Research and application of artificial intelligence in dentistry from lower-middle income countries – a scoping review

Introduction

Artificial Intelligence (AI) techniques, such as Machine Learning (ML) and Deep Learning (DL), hold promise in revolutionising dental diagnostics by improving accuracy and efficiency. Trained AI can detect teeth, identify pathologies, and anomalies from dental radiographs with greater accuracy and speed compared to human specialists. However, AI research and its application in dentistry face challenges, including the need for high-quality datasets, specialised human resources, and expensive hardware like graphic processing units (GPUs). These factors, combined with limited resources in Low and Middle-Income Countries (LMICs), may hinder the advancement of AI in healthcare. Ironically, LMICs could benefit the most from AI in standardising clinical judgement and improving access to dental care. Yet, AI trained on datasets from high-income countries may introduce biases when applied in LMICs due to differences in disease prevalence and population characteristics. To address these challenges, AI algorithms must be trained on context-specific datasets to ensure relevance and equality. This scoping review aims to analyse the utility and development of AI in dentistry within LMICs, focusing on dataset quality, challenges, maturity of AI models, and considerations for cost-effectiveness. Understanding AI’s potential in oral healthcare can help improve access and integration in areas with limited resources.

Methods

The scoping review followed PRISMA Extension for Scoping Reviews guidelines, employing a predefined protocol accessible through the Open Science Framework. Collaboratively developed with a medical information specialist, the search strategy targeted the intersection of dentistry and artificial intelligence (AI) in low-middle income countries (LMICs). A thorough literature search from January 2010 to February 2023 was conducted across major health sciences databases and supplemented by manual searches in Google Scholar and IEEE Xplore. The search terms encompassed variations of “Dentistry,” “Artificial Intelligence,” “Deep Learning” and “Machine Learning.”

Research articles included were primary quantitative and/or qualitative, published in English after 2010, focusing on AI implementation in dental health within LMICs. The following were excluded: protocols, conference proceedings, letters to editors, and policy documents. Screening involved title screening and conflict resolution through discussion. Data from selected studies were extracted based on predetermined criteria, including country of origin, dentistry field, dataset types, algorithms used, computational resources, and challenges in LMIC-based AI research.

Eligible studies focused on dentistry, utilised AI models such as machine learning and deep learning, and were published in English. The following were excluded: reviews, editorials, commentaries, conference papers, studies in languages other than English, and non-indexed studies. The goal was to provide a comprehensive overview of studies, including their origins, datasets, study designs, performance metrics, and the maturity of AI algorithms, with a focus on the current state of AI in LMICs.

Results

A total of 1578 articles were initially identified through electronic and manual literature searches. After removing duplicates, 1357 articles underwent further screening, resulting in 25 eligible studies for final analysis. Most studies were from India, focusing on prediction and identification of pathologies like dental caries and periapical lesions. Orthodontics was the specialty where AI was predominantly applied, followed by Endodontics. The majority of studies focused on validation and utilised a quantitative approach, particularly for diagnostic test accuracy. Orthopantomograms (OPGs) were the most commonly used datasets across specialties, while Restorative Dentistry utilised a variety of datasets including health records and histology patches. Most studies reached stage 2 of clinical AI development maturity, with only one study reaching stage 4. Limitations mentioned in the studies were mainly related to dataset size and generalisability, with no specific mention of difficulties in LMIC setups or cost-utility analysis. Heat map analysis indicated a diverse range of datasets used in different dental specialties, with Orthodontics utilising the most variety.

Discussion

The adoption of Artificial Intelligence (AI) in dentistry has gained traction, yet there’s a gap in acknowledging the contributions of researchers from developing countries, potentially leading to biases and health inequities. This review aims to shed light on the progress of dental AI research in low and middle-income countries (LMICs), analysing the focus, methodologies, and challenges faced.

Out of 1578 initially identified articles, only 25 met inclusion criteria, indicating a scarcity of research in LMICs. AI applications focused mainly on validation, with emphasis on diagnostic accuracy. Orthopantomograms (OPGs) were frequently used datasets. Deep learning techniques, particularly Convolutional Neural Networks (CNNs), were prevalent, though datasets were often small and not representative of the population. The lack of curation and annotation reporting raises concerns about dataset quality.

While most studies focused on algorithm-to-model stages, few progressed to device deployment in real-world settings, with limited patient recall. Notably, there’s a lack of cost-effectiveness analysis and consideration of patient experience in AI implementation.

The review emphasises the need for improved infrastructure and resource allocation to facilitate AI research in LMICs. Addressing dataset quality, regulation, and cost-effectiveness are essential for leveraging AI’s potential in dentistry. Despite limitations in making definitive conclusions due to study variability, this review provides valuable insights into the state of AI in dental research in LMICs.

Conclusion

This scoping review comprehensively gathered literature concerning the application and research of Artificial Intelligence (AI) in dentistry from Low and Middle-Income Countries (LMICs). The majority of studies, primarily from India, focused on Orthodontics, emphasising diagnostic test accuracy through validation studies with a quantitative approach. Orthopantomograms (OPGs) were the most commonly used datasets. However, the scarcity of datasets was notable across all included studies. Most studies reported AI maturity at level 2 (based on the stages of clinical AI development maturity), indicating a need for further development. Notably, there was no mention of limitations related to resource constraints in LMICs, despite the evident lack of research from these regions. Key gaps identified include the absence of cost-utility analysis and consideration of patient experience. The review highlights significant variability in result reporting. Overall, the limited research included in this review demonstrates the deficiency in AI application and research in dentistry within LMICs. With support from the global community, researchers in LMICs can access necessary resources to enhance AI utilisation in dentistry, leading to increased research publications in these countries.