Artificial Intelligence Applications in Pediatric Dentistry: A Systematic Scoping Review With Visual Evidence Mapping

ABSTRACT

Background and Aims

Artificial intelligence is increasingly applied to pediatric dental imaging, caries detection, orthodontic decision support, age estimation, and risk prediction. However, pediatric evidence remains fragmented across clinical tasks, model types, validation strategies, and reported metrics. This systematic scoping review mapped the clinical applications, methodological characteristics, and reported performance of AI in pediatric dentistry.

Methods

PubMed, Scopus, Web of Science, and Google Scholar were searched for English‐language studies published from January 2010 to May 2026. Original studies evaluating AI, machine learning, or deep learning in pediatric dentistry or pediatric‐relevant dental tasks were included. Screening was performed using Rayyan. Data extraction captured study characteristics, pediatric relevance, AI method, clinical task, outcomes, and validation approach. Risk of bias was assessed using QUADAS‐3 for diagnostic/classification studies and PROBAST + AI for prediction and model‐development studies. Findings were synthesized descriptively according to PRISMA‐ScR.

Results

Thirty‐five studies were included, comprising 19 (54.3%) pediatric‐only studies and 16 (45.7%) broader or mixed‐age studies with separately identifiable pediatric or pediatric‐relevant evidence. By primary application, 15 (42.9%) addressed orthodontic or craniofacial assessment and decision support, eight (22.9%) risk prediction or biomarker modeling, five (14.3%) diagnostic imaging or dental‐anomaly detection, five (14.3%) dental age estimation, and two (5.7%) caries or plaque detection. Validation was internal or test‐set based in 26 studies (74.3%), observer‐ or manual‐comparative in five (14.3%), and not clearly defined in four (11.4%); none reported independent external validation. For QUADAS‐3, non‐low‐risk judgments were most frequent in the analysis domain (8/19, 42.1%), whereas for PROBAST + AI they were most frequent in the participants and data‐sources domain (5/16, 31.3%).

Conclusion

AI shows promising pediatric dental applications, especially in image‐based and decision‐support tasks, but evidence remains heterogeneous, task‐specific, and limited by variable validation, incomplete reporting, and scarce external testing.

1. Introduction

Artificial intelligence (AI) has become an increasingly visible methodological approach in healthcare research, particularly in areas that rely on pattern recognition, image interpretation, risk prediction, and decision support [1, 2, 3]. In dentistry, AI and machine‐learning methods have been applied to radiographic interpretation, caries detection, cephalometric landmark identification, orthodontic assessment, dental age estimation, and prediction of oral health outcomes. These applications are especially relevant to pediatric dentistry because dental development, eruption stage, mixed dentition, patient cooperation, and age‐dependent disease patterns can complicate diagnosis and treatment planning [4, 5].

Pediatric dentistry differs from adult dental care in several important ways. Children may have limited tolerance for prolonged examination, variable cooperation during image acquisition, developing dentition, and changing craniofacial anatomy. Clinical decisions may also have long‐term consequences for occlusion, growth, function, and future orthodontic needs. For this reason, AI tools developed in adult or general dental populations cannot be assumed to perform similarly in children unless pediatric‐specific or pediatric‐relevant data are clearly reported. This is particularly important for radiographic applications, where image quality, anatomical maturity, tooth eruption, and radiation justification all require age‐sensitive interpretation [6].

Although AI research in dentistry has expanded rapidly, the pediatric evidence base remains fragmented. Existing studies vary widely in clinical task, input data type, AI model, reference standard, validation strategy, and reported performance metric. Some articles focus on image‐based diagnostic tasks, such as mesiodens detection, permanent tooth germ detection, caries identification, plaque detection, or cephalometric classification. Others address prediction‐oriented tasks, including early childhood caries risk, biomarker‐based modeling, microbiome features, or utilization patterns. Given the substantial clinical and methodological heterogeneity across the included studies, pooling model‐performance estimates across the entire evidence base was not undertaken in this review. Quantitative synthesis could, however, be considered for sufficiently homogeneous subsets of studies with comparable populations, clinical tasks, reference standards, validation designs, and outcome metrics [7].

A further challenge is that pediatric eligibility is not always clearly separated from broader dental or orthodontic datasets. In this review, studies were classified as pediatric‐only when all participants were aged 18 years or younger. The term pediatric‐relevant was reserved for studies with mixed‐age or incompletely reported populations when either (i) participants aged 18 years or younger were explicitly represented and their data or performance estimates were separately extractable, or (ii) the analyzed records represented an inherently pediatric developmental stage, such as primary or mixed dentition, tooth eruption or tooth‐germ development, dental‐age assessment within the pediatric age range, or craniofacial growth, and results for those records were separately identifiable. Adult‐only studies and mixed‐age studies without separable pediatric data were excluded; potential applicability of the AI task to pediatric practice alone was not considered sufficient for inclusion. This operational distinction, together with explicit mapping of population characteristics, validation approaches, and metric types, was used to avoid interpreting adult‐derived or internally validated performance as evidence of pediatric clinical readiness.

Previous evidence syntheses have progressively characterized this field. Khanagar et al. (2022) assessed the performance of AI models across 20 pediatric‐dentistry studies [8], whereas Mahajan et al. (2023) examined 13 predominantly diagnostic studies published through 2021 [9]. Rokhshad et al. (2024) subsequently synthesized 33 studies addressing diagnostic and treatment‐planning applications [10], and La Rosa et al. (2024) provided a broader scoping overview of AI uses in pediatric dentistry [11]. Karamüftüoğlu et al. (2026) included 32 studies, of which 15 contributed to meta‐analysis, and focused primarily on diagnostic performance in populations aged 18 years or younger. Collectively, these reviews established the breadth and promising task‐specific performance of pediatric dental AI, but their principal emphasis was on cataloging applications or synthesizing diagnostic accuracy.

