Explainable Artificial Intelligence in Dentistry: A Systematic Review of Its Trust and Translation.

Explainable Artificial Intelligence in Dentistry: A Systematic Review of Its Trust and Translation - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Int Dent J . 2026 May 25;76(4):109626. doi: 10.1016/j.identj.2026.109626 Explainable Artificial Intelligence in Dentistry: A Systematic Review of Its Trust and Translation Sermporn Thaweesapphithak a, b , Vivat Thongchotchat a , Hamid Alinejad-Rokny c, d , Lakshman Samaranayake a, e, f , Thanaphum Osathanon g , Thantrira Porntaveetus a, h, i, ⁎ Author information Article notes Copyright and License information a Center of Excellence in Precision Medicine and Digital Health, Department of Physiology, Faculty of Dentistry, Chulalongkorn University, Bangkok, Thailand b Department of Oral Biomedical Sciences, International College of Dentistry, Walailak University, Bangkok, Thailand c UNSW BioMedical Machine Learning Lab (BML), School of Biomedical Engineering, UNSW Sydney, Sydney, NSW 2052, Australia d Visiting Scholar (Collaborative Projects), Center of Excellence in Precision Medicine and Digital Health, Chulalongkorn University, Bangkok, Thailand e Faculty of Dentistry, University of Hong Kong, Hong Kong f Global Research Cell, Dr. D. Y. Patil Dental College and Hospital, Dr. D. Y. Patil Vidyapeeth, Pimpri, Pune, India g Center of Excellence for Dental Stem Cell Biology, Center of Artificial Intelligence and Innovation, Department of Anatomy, Faculty of Dentistry, Chulalongkorn University, Bangkok, Thailand h Chulalongkorn University Implant and Esthetic Center, Faculty of Dentistry, Chulalongkorn University, Bangkok, Thailand i Clinic of General-, Special Care and Geriatric Dentistry, Center for Dental Medicine, University of Zurich, Zurich, Switzerland ⁎ Corresponding author. Center of Excellence in Precision Medicine and Digital Health, Department of Physiology, Faculty of Dentistry, Chulalongkorn University, Bangkok 10330, Thailand. thantrira.p@chula.ac.th thantrira.porntaveetus@zzm.uzh.ch Collection date 2026 Aug. © 2026 The Authors This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). PMC Copyright notice PMCID: PMC13223827  PMID: 42184717 Abstract Introduction and aims Explainable artificial intelligence (XAI) is a set of methods and processes that make the decisions of artificial intelligence (AI) models understandable to those who are not conversant with the technology. This “black box” nature of complex AI models appears to be a primary barrier to their clinical adoption in health sciences, including dentistry. XAI is being touted as a solution to build clinician trust. This review critically assesses whether current dental XAI research is methodologically rigorous enough to substantiate claims of enhanced trustworthiness. Methods This systematic review followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 guidelines, searching PubMed, IEEE Xplore, medRxiv, and Ovid for dental XAI studies (2015-2025). We assessed the risk of bias and applicability using QUADAS-2 and PROBAST. Results Nineteen of the 100 identified studies met the inclusion criteria. Although these studies used diverse XAI techniques, including image-based saliency methods (eg, Grad-CAM), feature attribution approaches (eg, SHAP), and local approximation methods (eg, LIME) across various dental specialties, quality assessment exposed significant limitations. Most (14 of 19) exhibited a high risk of bias, driven by small retrospective datasets, lack of external validation, and weak reference standards. Interestingly, only 1 study has evaluated the impact of XAI on human understanding. Conclusion Current dental XAI research remains in a proof-of-concept phase, characterised by technical demonstrations based on low-quality evidence. The field has not yet substantiated claims that XAI enhances clinical trust or decision-making. To bridge this gap, future work must prioritise methodological rigour, external validation, and, most importantly, human-centred evaluations with dental professionals to measure the true impact of explainability on clinical workflows and patient care. Clinical Relevance Current dental XAI lacks the evidence quality required for clinical reliance. Practitioners should exercise caution, as these tools have not been proven to actually improve diagnostic accuracy or trust in daily practice. Until validated in real-world settings, XAI remains experimental technology rather than a standard of care. Key words: Explainable AI, Interpretable machine learning, Dentistry, Systematic review, Trust, Clinical adoption, QUADAS-2, PROBAST Introduction Artificial intelligence (AI) is poised to revolutionise dentistry, offering advancements in diagnostic imaging, treatment planning, and outcome prediction. 1 , 2 , 3 However, the clinical integration of these technologies, particularly complex deep learning models, is hampered by their so-called black box nature. When an AI system provides a diagnosis without a transparent rationale, it undermines clinician trust, complicates regulatory approval, and poses ethical challenges for patient care. 1 , 4 , 5 , 6 Explainable artificial intelligence (XAI) has therefore emerged as a critical field aimed at making AI decisions interpretable to humans. 7 , 8 A suite of techniques, from saliency maps like Grad-CAM (a gradient-based heatmap method for imaging) for imaging to model-agnostic methods like SHAP (feature attribution based on game theory) and LIME (local approximation of model behaviour) for tabular data, promises to open these “black boxes”. 9 , 10 In a clinical context, the purported value of XAI is not merely technical transparency but the fostering of trust, allowing a clinician to verify, understand, and confidently act upon an AI’s recommendation. 11 , 12 The expanding application of XAI to dental problems is clear, yet the mere provision of an explanation is an insufficient foundation for clinical trust. An explanation that rationalises a flawed or biased prediction is not just unhelpful; it is potentially dangerous. Consequently, the pivotal issue is not the presence of XAI but the quality of its implementation within methodologically robust research. From a clinical perspective, this translates to a single, pragmatic concern: “Was the AI system developed and validated with the rigour necessary to make its explanations both reliable and actionable for dental professionals?” This systematic review moves beyond cataloguing applications to critically appraise the evidence base for XAI in dentistry. We synthesise the current landscape but focus our analysis on a central thesis: the path to trustworthy clinical AI in dentistry requires more than just technical explanations; it demands that those explanations be built upon methodologically rigorous and clinically evaluated models. To operationalise this objective, established methodological frameworks were applied with distinct and complementary roles. Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were used to ensure transparent, reproducible, and unbiased reporting of the review process itself. 11 QUADAS-2 was applied to diagnostic accuracy studies to evaluate whether XAI explanations were grounded in diagnostically valid models with appropriate patient selection, reference standards, and evaluation pathways. 