Accuracy of Artificial Intelligence in Predicting Facial Changes Post-Orthognathic Surgery: A Comprehensive Scoping Review

Background Accurate prediction of facial soft tissue changes post-orthognathic surgery is crucial for treatment planning and patient communication. Current models pose limitations due to the complexity of facial biomechanics and individual variances. Artificial intelligence (AI) has emerged as an important tool in many disciplines, including the dental field. Objectives The aim of this scoping review is to assess the accuracy of AI in predicting facial changes post-orthognathic surgery in comparison to traditional models. Explore the strengths and limitations of the current AI models. Material and Methods Following PRISMA-DTA guidelines, a comprehensive search was conducted manually and through Medline, Embase, Web of Science, Scopus, and Google Scholar databases was conducted, focusing on studies that applied AI models with various machine learning and deep learning algorithms for post-surgical outcome prediction. Selection criteria were based on the PICO format, emphasizing studies that compared AI-predicted outcomes with actual post-surgical results. Literature was searched until January 31, 2024. Results The initial search result yielded 1579 records. After screening and assessment for eligibility, seven studies met the inclusion criteria, with publication dates ranging from 2009 to 2023. Several AI algorithms were evaluated on different orthognathic surgical procedures, revealing the high predictive accuracy of AI models across various facial regions. Conclusions AI demonstrates significant potential for enhancing the precision of facial outcome predictions following orthognathic surgery. However, despite the promising results, limitations such as small sample sizes and a lack of external validation were noted. Further research with larger, more diverse datasets and standardized validation methods is essential for optimizing AI’s clinical utility. Key words:Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL), Orthognathic Surgery, Facial Soft-tissue Prediction, Predictive Accuracy, Orthodontics.


Introduction
Artificial Intelligence (AI), defined as "a system's ability to interpret external data correctly, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation." has had a significant impact since it was established in the 1950s (1).Initially focused on creating "thinking machines" that mimic human intelligence and behaviour, AI has evolved to encompass a variety of technologies capable of replicating human decision-making and problem-solving abilities (2).This innovation in technological progress is practical, enhancing human productivity by efficiently completing tasks using extensive datasets to convert data into actionable information for specific tasks, such as the diagnostic processes in medical sciences, improving accuracy in diagnosis and patient care outcomes (3).Machine Learning (ML) is an integral subset of AI, incorporating algorithms that improve with exposure to more data.Deep Learning (DL), a subcategory of ML, Fig. 1: Infographic representation of the relationship between artificial intelligence (AI), machine learning (ML), artificial neural networks (ANNs), and deep learning (DL).Obtained from source: (38).uses neural networks to estimate complex non-linear associations between input and output variables (Fig. 1).These algorithms, capable of accomplishing tasks at a faster pace than humans, mark a significant advancement in AI (4).Deep learning applications, including image recognition, speech recognition, and natural language processing, demonstrate the efficiency of these algorithms.Deep learning has shown remarkable performance in computer vision tasks, and their application extends to various fields including healthcare, where they assist in diagnosis and decision-making processes (5,6).Artificial Neural Networks (ANN) and Convolutional Neural Networks (CNN) are subsets of DL.ANNs are inspired by the biological neural networks in the human brain and play a key role in machine learning, enabling the analysis of complex relationships within large datasets.Typically, an ANN is structured with at least three layers: an input layer, an output layer, and one or more hidden layers.These layers are interconnected, forming a network that processes information (Fig. 2) (7,8).CNNs, known for their exceptional performance in handling high-resolution images, are vital for deep learning.They are especially suited for image and pattern recognition tasks, thanks to their convolutional layers, pooling layers, and fully connected layers.Convolutional layers apply filters to input data to create feature maps, ideal for recognizing objects, shapes, and patterns.Pooling layers simplify the feature maps by preserving essential information but reducing their size, which helps in making the network more efficient.Finally, the fully connected layers integrate these insights for better decision-making, making CNNs superior to ANNs for tasks involving images (Fig. 3) (7,9).The integration of AI technologies, including Machine Learning ML and Deep Learning DL, into healthcare, specifically in fields like dentistry and orthodontics, represents a significant advancement.They have become increasingly prevalent due to their exceptional accuracy in handling large data, learning tasks, predictions and decision-making processes.This advancement has allowed them to match the performance of skilled healthcare professionals in various aspects of patient care.Notably, the application of ANNs and CNNs within orthodontics exemplifies this progress.Initially used for radiographic analysis such as radiographic lesions detection and automatic cephalometric landmarks identification (10)(11)(12).These technologies have evolved into more complex decision-making tools for treatment planning, showcasing their versatility and effectiveness in enhancing clinical outcomes (13)(14)(15)(16).In orthodontics, enhancing facial aesthetics is a key goal.Improving facial aesthetics is a primary motivation for patients undergoing orthognathic surgery, with the prediction of irreversible outcomes presenting significant challenges.However, accurately predicting post-treatment facial appearance is challenging due to complex biomechanics and minimal craniofacial changes.Indivi-dual factors like healing processes, bone structure, and soft tissue response, which are difficult to accurately predict, play a crucial role in these outcomes (17)(18)(19)(20).The acknowledged AI's capabilities are not to be overlooked.Furthermore, it could be an invaluable tool for precisely predicting the post-surgical facial appearance following extensive orthognathic structural changes.There is a lack of thorough reviews on the effectiveness of AI and its various models in predicting facial topology post-orthognathic surgery.Therefore, the objective of this scoping review is to systematically examine the literature on the precision of AI's capabilities in predicting changes in facial soft tissue following orthognathic surgeries.The relevant MeSH (Medical Subject Headings) terms used for the search strategy were: 1-"Artificial Intelligence"(MeSH) 2-"Machine Learning"(MeSH) 3-1 OR 2 4-"Orthognathic Surgical Procedures"(MeSH) 5-"Treatment Oucome"(MeSH) 6-"Prediction"(MeSH) 7-5 OR 6 8-3 AND 4 and 7