The added value of visual evidence mapping lies in converting a heterogeneous body of studies into complementary, study‐level displays rather than reducing it to a narrative catalog or a single pooled performance estimate. In this review, visual evidence mapping combines a thematic map of the five clinical domains and their subdomains, an article‐level performance map that preserves the metric originally reported, and separate QUADAS‐3 and PROBAST summaries, interpreted alongside structured information on population, data type, and validation approach. This framework makes visible where evidence is concentrated, which domains depend predominantly on retrospective internal testing, where pediatric‐specific evidence remains sparse, and whether apparently strong model performance is accompanied by methodological or validation concerns. These relationships are difficult to identify from narrative lists or by combining non‐equivalent clinical tasks and outcome measures. Accordingly, the visualizations are intended as descriptive evidence‐gap tools for research prioritization, not as rankings of algorithms or estimates of comparative clinical effectiveness. The review therefore aimed to map the breadth and methodological structure of pediatric dental AI applications and identify domains requiring pediatric‐specific, external, and prospective evaluation.

2. Methods

This systematic scoping review was conducted in accordance with the PRISMA Extension for Scoping Reviews (PRISMA‐ScR). The review was designed to map the scope, clinical domains, model types, reported performance metrics, and methodological characteristics of AI applications in pediatric dentistry rather than to generate pooled diagnostic or predictive effect estimates (Figure 1) details the study selection process. This review is registered with the Open Science Framework (OSF) (10.17605/OSF.IO/E2F49).

FIGURE 1.

PRISMA 2020 flow diagram of study selection. A total of 2536 records were identified from PubMed, Scopus, Web of Science, and Google Scholar. After removal of 378 duplicate records using Rayyan‐assisted duplicate detection followed by manual verification, 2158 records underwent title and abstract screening. Of these, 1973 records were excluded because they did not meet eligibility criteria. A total of 185 full‐text reports were assessed for eligibility. During full‐text review, 150 reports were excluded for predefined reasons: non‐pediatric population or no extractable pediatric‐relevant data, no AI‐based analysis or algorithm, review or conference article without original data, insufficient methodological or performance details, or non‐English/inaccessible full text. Finally, 35 articles were included in the qualitative synthesis.

2.1. Search Strategy

A systematic literature search was conducted in PubMed, Scopus, Web of Science, and Google Scholar from January 1, 2010, to May 1, 2026. The search combined terms for artificial intelligence and pediatric dental care. The search strategy was adapted for each database using appropriate field tags, Medical Subject Headings where available, truncation, and Boolean operators. No automation filter was used during the database search. The complete database‐specific search strings are provided in Table 1.

TABLE 1.

Curated search strategies across searched databases.

Database	Search strategy
Google Scholar	Google Scholar was searched as a supplementary source using simplified keyword combinations because it does not support fully reproducible field‐tagged Boolean searching in the same manner as PubMed, Scopus, or Web of Science. Results were sorted by relevance, and the first 1000 records were screened. Search combinations included: “artificial intelligence pediatric dentistry”; “machine learning pediatric dentistry”; “deep learning pediatric dentistry”; “artificial intelligence pedodontics”; “machine learning pedodontics”; “deep learning early childhood caries”; “artificial intelligence primary teeth”; “machine learning primary teeth”; “deep learning pediatric orthodontics”; “artificial intelligence dental age children”; “machine learning panoramic radiographs children dentistry”; and “artificial intelligence orthodontic extraction children”.
PubMed	(“Artificial Intelligence”[Mesh] OR “Machine Learning”[Mesh] OR “artificial intelligence”[tiab] OR “machine learning”[tiab] OR “deep learning”[tiab] OR “neural network*”[tiab] OR “convolutional neural network*”[tiab] OR “computer vision”[tiab] OR “expert system*”[tiab]) AND (“Pediatric Dentistry”[Mesh] OR “Child”[Mesh] OR “Pediatrics”[Mesh] OR pediatric*[tiab] OR pediatric*[tiab] OR child*[tiab] OR adolescent*[tiab] OR “pediatric dentistry”[tiab] OR “pediatric dentistry”[tiab] OR pedodontic*[tiab] OR “children's dentistry”[tiab] OR “early childhood caries”[tiab] OR “primary teeth”[tiab] OR “deciduous teeth”[tiab] OR “mixed dentition”[tiab]) AND (dent*[tiab] OR oral[tiab] OR orthodontic*[tiab] OR cephalometric*[tiab] OR caries[tiab] OR plaque[tiab] OR “dental anomaly”[tiab] OR “tooth germ”[tiab] OR mesiodens[tiab] OR “dental age”[tiab])
WOS	(TS = (“artificial intelligence”) OR TS = (“machine learning”) OR TS = (“deep learning”) OR TS = (“neural network*”) OR TS = (“convolutional neural network*”) OR TS = (“computer vision”) OR TS = (“expert system*”)) AND (TS = (pediatric*) OR TS = (pediatric*) OR TS = (child*) OR TS = (adolescent*) OR TS = (“pediatric dentistry”) OR TS = (“pediatric dentistry”) OR TS = (pedodontic*) OR TS = (“children's dentistry”) OR TS = (“early childhood caries”) OR TS = (“primary teeth”) OR TS = (“deciduous teeth”) OR TS = (“mixed dentition”)) AND (TS = (dent*) OR TS = (oral) OR TS = (orthodontic*) OR TS = (cephalometric*) OR TS = (caries) OR TS = (plaque) OR TS = (“dental anomaly”) OR TS = (“tooth germ”) OR TS = (mesiodens) OR TS = (“dental age”))
Scopus	(TITLE‐ABS‐KEY(“artificial intelligence”) OR TITLE‐ABS‐KEY(“machine learning”) OR TITLE‐ABS‐KEY(“deep learning”) OR TITLE‐ABS‐KEY(“neural network*”) OR TITLE‐ABS‐KEY(“convolutional neural network*”) OR TITLE‐ABS‐KEY(“computer vision”) OR TITLE‐ABS‐KEY(“expert system*”)) AND (TITLE‐ABS‐KEY(pediatric*) OR TITLE‐ABS‐KEY(pediatric*) OR TITLE‐ABS‐KEY(child*) OR TITLE‐ABS‐KEY(adolescent*) OR TITLE‐ABS‐KEY(“pediatric dentistry”) OR TITLE‐ABS‐KEY(“pediatric dentistry”) OR TITLE‐ABS‐KEY(pedodontic*) OR TITLE‐ABS‐KEY(“children's dentistry”) OR TITLE‐ABS‐KEY(“early childhood caries”) OR TITLE‐ABS‐KEY(“primary teeth”) OR TITLE‐ABS‐KEY(“deciduous teeth”) OR TITLE‐ABS‐KEY(“mixed dentition”)) AND (TITLE‐ABS‐KEY(dent*) OR TITLE‐ABS‐KEY(oral) OR TITLE‐ABS‐KEY(orthodontic*) OR TITLE‐ABS‐KEY(cephalometric*) OR TITLE‐ABS‐KEY(caries) OR TITLE‐ABS‐KEY(plaque) OR TITLE‐ABS‐KEY(“dental anomaly”) OR TITLE‐ABS‐KEY(“tooth germ”) OR TITLE‐ABS‐KEY(mesiodens) OR TITLE‐ABS‐KEY(“dental age”))