13 PROBAST was used to assess prediction model studies, allowing examination of whether XAI-supported predictions were built upon models with acceptable risk of bias, robustness, and clinical applicability. 14 , 15 Together, these frameworks were employed not as procedural requirements but as analytical tools to identify the gap between technically explainable AI outputs and the level of evidence required for meaningful clinical trust and adoption. Methods This review was conducted in accordance with the PRISMA 2020 statement. 16 A completed PRISMA 2020 checklist is provided in Supplementary File 1 . The protocol was designed to answer the following question: “What is the state of evidence for Explainable AI in dental diagnostics, and to what extent do current studies demonstrate the methodological rigour necessary for fostering clinical trust?” The protocol was submitted to PROSPERO (CRD420251182324) ( https://www.crd.york.ac.uk/PROSPERO/view/CRD420251182324 ). Eligibility criteria We included original studies that: 1. Applied a XAI technique (eg, saliency maps, Grad-CAM, SHAP, LIME, attention mechanisms, rule-based systems) or provided proxy explainability in the form of AI-generated, clinician-interpretable visual or textual cues directly linked to individual predictions 2. Focused on a dental or oral health diagnostic task 3. Were published in English between January 1, 2015, and August 30, 2025 We excluded reviews, editorials, conference abstracts without full text, and studies where the AI was not directly applied to a dental context or where explainability was not a central component. Definition and taxonomy of explainability in dental AI To ensure a consistent and transparent classification of explainability across the included studies, a priori taxonomy was established and applied during study selection and data extraction. This framework draws from established conceptual models in the XAI literature, 7 , 17 categorising systems into 2 primary groups: XAI and non-XAI. XAI (recognised and proxy): this category includes AI systems that provide supplementary information to support clinician understanding of model outputs. It encompasses recognised approaches that offer direct insight into model decision-making, such as saliency-based visualisations, feature attribution methods, and inherently interpretable models, 7 , 18 as well as systems that provide clinician-interpretable cues at the user-interface level, including visual highlights, bounding boxes, or textual markers linked to individual predictions. While interface-level cues may not be derived from explicit model attribution mechanisms, they were classified as proxy explainability because they contextualise model outputs and facilitate clinical interpretation without revealing the underlying reasoning of the model. 9 , 19 AI systems that provide only prediction outputs, such as binary classifications or probability scores, without accompanying explanatory cues, visualisations, or clinician-facing justifications, are classified as nonexplainable AI systems. 7 These models do not provide insight into the factors underlying their predictions and were therefore considered opaque from an explainability perspective. Information sources and search strategy A systematic search was performed in PubMed, IEEE Xplore, medRxiv, and Ovid. The search strategy was developed to capture the intersection of XAI, health care, and dentistry. The core query used was: ("explainable ai" or "interpretable machine learning" or "XAI") and ("healthcare" or "medical" or "clinical" or "diagnosis") and (dentistry) and (dental science) The full search strategy for each database is provided in Supplementary File 2 . Searches were executed on August 30, 2025. The search period was restricted to publications from 2015 onwards to ensure the inclusion of contemporary AI and XAI methodologies relevant to current dental clinical practice. Study selection The study selection process adhered to the PRISMA 2020 directives 11 ( Figure ) and followed a 4-stage process: identification, screening, eligibility, and inclusion. After deduplication, the titles and abstracts were independently screened by 2 reviewers (S.T. and V.T.) against the eligibility criteria. Full-text articles of potentially eligible studies were then assessed for inclusion. Any disagreements were resolved through discussion, and when consensus could not be reached, a third reviewer (T.P.) was consulted for final adjudication. Fig. Open in a new tab Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 flow diagram of the literature search and screening process. This diagram illustrates the flow of records through the systematic review process in accordance with PRISMA 2020. A total of 100 records were identified from PubMed (n = 14), IEEE Xplore (n = 12), medRxiv (n = 66), and Ovid (n = 8). After screening and eligibility assessment, 6 review articles and 1 survey study were excluded from PubMed, 6 studies without relevant keywords and 1 review article were excluded from IEEE Xplore, 59 manuscripts lacking XAI-related terms or proxy terms were excluded from medRxiv, and 5 studies without relevant keywords and 3 review articles were excluded from Ovid. Nineteen manuscripts were included in the final synthesis. XAI, explainable artificial intelligence. Data extraction Data extraction was independently performed by 2 reviewers (S.T. and V.T.) using a predefined extraction framework, with disagreements resolved by consensus or adjudication by a third reviewer (T.P.) when necessary. To ensure alignment with methodological standards and the objectives of this review, the data extraction framework was developed by integrating PRISMA guidance with recommendations from recent reviews in explainable/dental AI research. 18 , 20 , 21 , 22 For each included study, the following information was systematically extracted as follows: 1. Bibliographic information : author(s), year of publication, country of origin, and publishing journal or preprint platform 2. Study characteristics : study design (eg, retrospective, prospective, cross-sectional, experimental), sample size, and dataset type 3. Clinical domain : the clinical field addressed (eg radiology, oral pathology, orthodontics) 4. AI model : the machine learning or deep learning architecture employed (eg, convolutional neural network [CNN], transformer, ensemble methods, support vector machine) 5. Explainable AI methodology : the XAI techniques applied, including saliency maps (eg, Grad-CAM, heatmaps), feature attribution approaches (eg, SHAP, LIME), rule-based interpretable models, or other transparency-enhancing methods 6. Evaluation metrics : the performance outcomes reported, such as accuracy, area under the curve (AUC), sensitivity, specificity, F1-score, or interpretability and usability measures 7. Clinical relevance : the reported role of XAI within the dental or clinical context, including contributions to diagnosis, treatment planning, clinical decision support, or patient communication 8. Study limitations : author-identified challenges such as dataset size, validation strategy, generalisability, or interpretability constraints The extraction framework was developed to ensure that the collected data were pertinent to the review’s primary objectives. Quality assessment