Material and Methods
The search syntax using keywords with Booleans and truncations for each database searched is found in (Table 1).Furthermore, a manual search was conducted complementing the electronic search process, involving the examination and exploration of the reference lists of the initially chosen articles.
-Criteria for literature eligibility: Only original clinical studies that followed the designated PICO format were included in this review (randomised-and/or non-randomised clinical trials, longitudinal prospective or retrospective cohort clinical studies).Attempts were made to contact corresponding authors for any relevant inaccessible literature, such as 'abstract only' or missing full-text to obtain the full-text wherever possible.Any other literature such as Case reports, correspondence letters, commentaries, reviews were excluded.

-Search strategy results:
A comprehensive search of all databases and a manual search yielded the identification of 1579 articles.Following the implementation of the exclusion criteria and duplicates detection, the total of 132 studies were retained.Out of the total, 125 studies were eliminated following a thorough assessment of the title and/or abstract's relevance to the scope of this review.A total of seven studies were ultimately incorporated into the current review and subjected to data extraction (Fig. 4).-Characteristics of the included studies: Seven studies that have utilized the use of AI models in predicting facial soft tissue outcomes following orthognathic surgery were included in this scoping review, as indicated in (Table 2).The outcomes primarily assessed the precision of soft tissue prediction, highlighting mean errors for different facial areas and overall accuracy rates of prediction.All studies were retrospective by design.Three were case-control studies (22)(23)(24); similar were cohort studies (25)(26)(27); and one was an experimental proof-of-concept study design (28).The sample sizes within the studies ranged from five to 137 patients.Three studies reported surgical intervention types such as bimaxillary surgery, mandibular advancement, and maxillary advancement surgeries (22,24,26), while four   No significant difference between the two methods.

Speed improvement:
<2 mins for simulation.Additionally, the application of AI and DL in surgical planning provides a cost-effective alternative to conventional treatment simulations, potentially leading to better patient outcomes (34,35).
On the other hand, it is not without limitations, and they must be addressed to fully harness the potential of AI and DL.The reported studies demonstrated high variability in AI models' algorithms, which can lead to inconsistent results.This emphasizes the need for standardization in model development and validation processes.Small sample sizes in many studies limit the generalizability of findings, necessitating efforts to increase sample size and dataset diversity to improve model robustness.Additionally, the lack of external validity raises concerns regarding the performance of models on datasets different from the training data.Cross-validation and external validation are needed to be certain about the reliability and generalizability of the AI predictive accuracy.Moreover, the scarcity of big data for AI training and validation is a significant challenge that warrants the need for collaborative efforts to compile comprehensive databases for robust training and validation (36,37).