Google Scholar was used as a supplementary source to identify potentially relevant material not retrieved through PubMed, Scopus, or Web of Science. The simplified keyword combinations listed in Table 1 were entered separately, and results were displayed using Google Scholar's default relevance ranking rather than date‐based sorting. Records were examined sequentially in the order displayed, with screening stopped after a cumulative total of 1000 records across the predefined queries had been reviewed. This pragmatic stopping rule was selected because Google Scholar displays a limited number of viewable results and does not provide the same transparent Boolean functionality, export capability, or reproducibility as conventional bibliographic databases. Titles and available snippets were assessed using the same eligibility criteria applied to the database records. The screened citations were recorded, imported into Rayyan, and deduplicated against records retrieved from the other sources. Google Scholar was used only as a supplementary search source and not as a replacement for the structured database searches.

2.2. Record Management and Deduplication

All retrieved records were imported into Rayyan for record management, duplicate detection, and screening. Deduplication was performed before title and abstract screening. Rayyan's duplicate‐detection function was used to identify potentially duplicate records based on overlapping bibliographic information, including title, author names, publication year, journal name, and DOI when available. Automatically identified duplicates were then manually checked by the reviewers to confirm that the records referred to the same article and to avoid accidental removal of distinct articles with similar titles or overlapping author groups. When duplicate records were identified across databases, the most complete citation record was retained. After deduplication, 378 duplicate records were removed, and the remaining records proceeded to title and abstract screening.

2.3. Eligibility Criteria

The eligibility objective was to identify original studies evaluating artificial intelligence methods in populations, dental records, or clinical outcomes that directly represented pediatric dental care. Studies were classified as pediatric‐only when all observations were obtained from participants aged 18 years or younger at the time of imaging, examination, or outcome assessment. Mixed‐age studies were eligible only when participants aged 18 years or younger were explicitly included and their data and model‐performance estimates were reported separately or could be unambiguously extracted. When participant ages were incompletely reported, a study was eligible only if the analyzed records represented an inherently pediatric developmental stage and the corresponding results were separately identifiable. Developmental dental tasks were defined as assessments of primary or mixed dentition, tooth eruption or tooth‐germ development, dental‐age estimation based on developing dentition within the pediatric age range, craniofacial growth or maturation, or treatment decisions explicitly dependent on pediatric developmental timing. Adult‐only studies and mixed‐age studies without separable pediatric evidence were excluded, even when the evaluated AI method had potential applicability to pediatric dentistry; task‐level or methodological relevance alone was insufficient for inclusion. Eligible reports were English‐language, full‐text original research studies applying artificial intelligence, machine learning, deep learning, or a related computational model and reporting extractable study data. Reviews, editorials, letters, case reports, conference abstracts, studies without original data, non‐English articles, inaccessible full texts, and studies not applying an AI‐based method were excluded.

For transparency, each included study was assigned a pediatric‐eligibility classification during data extraction. Studies were classified as “pediatric‐only” when all analyzed participants or records were obtained from individuals aged 18 years or younger. Studies were classified as “mixed‐age/pediatric‐relevant” only when pediatric participants or pediatric‐specific model‐performance estimates were separately identifiable, or when the analyzed records represented an explicitly pediatric developmental stage with separately reportable results. Studies based exclusively on adult populations, as well as mixed‐age studies without separately identifiable pediatric evidence, were excluded from the pediatric synthesis.

2.4. Screening and Data Extraction

After deduplication, 2158 unique records were independently screened by two reviewers (MA and MK). Following independent assessment and resolution of disagreements through discussion or consultation with the third reviewer (MAA), 1973 records were excluded and 185 reports advanced to full‐text assessment. The 185 full‐text reports were independently evaluated against the eligibility criteria. Of these, 25 reports (13.5%) required consensus discussion. All disagreements were resolved before the final decision, resulting in the exclusion of 150 full‐text reports and inclusion of 35 studies in the qualitative synthesis.

During full‐text assessment, two reviewers independently recorded the reported participant age range at data acquisition, the dentition or developmental stage evaluated, and whether pediatric data and model‐performance estimates were separately extractable. A reported mean age of 18 years or younger was not considered sufficient when the sample could include adults and the pediatric subgroup could not be identified. Eligibility disagreements were resolved through reviewer discussion and, when necessary, consultation with the third reviewer.

2.5. Data Extraction

Data extraction included author, year, country, study type, participant demographics (age and sample size), AI methodologies used, the specific dental applications evaluated, study outcomes, and conclusions. Two reviewers (MB and SSRS) independently performed data extraction, and any disagreements were resolved by discussion or consultation with a third reviewer (MAA).

Validation and translation characteristics were additionally extracted at the study level, including internal or test‐set validation, observer‐ or manual‐comparative evaluation, independent external validation, prospective clinical evaluation, multicenter evaluation, and calibration reporting. Independent external validation was defined as evaluation using data from an institution, population, or clinical setting distinct from that used for model development. Prospective clinical evaluation required assessment of a prespecified model on prospectively accrued cases, whereas multicenter evaluation required model assessment across two or more independent clinical sites. For prediction and decision‐support models, calibration was considered reported when an explicit calibration measure or assessment, such as a calibration plot, calibration slope or intercept, Brier score, or equivalent comparison of predicted and observed risk, was provided.

2.6. Risk of Bias Assessment

Risk‐of‐bias tool selection was determined according to the intended clinical purpose and evaluation framework of each study rather than its AI architecture alone. QUADAS‐3 was applied to studies evaluating an AI system as a diagnostic index test against an independent clinical, expert‐derived, or accepted reference standard. These included eligible detection, classification, segmentation, cephalometric‐assessment, dental‐anomaly, caries or plaque‐detection, and dental‐age‐estimation studies conducted within a diagnostic or reference‐standard framework. The four QUADAS‐3 domains were participants, index test, target condition, and analysis.