The methodological quality and risk of bias of included studies were critically appraised using tailored criteria from QUADAS-2 14 for diagnostic accuracy studies and PROBAST 23 for prediction model studies. Studies were evaluated across key domains: participant selection, predictor and outcome definition, analysis, and overall risk of bias. This assessment was central to our synthesis, moving beyond performance metrics to evaluate the foundational trustworthiness of the evidence. Results Study selection and characteristics The PRISMA flow diagram ( Figure ) illustrates the screening process. From 100 initial records, 19 studies met the inclusion criteria. The included studies, 24-42 summarised in Table 1 , spanned a wide range of dental specialties, including cariology, periodontology, oral pathology, orthodontics, and forensic dentistry. The excluded studies, along with the reasons for exclusion, are summarised in Supplementary Table 1 . Table 1. Summary of included studies with condensed extraction (n = 19). Bibliographic information Study characteristics Dental domain AI model XAI method Evaluation metrics Clinical relevance Adnan et al 24 Validation study; 7465 intraoral photos, Aga Khan University Cariology YOLOv5s CNN, DeTR EigenCAM Precision 90.7%, sensitivity 85.6%, F1 88.0% Smartphone app detects caries comparably to junior dentists; scalable tool for low-resource settings Dai et al 27 Longitudinal; VicGen cohort (134 children), bariatric cohort (120 adults) Paediatric dentistry LP-Micro (ensemble RF, XGBoost, SVM, NN, LSTM, GRU, CNN-GRU) Permutation importance (PermFit) Improved vs baselines; Streptococcus mutans key ECC predictor Highlighted microbial biomarkers for ECC prevention Ikeda et al 38 Prospective longitudinal; 61,883 older adults, JAGES study Oral health/geriatric dentistry RF, XGBoost, logistic regression SHAP Accuracy 0.88, F1 0.89, AUC 0.88 Identified oral health factors (teeth number, dry mouth) as fall predictors Dai et al 28 Longitudinal microbiome (VicGen, bariatric cohorts) Paediatric dentistry Lasso, XGBoost, RF, SVM, NN (LSTM, GRU, CNN-GRU) Polynomial group lasso, permutation importance AUC up to 0.74, accuracy 70% Critical taxa/timepoints for ECC prediction Parola et al 34 Experimental; 567 oral lesion images Oral cancer YOLOv8, DETR, Faster R-CNN, ResNet, ensembles Informed DL, case-based reasoning mAP@50=0.732, accuracy 85.4% vs 77.7% Explainable oral cancer screening robust to noisy images Farook et al 39 In vitro; 56 cavity preps (mandibular molars), ICDAS 3-6 Restorative dentistry 3D-CNN Saliency maps Accuracy 94.4%, DSC >0.90 3D scans classified cavity depth reliably Kayadibi et al 31 Retrospective; 1317 panoramic radiographs (UESB dataset) Radiology (third molar detection) CNN (E-mTMCNN, GoogLeNet, ResNet18, MobileNetV2, etc) LIME Accuracy 87%, AUC 0.87 Accurate third molar detection for treatment planning Dangsungnoen et al 40 User study, 24 participants (dental + data science students) Forensic dentistry (sex prediction) DeepToothDuo (EfficientNetB0), YOLOv5 OPG-SHAP + Gemini RAG explanations Improved understanding (7.29), trust (3.85) Dual-modality XAI enhanced comprehension/trust in sex prediction Angelone et al 37 Prospective; 79 patients, intraoral photos (plaque ROIs) Periodontology RF, NB, SVM, KNN, NNET, CART, LDA Partial dependence plots Accuracy up to 80%, specificity 100% Explainable classification of moderate vs severe periodontitis Kamran et al 25 Dataset: 1593 periapical/panoramic x-rays Cariology /hypodontia Vision Transformer (ViT) LIME, SHAP, Grad-CAM Accuracy 96.6%, F1 96.6% High-performance caries/hypodontia detection with interpretability Milani et al 32 CBCT (SOS staging), lateral cephalometric (malocclusion) Orthodontics ResNet18/34/50, ConvNeXt (+Attn) Gradient attention maps, Grad-CAM Accuracy up to 90.9%, AUC 93.5% Verification of model attention alignment for orthodontics Taskin and Al Islam et al 35 Dataset: 12,320 images (6 oral disease classes) General dentistry MobileNetV2, DenseNet169, ResNet50V2, InceptionV3 Grad-CAM, LIME Accuracy 89.0%, F1 89.1% Robust interpretable multiclass oral disease diagnosis Zhu et al 41 NHANES data, >25,000 adults, nonradiographic multimodal Periodontology LR, RF, HGBT, SVM, 3-layer MLP Feature importance, misclassification analysis AUC 0.81, F1 0.74 Home-based periodontitis screening tool Rivera et al 42 Secondary dataset; 233 composites, 17 attributes, 7 outcomes Dental materials KNN, DT, RF, LR, SVM, XGBoost, Voting Regressor Feature importance Accuracy up to 0.90, ROC AUC 0.97, R ² 0.93 Supports optimisation of restorative composite design Pham 29 58 paediatric panoramic radiographs Paediatric dentistry VLM (CNN + transformer) Auto-generated text + saliency features Accuracy 90%, AUC 0.96 Feasible multimodal diagnosis with VLMs Pham 30 70 train + 29 test paediatric panoramic radiographs Paediatric dentistry 1D-CNN, LSTM, BERT; pretrained CNNs Proxy XAI: text descriptions via ChatGPT 1D-CNN accuracy 84%, BERT 76.7%, LSTM 56.7% Text-based classification improved interpretability Motmaen et al 36 Retrospective; 1,184 PANs (26,956 teeth) Oral surgery ResNet-50 CAMERAS (activation mapping) ROC-AUC 0.901, PR-AUC 0.749 AI > dentists for extraction prediction Devlin et al 26 RCT; 23 dentists, 24 bitewing radiographs Cariology /radiology AssistDent ML prompts On-screen saliency prompts Sensitivity 75.8% vs 44.3%, specificity 85.4% vs 96.3% AI significantly improved enamel-only caries detection Lee et al 33 Retrospective; 151 cephalometric cases Oral and maxillofacial surgery LLMs (LLAMA-2, GPT-3.5/4, Gemini-Pro) Prompt engineering, structured text conversion Balanced accuracy up to 67.7%, F1 up to 67.5% LLMs feasible for cephalometric diagnosis Open in a new tab Abbreviations: 1D, 1-dimensional; 3D, 3-dimensional; AI, artificial intelligence; AUC, area under the curve; BERT, bidirectional encoder representations from transformer; CART, classification and regression trees; CBCT, cone-beam computed tomography; CNN, convolutional neural network; DETR, detection transformer; DL, deep learning; DSC, dice similarity coefficient; DT, decision tree; ECC, early childhood caries; E-mTMCNN, explainable mandibular third molar convolutional neural network; GRU, gated recurrent unit; HGBT, histogram-based gradient boosting tree; ICDAS, international caries detection and assessment system; KNN, k-nearest neighbor; LDA, linear discriminant analysis; LIME, local interpretable model-agnostic explanations; LLM, large language model; LP-Micro, longitudinal microbiome-based interpretable machine-learning framework; LR, learning rates; LSTM, long short-term memory; ML, machine learning; MLP, multilayer perceptron; NB, naive bayes; NHANES, National Health and Nutrition Examination Survey; NN, neural networks; NNET, neural networks; OPG, orthopantomogram; PAN, panoramic radiograph; PR-AUC, panoramic radiography area under the curve; RAG, retrieval-augmented generation; R-CNN, region-based convolutional neural network; RCT, randomized controlled trial; RF, random forest; ROC, receiver operating characteristic; ROI, region of interest; SHAP, shapley additive explanations; SOS, spheno-occipital synchondrosis; SVM, support vector machine; UESB, universidade estadual do sudoeste da bahia; VLM, vision language model; XAI, explainable artificial