points.38-40 Additionally, normalization of images' intensity should be performed when planning to develop a model that classifies a disease, to avoid biasing the results, or building a model for image synthesis, for example, segmentation and transformation. Evidence demonstrated the accuracy of the image synthesis model is greatly affected by standardizing the intensity features across the training's input images.41, 42

Data mining involves analysing large data sets to extract unknown patterns and comprehensible information from large data sets.43 Several functionalities or tasks can specify the knowledge found in data mining, such as regression, clustering and classification.36 These findings can summarize the input data or be utilized in additional analyses such as machine learning and predictive analytics. Although algorithms are used in data mining and ML, outputs from data mining help optimize decisions, for example, detecting valuable clinical information will help the practitioner make better decisions and increase the quality of care. However, training a machine with the extracted knowledge will enable predicting the diagnosis or prognosis of a new patient.9

Dimensionality reduction is performed before machine training to eliminate irrelevant and redundant data, improve learning accuracy, and enhance output comprehensibility. The main types of dimensionality reduction are feature selection and feature extraction. Feature selection involves selecting data that contains the most relevant information for solving a particular problem; a subset of the original data is maintained and used for machine training.44 For instance, Bianchi et al,45 conducted a study aiming to detect temporomandibular joint osteoarthritis (TMJ OA). Only the most robust features were selected for machine training out of the collected 52 clinical, biological and radiomic markers and 1326 interactions. Similarly, in AI systems for two-dimensional cephalometric analyses, the rates of success of landmark detection and classification of skeletal and dental problems depend on the proper selection of the diagnostic variables.46 On the other hand, feature extraction transforms the original features into a new, smaller set of more significant features.44 In the study of Bianchi et al,45 twenty three-dimensional imaging features of bone texture and bone morphometry, that quantify the initial morphological changes in the condylar trabecular bone, were extracted and used for machine training instead of analysing the whole Cone-Beam Computed Tomography (CBCT) scan grey level voxels.

After implementing and tailoring the previous steps according to the aim of the ML model, the data are considered ready for advanced analysis and machine learning.

2.2.3 In-depth data analysis

Following data preparation, each clinical application of artificial intelligence in orthodontics requires the selection of the proper machine-learning methodology, training the ML model and evaluation of the developed model's performance.

The success of the ML algorithms depends on the thorough comprehension of what algorithms can provide, the limitations of algorithms, and how that will support and fit into clinical care. Hence, communication between data analysts, data scientists and clinicians is important during all phases of CDSSs development.47 Within dentistry, different ML algorithms have been utilized based on the size of the data, variables/features to analyse, and the objective of the model. ML can involve unsupervised or supervised learning.48

In unsupervised learning, algorithms detect hidden patterns within an unlabelled data set. That means all variables within the training data set are utilized as inputs, and the machine will automatically discover structures/patterns within that data set without receiving instructions about the desired outcomes.48, 49 Based on the problem at hand, unsupervised learning algorithms will split the data set into groups (clustering) or find rules representing the relationship between variables within a data set (association).49 For example, in a study conducted by Auconi et al,50 combinations of variables (inputs) were provided to the fuzzy cluster, which detected the best phenotypic factors to group a sample of Class-III patients into subjects with increased mandibular dimensions, subjects with increased maxillomandibular divergence, and subjects with intermediate characteristics between the two groups.

Supervised learning algorithms, on the other hand, analyse training data sets with predetermined inputs and outputs. Consequently, the inferred ML model can predict the outputs of new data. Common tasks of supervised learning algorithms include classification and regression.49 The classification task aims to detect a function (discrete value) that aids in splitting the data set into classes based on various parameters.51 For example, using a classification algorithm and a training data set consisting of patients' intra-oral and cephalometric findings, the supervised ML model can determine (i.e. classify) the cases that need or do not need a tooth extraction for orthodontic reasons.52 In the regression task, the correlation between dependent and independent variables is detected during the machine training, and the developed model can predict continuous variables.51 For instance, an ML model trained with a regression algorithm can predict dental age (continuous variable) from the pulp-to-tooth ratio of the canines.53 When training with supervised ML algorithms, it is possible to generate an overfit model that performs well only on the training data set. Thus, it is essential to evaluate the developed model's generalizability and tune the model parameters through validation methods, for example, cross-validation.54 Last, the performance of the model should be evaluated using the test data set. For that, various methods exist and should be selected based on the task of the ML algorithms (classification or regression) and the type of outcomes.55