PROBAST + AI was applied to studies developing or evaluating models intended to predict an individual‐level outcome or clinical decision, including caries risk, dental‐care utilization, biomarker‐ or microbiome‐based risk, and orthodontic treatment or extraction decisions. It was also applied to model‐development studies that did not use a diagnostic index‐test framework with an independent reference standard. The four PROBAST + AI domains were participants and data sources, predictors, outcome, and analysis.

Two reviewers independently applied the relevant instrument using its signaling questions and review‐specific guidance concerning dataset selection, training–validation–test separation, reference‐standard construction, missing‐data handling, overfitting, calibration, and external validation. Domain‐level and overall judgments were categorized as low risk, some concerns, high risk, or insufficient information. Disagreements were resolved through discussion or consultation with a third reviewer. QUADAS‐3 and PROBAST + AI judgments were presented in separate panels and were not combined into a numerical score or pooled across instruments.

2.7. Analysis and Synthesis

Because this was a scoping review, the objective was to map the breadth, distribution, and methodological characteristics of AI applications in pediatric dentistry rather than to perform a meta‐analysis. Therefore, no pooled estimates of diagnostic accuracy, sensitivity, specificity, AUC, or predictive performance were calculated.

Extracted data were synthesized descriptively within five standardized clinical application domains (1): diagnostic imaging and dental‐anomaly detection (2); caries and plaque detection (3); orthodontic and craniofacial assessment and decision support (4); dental age estimation; and (5) risk prediction and biomarker modeling. For quantitative summaries, each study was assigned to one primary domain according to its principal clinical objective, while secondary applications were retained in the narrative and visual mapping. Imaging studies were classified according to their primary clinical task rather than being automatically assigned to the diagnostic‐imaging domain. Within each domain, we summarized pediatric relevance, input data, AI method, reported outcomes, and validation approach.

Visual synthesis was performed using two descriptive tools: a thematic map and a performance metric map. The thematic map summarized the main clinical domains and subdomains represented in the evidence base. The performance metric map displayed the availability and magnitude of reported performance indicators, such as accuracy, precision, recall, F1‐score, AUC, mean absolute error, correlation, Dice coefficient, or mIoU, according to what each article reported. These metrics were not pooled across heterogeneous tasks and were interpreted only as descriptive indicators of reported model performance.

2.8. Descriptive Performance Metric Mapping

A descriptive performance metric map was generated to show which AI applications reported quantitative performance metrics and how those metrics varied across pediatric‐relevant clinical tasks. This visualization was not treated as a meta‐analysis because the included articles differed substantially in study design, clinical task, dataset size, imaging modality, reference standard, unit of analysis, validation strategy, and reported metric type.

Metrics were extracted as reported by each article. Accuracy, precision, recall, F1‐score, AUC, Dice coefficient, mIoU, correlation coefficients, and error‐based measures such as MAE or RMSE were not combined into a single pooled estimate because they represent different model behaviors and different clinical tasks. When comparable metrics were available within the same task category, they were presented descriptively with the contributing article names and ranges. Blank cells indicated that the metric was not reported or was not extractable for the pediatric‐relevant component.

2.9. Thematic Mapping of AI Applications

Thematic analysis was first performed to identify recurrent concepts and clinical foci within the included literature. The thematic coding was performed independently by two reviewers. Each reviewer examined the stated study objectives, clinical application, input data, and reported outcomes and independently assigned each study to a primary clinical domain and, where applicable, relevant subdomains. Disagreements in domain or subdomain assignment were resolved through discussion between the two reviewers; when consensus could not initially be reached, a third reviewer adjudicated the classification. The final domain and subdomain assignments were established by consensus before construction of the thematic map.

Inductive coding of study aims and outcomes yielded the same five standardized clinical application domains used for descriptive synthesis. The thematic diagram retained relevant subdomains and cross‐domain links when studies had secondary clinical or methodological applications. To visually synthesize the scope of AI applications in pediatric dentistry, a thematic diagram was developed. Key themes were derived from a narrative synthesis of the included studies and grouped based on shared clinical objectives or methodological similarities. The primary categories included diagnostic imaging, caries detection, orthodontic assessment, age estimation, and risk prediction. Each main category was then expanded into subdomains reflecting specific use cases and AI techniques (e.g., convolutional neural networks for mesiodens detection or salivary biomarkers for caries risk). The diagram was constructed using a directed graph layout implemented with NetworkX and Matplotlib in Python.

2.10. Statistical and Software Reporting

No meta‐analysis or inferential hypothesis testing was performed because this systematic scoping review was designed to map the scope, characteristics, and reported metrics of the evidence rather than to estimate pooled diagnostic or predictive performance. Descriptive statistics were used to summarize study counts, proportions, application domains, validation approaches, and reported model metrics. Proportions are reported with numerators and denominators where applicable. Reported performance values, including accuracy, precision, recall, F1‐score, AUC, mIoU, Dice coefficient, MAE, RMSE, correlation coefficients, and R ², were extracted as presented in the original articles and were not statistically pooled. Visualizations were generated using Python with Matplotlib and NetworkX. All analyses were descriptive and exploratory.

3. Results

3.1. Study Selection

The searches identified 2536 records across PubMed, Scopus, Web of Science, and Google Scholar. After removal of 378 duplicate records, 2158 unique records underwent title and abstract screening, of which 1973 were excluded. The remaining 185 reports were assessed in full text. Of the 185 reports assessed at full text, 150 were excluded: 62 because the population was non‐pediatric or pediatric‐relevant data were not separately extractable, 28 because no eligible AI‐based analysis or algorithm was evaluated, 24 because they were reviews, conference publications, or other reports without eligible original data, 21 because methodological or performance information required for eligibility was insufficient or non‐extractable, and 15 because the full text was non‐English or inaccessible. The remaining 35 studies were included in the final qualitative synthesis (Figure 1).

3.2. Study Characteristics

Of the 35 included studies, 19 (54.3%) were classified as pediatric‐only, with all analyzed participants or records derived from individuals aged 18 years or younger. The remaining 16 studies (45.7%) were classified as broader or mixed‐age pediatric‐relevant studies, in which the eligible evidence consisted of separately identifiable pediatric participants, pediatric‐specific outcomes, or developmental dental outcomes relevant to the pediatric age range. For these studies, only the pediatric or pediatric‐relevant component was incorporated into the synthesis. Studies without separately identifiable pediatric data or pediatric‐relevant developmental outcomes were not considered eligible on the basis of general methodological or clinical applicability alone.