intelligence; XGBoost, extreme gradient boosting. CNNs were the predominant AI architecture. The most common XAI techniques were Grad-CAM, SHAP, and LIME. The main study characteristics are summarised in Table 1 , while complete details, including dataset properties, performance metrics, and XAI methods, are available in Supplementary Table 2 . Cariology In cariology, several studies focused on caries detection from radiographs or photographs. Adnan et al 24 validated YOLOv5s- and DeTR-based detectors of 7465 intraoral images, demonstrating high precision (90.7%) and an F1-score (88.0%) for caries detection via a smartphone application. Kamran et al 25 used Vision Transformer models with SHAP, LIME, and Grad-CAM to detect caries and hypodontia from periapical and panoramic radiographs, achieving 96.6% accuracy. Devlin et al 26 conducted a randomised controlled trial using AssistDent, reporting a marked increase in sensitivity for detecting enamel-only caries compared to dentists without AI support. Paediatric dentistry In paediatric dentistry, Dai et al 27 , 28 used ensemble machine learning and longitudinal microbiome data from the VicGen and bariatric cohorts to identify microbial biomarkers predictive of early childhood caries. Pham 29 , 30 investigated paediatric panoramic radiographs using multimodal and text-based methods. A vision-language model combining CNN and transformer layers achieved 90% accuracy and an AUC of 0.96, while proxy-based XAI models using text descriptions demonstrated that a 1-dimensional CNN outperformed pretrained CNNs and a pretrained bidirectional transformer language model (BERT) in interpretability for binary disease classification. Radiology and imaging Radiology and imaging studies were also prominent. Kayadibi et al 31 trained CNN architectures (eg, ResNet18, MobileNetV2) of 1317 panoramic radiographs for third molar detection, applying LIME to provide feature explanations and achieving 87% accuracy. Milani et al 32 applied ResNet and ConvNeXt models to cephalometric and cone-beam computed tomography (CBCT) datasets, integrating attention maps and Grad-CAM to verify orthodontic staging with an AUC of up to 93.5%. Lee et al 33 explored the feasibility of large language models (LLMs), including GPT-4 and Gemini-Pro, for cephalometric diagnosis, reporting balanced accuracies of 67.7%. Oral pathology and cancer detection In oral pathology and cancer detection, Parola et al 34 evaluated YOLOv8, DETR, and Faster R-CNN models with case-based reasoning for oral lesion recognition, achieving a mAP@50 of 0.732. Taskin and Al Islam 35 developed an interpretable multiclass model (MobileNetV2, DenseNet, ResNet, Inception) to classify 6 oral diseases with 89% accuracy. Motmaen et al 36 trained a ResNet-50 on 26,956 teeth from 1184 panoramic radiographs to predict tooth extraction decisions; their model, which used CAMERAS for activation mapping, outperformed dentists (receiver operating characteristic AUC = 0.901). Other dental specialties Other dental specialties were also included in our review. Angelone et al 37 applied interpretable machine learning (ML) models, with partial dependence plots, to intraoral plaque biofilm images, achieving 80% accuracy for classifying periodontal disease severity. Ikeda et al 38 leveraged SHAP to explain risk factors for falls among 61,883 older adults, demonstrating the clinical utility of oral health data in geriatric dentistry. Farook et al 39 developed a 3-dimensional CNN with saliency maps to classify cavity depth in restorative dentistry, achieving 94.4% accuracy. Dangsungnoen et al 40 proposed DeepToothDuo with SHAP and retrieval-augmented explanations for forensic sex prediction, which improved participant trust and understanding. Other reviewed studies included those of Zhu et al, 41 who trained multimodal ML models on National Health and Nutrition Examination Survey data to predict periodontitis with an AUC of 0.81, suggesting a pathway to home-based screening, and Rivera et al, 42 who examined dental composite performance using feature importance from multiple ML models to support material optimisation. Together, these studies demonstrate the diversity of XAI applications in dentistry, spanning diagnostic imaging, preventive care, restorative materials, and clinical decision support. Synthesis of AI and XAI methodological trends A clear pattern has emerged in the pairing of XAI techniques with AI models and data types. Saliency-based methods (Grad-CAM, heatmaps) were almost exclusively applied to CNN-based image analysis. 24 , 25 , 31 , 32 , 35 , 36 , 43 In contrast, feature attribution methods like SHAP were typically used with tree-based models or tabular data. 27 , 28 , 38 , 41 , 42 This reflects a technically appropriate but largely formulaic application of XAI. Only 1 study by Dangsungnoen et al 40 explicitly designed and evaluated a multimodal explanation system (OPG-SPELL) with end users, measuring its impact on understanding and trust. Critical appraisal of methodological rigour and risk of bias The quality assessment of the evaluated studies reveals a field in its infancy, grappling with fundamental methodological challenges. Most studies (14 of 19) were judged to have a high overall risk of bias as discussed below. The methodological quality of the 19 included studies was appraised using QUADAS-2 for diagnostic accuracy and PROBAST for prediction model studies. Across the body of evidence, the risks of bias and applicability concerns varied considerably depending on study design, population, and reference standards. The domain-level assessments are summarised in Table 2 , with complete details available in Supplementary Table 3 . Table 2. Quality assessment of included studies (n = 19). Bibliographic information QUADAS-2 * PROBAST † Overall risk of bias ‡ Patient selection Index test Reference standard Flow and timing Applicability concerns Participants Predictors Outcomes Analysis Adnan et al 24 High Low High Low High High Low High High High Dai et al 27 Low Low Low Low Fair High Low Low High Fair Ikeda et al 38 Low Low Low Low Low Low Low High Low Fair Dai et al 28 Low Low Low Low Low Low Low Low Fair Low Parola et al 34 High Fair High Fair High High Fair Fair High High Farook et al 39 High Low Low Low High High Low Low High High Kayadibi et al 31 Low Low Low Limited Low High Low Low High Fair Dangsungnoen et al 40 High Low High Low High High Low High High High Angelone et al 37 High Low High Low Fair High Low High High High Kamran et al 25 High Low High High High High Low High High High Milani et al 32 High Low Fair Limited Fair High Low Fair Fair Fair Taskin and Al Islam et al 35 High Low High Limited High High Low Fair High High Zhu et al 41 Low Low High High Low Low Low High Low Fair Rivera et al 42 High Low High High High High Low Low High High Pham 29 High Low High Limited Fair High Low Fair High High Pham 30 High Low Low High High High Low Low High High Motmaen et al 36 High Low Fair Low Fair High Low Fair High High Devlin et al 26 Low Low Low Low Low Low Low Low Low Low Lee et al 33 High High High High Limited High High High High High Open in a new tab ⁎ For all QUADAS-2 assessment. “Low” means low risk of bias with clearly defined methodology and appropriate reference standards; “Fair” means some methodological concerns with potential bias but not