Interestingly, a review conducted by Asiri et al9 revealed that most of the ML applications in orthodontics have utilized supervised machine learning algorithms to automate clinical procedures that execute or assist in diagnosis and treatment planning. Radiographs were commonly targeted in developing those ML models as they are considered essential tools for orthodontic diagnosis, treatment planning and evaluation of treatment outcomes. Discrepancies in landmarks identification have been recognized as a critical source of error in cephalometric analyses. Since the analysis' diagnostic value relies on the reproducibility and precision of landmarks identification, interests to develop an automated approach have increased to reduce the laboriousness of the task and subjectivity of the analysts. Park et al56 reported that training the machine with the YOLOv3 algorithm for automatic labelling of 80 landmarks resulted in small error plots and 5% improvements in the accuracy compared to top benchmarks reported in the literature. Additionally, the mean computational time consumed per image was only 0.05 seconds. Similarly, Kunz et al57 reported promising results obtained with an AI algorithm that can analyse new cephalometric X‑rays with precision comparable to the gold standard, that is, experienced human examiners. Furthermore, Lee et al58 presented an automated framework for cephalometric landmarks detection with the implementation of confidence regions (95%) around the estimated positions of the landmarks. This will allow clinicians to gauge the accuracy of the size and location of the calculated landmarks.

Besides using the two-dimensional radiographs for the cephalometric analysis, accurate measurements can be attained with CBCT imaging modality; CBCT provides an accurate three-dimensional spatial representation of the oral and craniofacial structures. However, the accuracy of manual landmarks plotting on the CBCT requires substantial effort, experience, and time. Gupta et al reported an automatic knowledge-based landmark detection algorithm able to produce accurate cephalometric measurements comparable to those computed from manual identification.59 Automation of the radiographic analysis has also involved attempts to estimate the skeletal maturation that provides the best estimate of the individual's biological age and aid treatment planning.60 Kashif et al61 used a classifier algorithm and developed a tool that can help predict the bone age from hand radiographs with a mean error of 0.6 years compared to the average reading of two experienced radiologists. Kok et al,62 on the other hand, reported different algorithms that predict the skeletal maturation from the cervical vertebrae on lateral cephalograms. ML algorithms have also been utilized to develop an automated imaging system that provides objective morphological facial assessment during the orthodontic diagnosis process.63 Furthermore, various methods for automatic volumetric segmentation of CBCT images were developed using ML algorithms. This will allow objective generation of three-dimensional models for advanced diagnosis and treatment planning whilst saving the time and efforts required for manual segmentation.64

Growth and development of orofacial complex are influenced by the interaction of the genetic and environmental factors. Orthodontists are mainly using phenotype-driven diagnostic tools like cephalometric analyses to predict the growth in individuals with Class-II or class-III skeletal malocclusions. However, future studies of genetics, epigenetics and metabolic pathways that utilize advanced machine learning tools will transform the process of orthodontic diagnosis and treatment planning.47, 65

2.2.4 Data communication

The next step after developing the ML-based model is to communicate the ML findings with the clinicians. Production of predictive models in a real-world healthcare setting is more challenging than developing models in an experimental environment. Therefore, it is important for clinical experts who were not involved in tools development to test and validate the system's performance. Following the deployment of the ML model, it should be monitored for reliability and correction of errors since clinical protocols and populations are changing over time.66

The CDSSs, reviewed in this manuscript, will improve Orthodontic care only if clinicians utilize ML and AI tools to analyse the inter-relationships among the dentition, craniofacial skeleton and soft tissues. Then, translate the acquired knowledge towards the advancement of orthodontic diagnosis, treatment planning, evaluation of growth and development, assessment of treatment progression, outcomes and stability.