The included studies were categorized into the same five primary domains: diagnostic imaging and dental‐anomaly detection, caries and plaque detection, orthodontic and craniofacial assessment and decision support, dental age estimation, and risk prediction and biomarker modeling (Table 2).

TABLE 2.

Concise characteristics of included studies evaluating artificial intelligence applications in pediatric dentistry.

Study	Year	Country	Population/sample	Age	Data type	AI model	Clinical task	Validation type	Key reported metric	Main limitation
Çetin et al. [6]	2026	Turkey	176 panoramic radiographs; 88 extraction‐indicated and 88 non‐indicated cases	8–10 years	Panoramic radiographs	Gabor–HOG–SVM	Permanent first molar extraction indication	Test‐set evaluation	Accuracy 77.78%; AUC 0.77; extraction‐indicated recall 0.81	Single‐center retrospective dataset; Demirjian‐based timing label; no external validation
Kaya et al. [12]	2022	Turkey	4518 pediatric panoramic radiographs	5–13 years	Panoramic radiographs	YOLOv4	Permanent tooth germ detection	Internal validation	Average precision 94.16%; F1‐score 0.90	Single‐country imaging dataset; external validation not reported
Kim et al. [13]	2022	South Korea	988 pediatric panoramic radiographs	6–12 years	Panoramic radiographs	CNN with automated ROI segmentation	Mesiodens diagnosis	Internal validation	Accuracy, precision, recall, F1‐score, and AUC 0.971	Retrospective imaging dataset; limited external generalizability
Kuwada et al. [14]	2021	Japan	383 panoramic radiographs	Not clearly reported	Panoramic radiographs	Deep learning object‐detection model	Cleft alveolus detection/classification	Internal observer comparison	False positives reduced from 12/30 to 1/30 after adding normal data	Population details incompletely reported; external validation not reported
Lin et al. [15]	2021	Taiwan	895 dental panoramic radiographs	Adult source dataset	Panoramic radiographs	CNN‐based two‐phase classifier	Tooth numbering and dental condition recognition	Internal validation	Tooth numbering accuracy 95.62%; condition accuracy 98.33%	Adult source dataset with subgroups of pediatric relevance
Ragodos et al. [16]	2022	USA	4084 intraoral photographs	Not clearly reported	Intraoral photographs	ResNet‐18 with transfer learning	Dental anomaly classification	Internal validation	Mean F1‐score 0.437–0.561; median AUC 0.683–0.872	Age reporting incomplete; moderate performance; external validation not reported
Xie et al. [17]	2010	USA	200 orthodontic patients	11–15 years	Clinical orthodontic variables	Artificial neural network	Orthodontic extraction decision support	Training/test evaluation	80% accuracy in untrained cases	Small test set; older model; limited contemporary validation
You et al. [18]	2020	USA	886 primary‐tooth images	6–12 years	Intraoral photographs	Deep learning segmentation model	Plaque detection	Comparison with pediatric dentist assessment	mIoU 0.736 ± 0.174	Metric not directly comparable with diagnostic accuracy; external validation not reported
Zaborowicz et al. [19]	2022	Poland	619 participants	4–12 years	Dental imaging‐derived tooth and bone indices	Deep neural network	Chronological age estimation	Internal validation	MAE 2.34–4.61 months; RMSE 5.58–7.49 months; R 2 0.92–0.96	Proprietary indices; external validation not reported
Zheng et al. [20]	2021	China	180 participants	Not clearly reported	CBCT/pulp chamber segmentation	Deep learning plus level‐set method	Age estimation from pulp chamber volume	Training/validation/test evaluation	Dice 87.8%; age correlation r = 0.74	Age range unclear in extracted table; limited population generalizability
Zhou et al. [21]	2022	China	304 pediatric panoramic radiographs	Children	Panoramic radiographs	Context‐aware CNN	Children's caries diagnosis	Internal validation	Accuracy 82.72%; precision 85.38%; recall 87.70%; AUC 0.9005	Limited sample size; single‐task validation; external testing not reported
Bunyarit et al. [22]	2022	Malaysia	1015 children/adolescents	5–18 years	Dental maturity scores	ANN‐MLP	Dental age estimation	Internal validation	Chronological–dental age difference: 12 days in boys and 25 days in girls	Population‐specific formula; generalizability uncertain
Panesar et al. [23]	2023	USA	30 cephalometric cases	≥ 18 years	Cephalometric landmarks	Deep learning AI with/without human augmentation	Cephalometric analysis	Comparative observer evaluation	Examiner precision improved by 10.47%; accuracy improved by 27.27%	Mixed age
Nauwelaers et al. [24]	2021	Belgium	1324 3D maxillary dental casts	10–80 years	3D dental casts	Singular autoencoder	3D palatal shape modeling	Internal comparative evaluation	SAE comparable with PCA and autoencoders	Mixed‐age dataset
Park et al. [25]	2021	South Korea	4195 children	1–5 years	Clinical and behavioral data	XGBoost, random forest, LightGBM, logistic regression	Early childhood caries prediction	Internal validation	AUROC 0.774–0.785	External validation and calibration reporting limited
Peluso et al. [26]	2023	Italy	18 cephalograms from CBCT images	9–19 years	CBCT‐derived cephalograms	3D Enlow neutral‐track analysis	Cephalometric/craniofacial assessment	Comparative 2D/3D evaluation	3D analysis comparable with 2D Enlow method	Very small sample; limited external validation
Peng et al. [27]	2021	USA	24,223 Medicaid‐insured children	24–60 months	Dental‐care utilization data	Machine‐learning pattern discovery	Early childhood dental‐care utilization patterns	Retrospective cohort modeling	Regular preventive care 67.7%; zero utilization 23.0%	Utilization‐focused outcome; not direct diagnostic model