likely to alter results substantially; “High” means high risk of bias due to major flaws such as limited sample size, weak reference standards, or lack of external validation; and “Limited” means insufficient information reported to permit a clear judgement. † For all PROBAST assessment. “Low” indicates low risk of bias with adequately described participants, predictors, outcomes, and analyses; “Fair” indicates some concerns, generally robust but with restricted generalisability; and “High” indicates high risk of bias due to methodological shortcomings, poor representativeness, or risk of overfitting. ‡ For overall risk of bias assessment. “Low” means low concern across all domains; “Fair” means some concerns that may limit generalisability but unlikely to invalidate findings; “High” means substantial concerns likely to compromise validity. Several studies were judged to have a high overall risk of bias, primarily due to small, single-centre datasets and nonstandard outcome definitions. For instance, Adnan et al 24 and Parola et al 34 relied on convenience samples and nonexpert or imperfect reference standards, raising concerns in the outcome domains of both QUADAS-2 and PROBAST. Similarly, Farook et al 39 and Pham 29 , 30 employed very small in vitro or paediatric datasets, which limited their external validity. Studies by Kamran et al, 25 Taskin and Al Islam, 35 Rivera et al, 42 Motmaen et al, 36 and Lee et al 33 also showed high risks, stemming from their reliance on public or retrospective datasets, a lack of external validation, or insufficiently representative participant populations. In contrast, a subset of studies demonstrated low to moderate risks of bias, often supported by larger or more representative cohorts. Dai et al 27 , 28 achieved low-risk ratings across most QUADAS-2 and PROBAST domains by employing prospective or multicohort microbiome datasets with well-defined predictors and outcomes, although concerns remained regarding limited generalisability. Ikeda et al 38 also found low risk in most domains, although the use of self-reported outcomes introduced moderate concerns. Zhu et al 41 leveraged a large, nationally representative dataset (National Health and Nutrition Examination Survey) that was rated low risk in participant and predictor domains; however, its outcome definitions did not align with gold-standard periodontal diagnoses. Other studies presented methodologically balanced designs but were limited in scope. Kayadibi et al 31 and Milani et al 32 featured clearly defined predictors and robust internal validation but were confined to single-institution data and lacked external testing. Similarly, Angelone et al 37 and Dangsungnoen et al 40 showed innovative applications of XAI but were hampered by small sample sizes and nonstandard outcome measures. Conversely, the ADEPT study 26 achieved an overall low risk of bias, benefitting from its randomised design, expert gold-standard comparators, and consistent participant assessment, despite a modest sample size. Overall, most studies were hampered by concerns regarding external validity, dataset representativeness, and a lack of standardised reference standards. While a minority demonstrated robust methodological practices with low risks of bias, most presented high risks in at least 1 domain, particularly in outcome assessment and external validation. Patient selection : A predominant issue was the use of small, retrospective, single-centre datasets, often from public repositories, leading to high concerns about representativeness and generalisability. Reference standard and outcomes : Many studies used nonexpert annotations or imperfect reference standards, casting doubt on the validity of the “ground truth” used to train and evaluate the models. Analysis : A near-universal limitation was the absence of external validation. Models were typically trained and tested on data from the same source, creating a high risk of overfitting and severely limiting claims of generalisability. Only a single study, the ADEPT randomised controlled trial, 26 was rated as having a low risk of bias across all domains, demonstrating the feasibility of robust study design in this field. Discussion This systematic review synthesises the current landscape of XAI in dentistry and delivers a critical message: while the technical application of XAI is growing, the evidence base for its ability to foster genuine clinical trust remains weak. Our analysis, anchored by a formal quality assessment, identifies several critical themes that must be addressed for the field to advance. In general, we submit that there was a deficit of methodological rigour in most of the 19 studies reviewed. CNNs were the most frequently adopted architecture, with interpretability commonly enabled through saliency-based visualisation tools such as Grad-CAM and heatmaps. 43 , 44 Other approaches, including model-agnostic techniques such as SHAP and LIME and attention mechanisms in transformer-based models, have also been reported to support interpretability of predictions derived from both imaging and structured data. By appraising these studies, this review demonstrates that while XAI has begun to permeate dental AI research, its implementation is uneven and frequently limited to proof-of-concept applications. Some investigations provided robust interpretability frameworks that aligned computational outputs with clinical reasoning, thereby supporting clinician trust in diagnostic decisions. 26 , 27 , 28 , 38 , 41 Others offered only partial or rudimentary explanations, raising concerns about transparency and reproducibility. 24 , 25 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 39 , 40 , 42 Synthesising these findings allows us to assess not only the extent of XAI adoption but also the degree to which current evidence addresses the core challenge of trustworthiness in dental AI. In doing so, the review fulfils its objective of identifying, appraising, and critically evaluating XAI applications in dentistry to determine whether explainability meaningfully enhances trust in diagnostic decision-making. The most significant finding of this review is the pervasive high risk of bias in the included studies. This has a direct and profound implication for trustworthiness. An explanation, no matter how visually compelling or intuitively appealing, is meaningless if it is explaining the behaviour of a flawed or nongeneralisable model. For instance, a Grad-CAM heatmap highlighting a specific region on a radiograph provides a false sense of security if the underlying model was trained on a small, biased dataset and never validated externally. The dental XAI community must first conquer these foundational challenges of dataset quality, robust validation, and outcome design before the nuanced benefits of explainability can be realised. Trust in clinical AI is built on proven reliability, for which transparency is a complement, not a substitute. The chasm between technical and clinical explainability There is a stark disconnect between the technical implementation of XAI and its clinical evaluation. The literature is dominated by post hoc explanations applied to black box models, with the validity of the explanation often assumed rather than tested. Very few studies ventured beyond technical validation to ask: Does this explanation actually help a dentist? Does a SHAP value or a LIME plot improve diagnostic confidence? Does it reduce error rates? Does it facilitate communication with a patient? The study by Dangsungnoen et al 40 is a notable exception, demonstrating that designed explanations can improve user understanding. The field must shift from demonstrating that an AI model can be explained to proving that the explanation is useful in a clinical workflow. This requires a new paradigm of human-centred design and evaluation, involving dental professionals as end-users throughout the development process. The unexplored potential of inherently interpretable models The current dental XAI landscape is overwhelmingly focused on explaining complex models (“post hoc explainability”). The debate in the wider XAI literature, however, questions whether this approach can ever be fully trustworthy. 