Conclusions
In conclusion, AI holds great promise for improving the prediction of facial soft-tissue changes following orthognathic surgery.While current models have demonstrated a high level of accuracy and reliability, ongoing research and development are essential to overcome existing limitations and fully realize the potential of AI in this field.By addressing these challenges, AI can become an invaluable tool for clinicians, enabling more precise and personalized treatment planning and ultimately improving patient outcomes.

-
Search strategy: This structured scoping review was conducted following the guidelines outlined in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Diagnostic Test Accuracy (PRISMA-DTA) (21).A comprehensive search was conducted in Medline via Pubmed, Embase, Web of Science, Scopus and Google Scholar electronic databases to identify and select the literature for this paper.The retrieved results encompassed all indexed literature within each database until January 31st, 2024, as no specific date range was specified.The search strategy in this review was conducted in line with the (Participant/Population, Intervention, Control/ Comparison and Outcome/Result) (PICO) format; Patients undergoing orthognathic or cranio-facial surgery was identified as the population (P).Intervention (I) was identified as the use of AI including different ML and DL algorithms for the prediction of post-surgical soft-tissue outcome.Control (C) classical clinical-based or computer-based outcome prediction without AI involvement.The Outcome (O) the level of diagnostic test accuracy, sensitivity and specificity between AI-predicted and post-surgical obtained actual soft-tissue outcomes.

Table 1 :
Search strategy used for electronic databases.
PubMed 291 (artificial intelligence OR machine learning OR deep learning OR neural network* OR algorithm* OR machine intelligence OR ANN OR CNN) AND (orthognathic OR orthognathic surgery OR jaw surgery OR mandibul* OR maxill* OR facial OR cranio*) AND (prediction OR simulation OR outcome OR plan OR assessment OR forecast) AND (soft-tissue OR facial) Web of Science 400 ((ALL= (artificial intelligence OR machine learning OR deep learning OR neural network* OR algorithm* OR machine intelligence OR ANN OR CNN)) AND ALL= (orthognathic OR orthognathic surgery OR jaw surgery OR mandibul* OR maxill* OR facial OR cranio*)) AND ALL=(prediction OR simulation OR outcome OR plan OR assessment OR forecast) AND ALL=(soft-tissue OR facial) Embase 146 TX (artificial intelligence or machine learning or deep learning or neural network or ANN or CNN) AND TX (prediction or outcome or simulation or forecast) AND TX (orthognathic or mandibl* or maxill* or cranio*) AND TX (soft-tissue or facial) Scopus 43 TX (artificial intelligence or machine learning or deep learning or neural network or ANN or CNN) AND TX (prediction or outcome or simulation) AND TX (orthognathic or mandibl* or maxill* or cranio*) AND TX (soft-tissue or facial) Google Scholar 699 (artificial intelligence OR machine learning OR deep learning OR neural network* OR algorithm* OR machine intelligence OR ANN OR CNN) AND (orthognathic OR orthognathic surgery OR jaw surgery OR mandibul* OR maxill* OR facial) AND (prediction OR simulation OR outcome OR plan OR assessment) AND (soft-tissue OR facial)

Table 2 :
Characterestics of the included studies.

Table 2 :
(21,(30)(31)(32)on errors, particularly within the lower facial third and the lips(30,31).Research indicated that facial morphology is affected by a multitude of factors, including age, gender, and dental issues, suggesting a need for personalized prediction models(21,(30)(31)(32).To overcome the limitations of legacy models, Lu et al. ted that AI and DL models offer at least similar if not an enhanced accuracy compared to traditional prediction methods.This improved accuracy in treatment objective visualization is crucial for patient communication and realistic expectation management.Furthermore, these models significantly reduce computation times, enabling rapid feedback and multiple simulations, thereby enhancing clinical workflows and patient experience.