Popova et al. [28]	2023	Germany	430 cephalometric images	< 6 and > 13 years	Cephalometric radiographs	Customized CNN	Automated cephalometric landmark recognition	Internal validation	Developmental stage affected selected landmark accuracy	Heterogeneous age groups; device and subgroup generalizability uncertain
Prasad et al. [29]	2023	India/Australia	700 orthodontic cases	10–30 years	Cephalometric and orthodontic diagnostic data	Decision tree, random forest, XGBoost	Orthodontic diagnosis and treatment planning	Internal validation/expert comparison	Overall accuracy 84%; top algorithms 87%–93%	Mixed‐age sample
Qu et al. [30]	2022	China	745 children	12–60 months	Behavioral and non‐biological risk factors	Machine‐learning caries risk prediction model	New‐onset caries risk prediction	Internal validation	66/745 children developed new caries; model showed good internal performance	External validation and calibration reporting limited
Ristau et al. [31]	2022	USA	60 cephalometric cases	Not clearly reported	Cephalometric radiographs	CNN‐based automated tracing software	Cephalometric landmark identification	Human–AI comparison	No major difference between orthodontists and automated tracing for most landmarks	Age/pediatric relevance incompletely reported; small sample
Rousseau and Retrouvey[32]	2022	Canada	45 facial images/cases	15–25 years	Facial images/vertical dimension measures	Automated facial analysis with machine learning	Orthodontic vertical‐dimension assessment	Manual vs. automated comparison	Automated method showed agreement with manual method	Mixed adolescent/adult sample; narrow orthodontic task
Seo et al. [33]	2021	South Korea	600 lateral cephalometric radiographs	6–19 years	Lateral cephalometric radiographs	CNN models with Grad‐CAM	Cervical vertebral maturation classification	Internal validation	All models > 90% accuracy; Inception‐ResNet‐v2 best‐performing	External validation not reported; single imaging task
Suhail et al. [34]	2020	USA	287 orthodontic cases	< 18 years	Orthodontic diagnostic features	Random forest and ensemble classifiers	Orthodontic extraction diagnosis/treatment prediction	Internal model evaluation	Ensemble performance close to inter‐orthodontist agreement	Age details unclear; external validation not reported
Ueda et al. [35]	2023	USA	220 cephalometric cases	≥ 18 years	Cephalometric measurements	Random forest, logistic regression, SVC	Maxillofacial morphology classification	Internal validation	Anteroposterior classification accuracy 0.986 ± 0.011	Adult source sample with pediatric subgroup; not directly pediatric
Wu et al. [36]	2022	USA/Thailand cohort	177 mother–child dyads	3 years	Salivary microbiome and behavioral data	Statistical machine‐learning models	Caries‐associated factor identification	Cross‐sectional/internal modeling	Mode of delivery and snack consumption outranked microbiome features	Observational design; causal inference and external validation limited
Guo et al. [37]	2021	Multicountry	10,257 OPG images	5–24 years	Orthopantomogram images	Deep CNN	Age classification	Internal validation	Deep CNN outperformed manual methods	Mixed‐age dataset
Heimisdottir et al. [38]	2021	USA	300 participants	3–5 years	Supragingival biofilm metabolomics	TPOT/automated machine learning	ECC‐associated metabolomic markers	Cross‐sectional modeling	16 metabolites associated with ECC after correction	Extracted age appears inconsistent with ECC focus
de Jesus et al. [39]	2020	Canada	80 children	45.9 ± 12.8 months	Plaque microbiota	TPOT/machine‐learning analysis	Sex‐based plaque microbiota in severe ECC	Comparative microbiome analysis	Sex‐based fungal composition differences identified	Small sample; biomarker findings require validation
Karhade et al. [40]	2021	USA	6404 children	36–71 months	Clinical, demographic, and parent‐reported data	Automated machine‐learning classifier	Early childhood caries classification	Internal validation	Parsimonious model AUC 0.74; sensitivity 0.67; PPV 0.64	External validation and calibration limited
Jung and Kim[41]	2016	South Korea	156 orthodontic patients	Mean age 24 years	Orthodontic diagnostic variables	Neural network models	Orthodontic extraction diagnosis	Internal validation	93% extraction/non‐extraction accuracy; 84% extraction‐pattern accuracy	Adult/young‐adult mean age
Koopaie et al. [42]	2021	Iran	40 children	48–72 months	Salivary biomarker and clinical data	Logistic regression/machine‐learning‐supported analysis	ECC biomarker‐based classification	Case‐control analysis	Lower cystatin S in ECC group; p = 0.03	Small sample; case‐control design; external validation not reported
Lee et al. [43]	2022	South Korea	471 participants	11–69 years	Panoramic radiomorphometric parameters	Machine‐learning algorithms	Age‐group prediction	Internal validation	AUC 0.73–0.90 across age groups	Broad mixed‐age dataset; pediatric subgroup
Kök et al. [44]	2019	Turkey	300 orthodontic patients	8–17 years	Cervical vertebral maturation data	k‐NN, Naive Bayes, decision tree, ANN, SVM, RF, logistic regression	Growth and development classification	Comparative internal validation	ANN showed stable/high accuracy across stages	External validation not reported
Kök et al. [45]	2021	Turkey	419 orthodontic patients	8–17 years	Cervical vertebral and hand‐wrist maturation data	Artificial neural network	Growth period and sex determination	Internal validation	ANN‐7 accuracy 0.9427; gender accuracy 0.8950	Retrospective dataset; external validation not reported