45 An alternative path is to use inherently interpretable models (eg, decision trees, linear models, rule-based systems) that are transparent by design. While these may sometimes sacrifice marginal performance, their decisions are inherently understandable. With the exception of a few studies using models such as random forests or logistic regression with SHAP, 37 , 38 , 41 this approach is underexplored in dentistry. For many clinical tasks, the high-stakes nature of decision-making may favour the certifiable trust of an interpretable model over the uncertain explanation of a more complex one. The findings of this review suggest that while XAI methods are increasingly incorporated into dental diagnostic research, their readiness for clinical adoption remains limited. Across the 19 included studies, saliency-based methods such as Grad-CAM and heatmaps were the most frequently applied, often serving to visually justify CNN predictions. Other studies adopted feature attribution techniques, such as SHAP or LIME, or attention-based mechanisms embedded within transformers. These strategies provide an important first step towards transparency, offering clinicians insight into why predictions are made. However, the evidence presented remains largely experimental: most studies relied on retrospective datasets with limited diversity and modest sample sizes, and only a minority reported external validation. Equally important, few studies explicitly evaluated clinician-centred outcomes, such as whether XAI explanations improved diagnostic confidence, reduced error rates, or enhanced communication with patients. Without prospective, real-world testing and without clear demonstration that interpretability meaningfully improves decision-making or workflow efficiency, the ability of current XAI implementations to enhance trust and support routine adoption remains uncertain. In the wider medical literature, several reviews have directly addressed barriers to the clinical adoption of XAI. Tonekaboni et al 19 emphasised that interpretability must be evaluated in terms of its impact on clinical decision-making and patient outcomes rather than technical plausibility alone. Amann et al 17 highlighted regulatory, ethical, and usability challenges that hinder XAI integration into health care workflows, while Rudin et al 46 argued that inherently interpretable models may offer stronger trust foundations than post hoc explanation tools. Compared to these health care–focused reviews, the current dental XAI landscape remains at an earlier stage: none of the dental studies included incorporated prospective clinical validation, usability testing with practitioners, or patient-centred measures of interpretability. This gap underscores the need for future dental research to move beyond proof-of-concept demonstrations and adopt evaluation frameworks that explicitly measure whether XAI explanations improve clinical adoption and trust in practice. Limitations of the review Our review has a few limitations. The restriction to English-language studies and the specific databases searched may have omitted relevant work. The rapid evolution of AI means that some very recent advances may not be captured. Furthermore, the application of QUADAS-2 and PROBAST, while appropriate, requires adaptation for some AI-specific studies, which may introduce subjectivity. Future directions This review demonstrates that the application of XAI in dentistry is widespread but shallow. The field is currently characterised by technically focused proofs of concept that are built upon methodologically weak foundations. To transition from a promising concept to a trusted clinical tool, we propose a concerted shift in research priorities such as universal adoption of standardised reporting measures, prospective validation, and real-world integration, focusing on trust, reproducibility, and clinically grounded evaluation. Prioritising robustness : Future studies must adhere to higher methodological standards. This includes using larger, prospectively collected, and diverse datasets; implementing rigorous external validation; and employing clinically accepted reference standards. The reporting of studies should be standardised to include key details on data provenance, model selection, and hyperparameter. Development of standardised metrics : Research organisations such as the International Association for Dental, Oral, and Craniofacial Research (IADR) and nongovernmental organisations akin to the International Dental Federation must, in partnership, develop and adopt standardised metrics and frameworks for evaluating XAI’s impact on human users. This involves moving beyond accuracy and AUC to measure outcomes such as diagnostic confidence, time to decision, error detection, and user satisfaction through controlled studies with dentists. Human-centred design : XAI systems should be codesigned with clinicians to ensure that the explanations provided are relevant, intuitive, and integrated into the clinical workflow. Research should explore what information dentists need, in what format, and at what point in the decision-making process. Human-centred evaluation remains underexplored in current XAI research. Future studies should incorporate structured comparisons between explainable AI systems and human examiners to better contextualise model explanations, clinical trust, and decision-making relevance. Reconsidering the model paradigm : Researchers should critically evaluate whether a complex, post hoc explained model is always necessary. For many tasks, inherently interpretable models may provide a more direct and trustworthy path to clinical adoption. In other words, very often a simpler, easier-to-understand model is the smarter and more trustworthy choice for a clinical setting. Conclusion In conclusion, the journey toward trustworthy AI in dentistry is a marathon, not a sprint. The widespread but often superficial application of XAI techniques is the first step. The next, more challenging step is to build these techniques upon an unshakeable foundation of methodological rigour and to prove their value not just in silico, but in the hands of the clinicians they are meant to serve. Author contributions S.T., V.T.: investigation, data analysis, and drafting the manuscript. H.A.R, L.S., and T.O.: data analysis and edited the manuscript. T.P.: study conception, data acquisition and interpretation, drafted the manuscript. All