Note: This table summarizes the core features of the included pediatric or pediatric‐relevant dental AI studies, including author, year, country, population, age range, data type, AI model, clinical task, validation approach, key reported performance metric, and main limitation. Pediatric eligibility was determined using the prespecified age‐ and development‐based criteria described in the Methods. Mixed‐age studies were retained only when pediatric‐specific data, performance estimates, or explicitly pediatric developmental outcomes were separately identifiable; adult‐only evidence was not included in the synthesis.

3.3. Validation and Translation Characteristics

Validation was predominantly confined to development‐stage evaluation. Of the 35 included studies, 26 (74.3%) used internal or held‐out test‐set validation, five (14.3%) used observer‐ or manual‐comparative evaluation, and four (11.4%) did not clearly define the validation approach. Independent external validation was identified in 0 of 35 studies (0%), and no study evaluated a prespecified AI system prospectively within a routine pediatric dental workflow (0/35, 0%). Likewise, independent multicenter external evaluation was not identified (0/35, 0%). Among the 16 Prediction or model‐development studies assessed using PROBAST + AI, formal calibration assessment was not identified in the extracted reporting (0/16, 0%). These findings indicate that the available evidence primarily characterizes model development and internal performance rather than transportability or prospective clinical performance.

3.4. Risk‐of‐Bias Findings

Nineteen diagnostic index‐test studies were assessed using QUADAS‐3. In the participants domain, 13 studies (68.4%) were judged as low risk, five (26.3%) as having some concerns, and one (5.3%) as high risk. For the index‐test domain, 13 studies (68.4%) were judged as low risk, five (26.3%) as having some concerns, and one (5.3%) as having insufficient information. In the target‐condition domain, 16 studies (84.2%) were judged as low risk, two (10.5%) as having some concerns, and one (5.3%) as high risk. The analysis domain showed the greatest concentration of non‐low‐risk judgments: 11 studies (57.9%) were judged as low risk, five (26.3%) as having some concerns, and three (15.8%) as high risk. Overall, 16 studies (84.2%) were judged as low risk and three (15.8%) as having some concerns.

Sixteen prediction or model‐development studies were assessed using PROBAST + AI. In the participants and data‐sources domain, 11 studies (68.8%) were judged as low risk, four (25.0%) as having some concerns, and one (6.2%) as having insufficient information. In the predictors domain, 12 studies (75.0%) were judged as low risk, one (6.2%) as having some concerns, one (6.2%) as high risk, and two (12.5%) as having insufficient information. In the outcome domain, 14 studies (87.5%) were judged as low risk, one (6.2%) as having some concerns, and one (6.2%) as having insufficient information. In the analysis domain, 12 studies (75.0%) were judged as low risk, one (6.2%) as having some concerns, two (12.5%) as high risk, and one (6.2%) as having insufficient information. Overall, 12 studies (75.0%) were judged as low risk, three (18.8%) as having some concerns, and one (6.2%) as having insufficient information. Complete study‐level judgments are presented in Figure 2.

FIGURE 2.

Tool‐specific risk‐of‐bias assessment of the included studies. Panel A presents QUADAS‐3 judgments for diagnostic index‐test studies across the participants, index test, target condition, and analysis domains. Panel B presents PROBAST + AI judgments for prediction and model‐development studies across the participants and data sources, predictors, outcome, and analysis domains. Overall judgments are displayed separately. Green circles indicate low risk, yellow triangles indicate some concerns, red squares indicate high risk, and blue diamonds indicate insufficient information. No numerical scoring or pooling across the two tools was performed.

3.5. Visual Evidence Mapping of AI Applications and Reported Metrics

To summarize the mapped evidence, two descriptive visual tools were generated: a thematic map and a descriptive performance metric map. These figures were designed to show the distribution of AI applications and the types of performance metrics reported across pediatric or pediatric‐relevant dental tasks. They were not intended to provide pooled estimates of diagnostic accuracy or clinical effectiveness.

Figure 3 summarizes the five mutually exclusive primary application domains identified through thematic coding: diagnostic imaging and dental‐anomaly detection; caries and plaque detection; orthodontic and craniofacial assessment and decision support; dental age estimation; and risk prediction and biomarker modeling. For quantitative domain counts, each included study was assigned to a single primary domain according to its principal clinical objective. Secondary applications were retained in the narrative synthesis and thematic subcategories and could span more than one clinical domain. Figure 4 presents article‐level reported performance metrics for all eligible studies with extractable quantitative performance data for the pediatric or pediatric‐relevant component. Because the included articles differed in clinical task, input data, validation strategy, reference standard, and metric type, reported values were mapped descriptively rather than pooled.

FIGURE 3.

Hierarchical thematic map of artificial intelligence applications in pediatric dentistry. The 35 included studies were assigned to one mutually exclusive primary clinical domain according to their principal clinical objective. Numbers and percentages show the distribution of studies across the five primary domains, while the nested boxes summarize the principal clinical subcategories represented within each domain. Secondary applications were retained in the narrative synthesis but were not included in the primary‐domain counts.

FIGURE 4.

Descriptive performance metric map of AI applications in pediatric dentistry. The figure presents article‐level quantitative performance measures from all included studies meeting the prespecified visualization criterion. Values are displayed as originally reported and were not pooled or averaged. Metrics expressed on directly interpretable classification‐performance scales, including accuracy, precision or average precision, recall or sensitivity, F1‐score, and AUC, are displayed alongside separately presented non‐equivalent metrics such as mIoU, Dice coefficient, correlation coefficients, MAE, RMSE, and R ². Studies without an extractable quantitative performance measure for the eligible pediatric or pediatric‐relevant component were not displayed. The visualization is intended as a descriptive evidence map rather than a ranking of algorithms or a comparative assessment of clinical effectiveness.

3.6. Descriptive Performance Metric Map

Figure 4 shows the distribution of reported performance metrics across all included studies meeting the prespecified visualization criterion. Because different studies reported different metric types, values are displayed descriptively at the article level and are not intended for direct cross‐study comparison. Kim et al. (2022) reported accuracy, precision, recall, F1‐score, and AUC of 0.971 for automated mesiodens diagnosis. Kaya et al. (2022) reported an average precision of 94.16% and an F1‐score of 0.90 for permanent tooth germ detection. Zhou et al. (2022) reported accuracy of 82.72%, precision of 85.38%, recall of 87.70%, F1‐score of 86.52%, and AUC of 0.9005 for caries detection on pediatric panoramic radiographs. You et al. (2020) reported mIoU values for plaque detection rather than conventional diagnostic accuracy metrics.

Orthodontic and extraction‐related decision‐support metrics were reported by Xie et al. (2010), Jung and Kim (2016), and Çetin et al. (2026b). Xie et al. (2010) reported 80% accuracy for orthodontic extraction decision‐making in untrained cases. Jung and Kim (2016) reported 93% accuracy for extraction vs. non‐extraction diagnosis and 84% accuracy for detailed extraction‐pattern diagnosis. Çetin et al. (2026b) reported 77.78% accuracy and an AUC of 0.77 for classifying timing‐based permanent first molar extraction indication in children using a Gabor–HOG–SVM model.