authors critically revised the manuscript and gave their final approval for publication. Conflict of interest The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Data availability statement Original data generated and analysed during this study are included in this published article or in the supplementary material. Acknowledgments Funding This research is funded by Health Systems Research Institute (69-143, 69-145), Faculty of Dentistry (DRF69_015), and Thailand Science Research and Innovation Fund Chulalongkorn University (HEA_FF_69_036_3200_003). SermpornThaweesapphithak was supported by the Faculty of Dentistry, Chulalongkorn University. Lakshman Samaranayake and Hamid Alinejad Rokny were partly supported by the Second Century Fund (C2F), Chulalongkorn University. Acknowledgements L.S. was supported by the Chulalongkorn Umiversity Second Century (C2) high potential professoriate fund at its Faculty of Dentistry, Bangkok, Thailand. Footnotes Supplementary material associated with this article can be found in the online version at doi:10.1016/j.identj.2026.109626 . Appendix. Supplementary materials Supplementary File 1: PRISMA 2020 checklist mmc1.docx (270.1KB, docx) Supplementary File 2: Database search strategy mmc2.docx (15.5KB, docx) Supplementary Table 1: Full-text articles excluded after eligibility assessment and reasons for exclusion. mmc3.docx (30.7KB, docx) Supplementary Table 2: Detailed data extraction of included studies. mmc4.docx (49.3KB, docx) Supplementary Table 3: Detailed explanations for quality assessment of included studies. mmc5.docx (52.3KB, docx) References 1. Rewthamrongsris P., Burapacheep J., Trachoo V., Porntaveetus T. Accuracy of large language models for infective endocarditis prophylaxis in dental procedures. Int Dent J. 2025;75(1):206–212. doi: 10.1016/j.identj.2024.09.033. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Marya A., Selvaraj S., Okazaki K., Wang D.H., Kanetaka H., Porntaveetus T. The role of extended reality in orthodontic treatment planning and simulation—a scoping review. Int Dent J. 2025;75(6) doi: 10.1016/j.identj.2025.103855. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Termteerapornpimol K, Kulvitit S, Prommanee S, Khurshid Z, Porntaveetus T. Comparative benchmark of seven large language models for traumatic dental injury knowledge. Eur J Dent. 2025, [Epub ahead of print]. [ DOI ] [ PMC free article ] [ PubMed ] 4. Khurshid Z., Faridoon F., Silkosessak O.C., et al. Multi-regional deep learning models for identifying dental restorations and prosthesis in panoramic radiographs. BMC Oral Health. 2025;25(1):1750. doi: 10.1186/s12903-025-07138-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Hennocq Q., Bongibault T., Marlin S., et al. AI-based diagnosis in mandibulofacial dysostosis with microcephaly using external ear shapes. Front Pediatr. 2023;11 doi: 10.3389/fped.2023.1171277. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Samaranayake L., Tuygunov N., Schwendicke F., et al. The transformative role of artificial intelligence in dentistry: a comprehensive overview. Part 1: fundamentals of AI, and its contemporary applications in dentistry. Int Dent J. 2025;75(2):383–396. doi: 10.1016/j.identj.2025.02.005. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Barredo Arrieta A., Díaz-Rodríguez N., Del Ser J., et al. Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI. Inform Fusion. 2020;58:82–115. [ Google Scholar ] 8. Hosseinzadeh M., Haider A., Malik M.H., et al. Enhanced heart sound classification using Mel frequency cepstral coefficients and comparative analysis of single vs. ensemble classifier strategies. PLoS ONE. 2024;19(12) doi: 10.1371/journal.pone.0316645. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Holzinger A., Langs G., Denk H., Zatloukal K., Müller H. Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 2019;9(4) doi: 10.1002/widm.1312. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Hosseinzadeh M., Haider A., Rahmani A.M., et al. SDN-based NFV deployment for multi-objective resource allocation in edge computing: a deep reinforcement learning for IoT workload scheduling. Sustainable Comput Informatics Syst. 2025;48 [ Google Scholar ] 11. Page M.J., McKenzie J.E., Bossuyt P.M., et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi: 10.1136/bmj.n71. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Hosseinzadeh M., Tanveer J., Ali S., et al. An energy-focused model for batteryless IoT: vortex wireless power transfer and fog computing in 6G networks. Internet Things. 2025;32:101657. [ Google Scholar ] 13. Whiting P.F., Rutjes A.W., Westwood M.E., et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011;155(8):529–536. doi: 10.7326/0003-4819-155-8-201110180-00009. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Wolff R.F., Moons K.G.M., Riley R.D., et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170(1):51–58. doi: 10.7326/M18-1376. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Hennocq Q., Willems M., Amiel J., et al. Next generation phenotyping for diagnosis and phenotype–genotype correlations in Kabuki syndrome. Sci Rep. 2024;14(1):2330. doi: 10.1038/s41598-024-52691-3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Fakhruddin K.S., Kamal I., Maglaperidze T., Marya A., Samaranayake L., Porntaveetus T. Autism and the oral microbiome: a systematic review of host-microbial interactions and diversity. Int Dent J. 2025;75(6) doi: 10.1016/j.identj.2025.100957. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Amann J., Blasimme A., Vayena E., Frey D., Madai V.I. the Precise Qc. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Med Informatics Decis Making. 2020;20(1):310. doi: 10.1186/s12911-020-01332-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Kalasampath K., Spoorthi K.N., Sajeev S., Kuppa S.S., Ajay K., Maruthamuthu A. A literature review on applications of explainable artificial intelligence (XAI) IEEE Access. 2025;13:41111–41140. [ Google Scholar ] 19. Tonekaboni S, Joshi S, McCradden MD, Goldenberg A. What clinicians want: contextualizing explainable machine learning for clinical end use. In: Proceedings of the 4th Machine Learning for Healthcare Conference; Proceedings of Machine Learning Research; Toronto. 2019;359–380. 20. Liu X., Faes L., Kale A.U., et al. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health. 2019;1(6):e271–ee97. doi: 10.1016/S2589-7500(19)30123-2. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Nagendran M., Chen Y., Lovejoy C.A., et al. Artificial intelligence versus clinicians: systematic review of design, reporting standards, and claims of deep learning studies. BMJ. 2020;368:m689. doi: 10.1136/bmj.m689. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Shan T., Tay F.R., Gu L. Application of artificial intelligence in dentistry. J Dent Res. 2021;100(3):232–244. doi: 10.1177/0022034520969115. [ DOI ] [ PubMed ] [ Google Scholar ] 23. Falagas M.E., Pitsouni E.I., Malietzis G.A., Pappas G. Comparison of PubMed, Scopus, Web of Science, and Google Scholar: strengths and weaknesses. FASEB J. 2008;22(2):338–342. doi: 10.1096/fj.07-9492LSF. [ DOI ] [ PubMed ] [ Google Scholar ] 24. Adnan N., Faizan Ahmed S.M., Das J.K., et al. Developing an AI-based application for caries index detection on intraoral photographs. Sci Rep. 2024;14(1) doi: 10.1038/s41598-024-78184-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Kamran M, Waseem M, Shafi N, Zahid A. In: 2024 18th International Conference on Open Source Systems and Technologies (ICOSST) IEEE; Lahore, Pakistan: 2024. Detection of caries and hypodontia using deep learning and explainable AI; pp. 1–6. [ Google Scholar ] 26. Devlin H., Williams T., Graham J., Ashley M. The ADEPT study: a comparative study of dentists’ ability to detect enamel-only proximal caries in bitewing radiographs with and without the use of AssistDent artificial intelligence software. Br Dent J. 2021;231(8):481–485. doi: 10.1038/s41415-021-3526-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Dai Y, Qian Y, Qu Y, et al. Longitudinal microbiome-based interpretable machine learning for identification of time-varying biomarkers in early prediction of disease outcomes. bioRxiv [Preprint]. 2024, 10.18.619118. Available from: https://www.biorxiv.org/content/10.1101/2024.10.18.619118 . 28. Dai Y., Qian Y., Qu Y., et al. Decoding longitudinal microbiome trajectories: an interpretable machine learning approach for biomarker discovery and prediction. Brief Bioinform. 2025;26(4):bbaf374. doi: 10.1093/bib/bbaf374. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Pham T.D. A deep learning vision-language model for diagnosing pediatric dental diseases. medRxiv 2025.05.21.25328098 [Preprint]. Available from: https://www.medrxiv.org/content/10.1101/2025.05.21.25328098v1 . 30. Pham T.D. Classification of pediatric dental diseases from panoramic radiographs using natural language transformer and deep learning models. medRxiv 2025.01.30.25321418 [Preprint]. Available from: https://www.medrxiv.org/content/10.1101/2025.01.30.25321418v1 . [ DOI ] [ PMC free article ] [ PubMed ] 31. Kayadibi İ, Köse U., Güraksın G.E., Çetin B. An AI-assisted explainable mTMCNN architecture for detection of mandibular third molar presence from panoramic radiography. Int J Med Inform. 2025;195 doi: 10.1016/j.ijmedinf.2024.105724. [ DOI ] [ PubMed ] [ Google Scholar ] 32. Milani O.H., Nikho A., Mills L., Tliba M., Cetin A.E., Elnagar M.H. In: 2025 IEEE 43rd VLSI Test Symposium (VTS) IEEE; Tempe, AZ, USA: 2025. Gradient attention map based verification of deep convolutional neural networks with application to X-ray image datasets. [ Google Scholar ] 33. Lee J, Xu X, Kim D, et al. Large language models diagnose facial deformity. medRxiv 2024.07.11.24310274 [Preprint]. Available from: https://www.medrxiv.org/content/10.1101/2024.07.11.24310274v1 . 34. Parola M., Galatolo F.A., La Mantia G., Cimino M., Campisi G., Di Fede O. Towards explainable oral cancer recognition: screening on imperfect images via informed deep learning and case-based reasoning. Comput Med Imaging Graph. 2024;117 doi: 10.1016/j.compmedimag.2024.102433. [ DOI ] [ PubMed ] [ Google Scholar ] 35. Taskin S., Al Islam F. 2024 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON) Information Technology for Health (BECITHCON); Dhaka, Bangladesh: 2024. Transfer learning-based fine tuned MobileNetV2 model with explainable artificial intelligence for identifying dental diseases. [ Google Scholar ] 36. Motmaen I, Xie K, Schönbrunn L, et al. Insights into predicting tooth extraction from panoramic dental images: artificial intelligence vs. dentists. medRxiv 2024.04.22.24306189 [Preprint]. Available from: https://www.medrxiv.org/content/10.1101/2024.04.22.24306189v2.full . [ DOI ] [ PMC free article ] [ PubMed ] 37. Angelone F., Franco A., Pisani N., et al. In: 2025 IEEE 13th International Conference on Healthcare Informatics (ICHI) IEEE; Rende, Italy: 2025. An explainable AI approach for periodontitis severity classification from intraoral images. [ Google Scholar ] 38. Ikeda T., Cooray U., Hariyama M., et al. An interpretable machine learning approach to predict fall risk among community-dwelling older adults: a three-year longitudinal study. J Gen Intern Med. 2022;37(11):2727–2735. doi: 10.1007/s11606-022-07394-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Farook T.H., Ahmed S., Rashid F., et al. Application of 3D neural networks and explainable AI to classify ICDAS detection system on mandibular molars. J Prosthet Dent. 2025;133(5):1333–1341. doi: 10.1016/j.prosdent.2024.09.014. [ DOI ] [ PubMed ] [ Google Scholar ] 40. Dangsungnoen L., Thongprasant K., Wiratchawa K., et al. OPG-SPELL: visual and textual explanations that influence understanding of the AI model for sex prediction from orthopantomograms. IEEE Access. 2025;13:146690–146706. [ Google Scholar ] 41. Zhu ZX, Wu X, Zhu L, et al. Development of a machine learning tool for home-based assessment of periodontitis. medRxiv 2025.03.10.25323689 [Preprint]. Available from: https://www.medrxiv.org/content/10.1101/2025.03.10.25323689v1 . 42. Rivera K.P., Whang K., Joshi K., Son H., Kim Y.S., Flores M. Dental composite performance prediction using artificial intelligence. J Dent Res. 2025;104(5):513–521. doi: 10.1177/00220345241311888. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Hosseinzadeh M., Ali U., Ali S., et al. Improving phishing email detection performance through deep learning with adaptive optimization. Sci Rep. 2025;15(1) doi: 10.1038/s41598-025-20668-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Goh L.H., Chong B., van der Lubbe S.C.C., et al. The epidemiology and burden of cardiovascular diseases in countries of the Association of Southeast Asian Nations (ASEAN), 1990–2021: findings from the Global Burden of Disease Study 2021. Lancet Public Health. 2025;10(6):e467–ee79. doi: 10.1016/S2468-2667(25)00087-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Raman R., Singh P., Singh V.K., Vinuesa R., Nedungadi P. Understanding the bibliometric patterns of publications in IEEE access. IEEE Access. 2022;10:35561–35577. [ Google Scholar ] 46. Cynthia R., Chaofan C., Zhi C., Haiyang H., Lesia S., Chudi Z. Interpretable machine learning: fundamental principles and 10 grand challenges. Stat Surv. 2022;16:1–85. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary File 1: PRISMA 2020 checklist mmc1.docx (270.1KB, docx) Supplementary File 2: Database search strategy mmc2.docx (15.5KB, docx) Supplementary Table 1: Full-text articles excluded after eligibility assessment and reasons for exclusion. mmc3.docx (30.7KB, docx) Supplementary Table 2: Detailed data extraction of included studies. mmc4.docx (49.3KB, docx) Supplementary Table 3: Detailed explanations for quality assessment of included studies. mmc5.docx (52.3KB, docx) Data Availability Statement Original data generated and analysed during this study are included in this published article or in the supplementary material. Articles from International Dental Journal are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (561.5 KB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Artificial Intelligence in Orthodontics: From Laboratory Benchmarks to Clinical Care.#385168
• Artificial intelligence for teaching, training, and assessment in dental education: a domain-based scoping review.#15168
• Artificial intelligence and machine learning in orthodontics#478098
• ARTIFICIAL INTELLIGENCE IN ORTHODONTICS: DIAGNOSIS, DISPUTE RESOLUTION, AND OUTCOME PREDICTION#667368
• Preparing future dentists for artificial intelligence: a cross-sectional study of perceptions and educational needs in China.#255466