Dental age‐estimation articles reported heterogeneous outcomes. Zheng et al. (2021) reported Dice similarity and correlation‐based performance, Bunyarit et al. (2022) reported chronological–dental age differences, and Zaborowicz et al. (2022) reported MAE, RMSE, and R ². Because these metrics are not directly interchangeable, they were retained in their original form and interpreted as reported article‐level outcomes. Reported performance metrics varied substantially across clinical tasks and metric types and were therefore retained as article‐level descriptive outcomes without interpretation of comparative effectiveness or clinical utility.

Studies were eligible for inclusion in the performance metric map when at least one quantitative model‐performance measure was explicitly reported and extractable for the pediatric or pediatric‐relevant component. Studies were excluded from the visualization only when no quantitative performance measure was reported or when the relevant performance result could not be separated from an ineligible broader population. Selection for the map was not based on the magnitude of reported performance, model architecture, publication year, clinical domain, or perceived study importance.

3.7. Thematic Analysis of AI Applications

3.7.1. Diagnostic Imaging

Image‐based AI applications were identified across several clinical domains, including dental‐anomaly detection, caries and plaque detection, cephalometric assessment, craniofacial growth evaluation, and dental‐age estimation. Imaging‐based articles included Kaya et al. (2022) for permanent tooth germ detection, Kim et al. (2022) for mesiodens diagnosis, Kuwada et al. (2021) for cleft alveolus detection/classification, Ragodos et al. (2022) for dental anomaly classification from intraoral photographs, Zhou et al. (2022) for caries detection, You et al. (2020) for plaque detection, Seo et al. (2021) for cervical vertebral maturation classification, Popova et al. (2023) for cephalometric landmark recognition, and Kök et al. (2019, 2021) for growth and development assessment.

These articles primarily evaluated model performance on retrospective, curated, or internally validated datasets. Therefore, the findings indicate that image‐based AI tasks are the most developed area of pediatric dental AI research, but they do not by themselves establish prospective clinical effectiveness [13] (Figure 3).

3.7.2. Caries Detection and Risk Prediction

Caries‐related AI applications included both image‐based detection and risk‐prediction approaches. Zhou et al. (2022) evaluated a context‐aware CNN for children's caries diagnosis on panoramic radiographs, whereas You et al. (2020) evaluated AI‐based plaque detection on primary teeth using mIoU. Prediction‐oriented articles included Park et al. (2021), Qu et al. (2022), Karhade et al. (2021), Koopaie et al. (2021), Wu et al. (2022), Heimisdottir et al. (2021), and de Jesus et al. (2020), which used clinical, behavioral, salivary, microbiome, metabolomic, or plaque‐related predictors.

Across the full evidence base, internal or test‐set validation was used in 26 of 35 studies (74.3%), observer‐ or manual‐comparative evaluation in five (14.3%), and the validation approach was not clearly defined in four (11.4%). Independent external validation was reported in none of the 35 included studies (0%) [18].

3.7.3. Orthodontic Assessments

Orthodontic and extraction‐related decision support was represented by Xie et al. (2010), Jung et al. (2016), and Çetin et al. (2026b). Xie et al. (2010) used artificial neural networks to support orthodontic extraction decisions in patients aged 11–15 years and reported 80% accuracy in untrained cases. Jung et al. (2016) developed neural network models for orthodontic extraction diagnosis and reported 93% accuracy for extraction vs. non‐extraction decisions and 84% accuracy for detailed extraction‐pattern diagnosis.

Çetin et al. (2026b) evaluated a Gabor–HOG–SVM model for permanent first molar extraction indication in children aged 8–10 years using panoramic radiographs. The model achieved 77.78% accuracy and an AUC of 0.77. The reference classification was based on Demirjian E–F developmental stages of the permanent second molar [17, 23, 31].

3.7.4. Age Estimation

Several studies evaluated AI‐based dental‐age estimation using radiographic or imaging‐derived features. Zheng et al. reported a correlation of r = 0.74 between predicted and reference age estimates based on pulp‐chamber volume. Bunyarit et al. reported differences between estimated dental age and chronological age of approximately 12–25 days. Other studies used error‐based measures, including MAE, RMSE, and R ², reflecting substantial heterogeneity in outcome reporting across age‐estimation models [20, 22].

3.7.5. Predictive Modeling and Risk Assessment

Machine‐learning approaches were also evaluated for pediatric caries‐risk prediction. Park et al. developed early‐childhood caries prediction models using XGBoost and random‐forest algorithms applied to clinical and behavioral variables. Qu et al. developed a caries‐risk prediction model using behavioral and non‐biological risk factors in children aged 12–60 months. These studies differed in predictor composition, modeling strategy, and validation approach [25, 30].

3.7.6. Diagnostic and Predictive Biomarkers

Biomarker‐oriented studies evaluated biological features associated with pediatric caries outcomes. Koopaie et al. identified salivary cystatin S as a variable differentiating children with and without early childhood caries, whereas Wu et al. evaluated salivary microbiome characteristics together with behavioral and clinical variables in relation to caries outcomes.

Across the 35 included studies, AI applications encompassed image‐based diagnosis, caries‐related prediction, orthodontic and craniofacial assessment, dental‐age estimation, and biomarker‐informed modeling. Validation was internal or test‐set based in 26 studies (74.3%), observer‐ or manual‐comparative in five (14.3%), and not clearly defined in four (11.4%); independent external validation was reported in none of the included studies [36, 42].

4. Discussion

4.1. Principal Findings and Evidence Concentration

This review identified a pediatric dental AI evidence base that is broad in clinical scope but unevenly distributed across application domains. Orthodontic and craniofacial assessment and decision support accounted for 15 of 35 studies (42.9%), followed by risk prediction and biomarker modeling in eight (22.9%), while diagnostic imaging and dental‐anomaly detection and dental‐age estimation each accounted for five studies (14.3%) and caries or plaque detection for two (5.7%). More importantly, expansion in the range of clinical applications was not accompanied by comparable maturation of validation practices: 26 of 35 studies (74.3%) relied on internal or test‐set evaluation, and none performed independent external validation. The evidence base therefore reflects substantial methodological exploration but considerably less evidence regarding transportability, prospective performance, or routine clinical implementation.

Earlier reviews primarily cataloged pediatric dental AI applications or synthesized diagnostic‐performance evidence across selected clinical tasks [4, 8, 9, 10, 11]. The present review extends this literature by applying an explicit pediatric‐eligibility framework, distinguishing pediatric‐only from pediatric‐relevant mixed‐age evidence, assigning studies to mutually exclusive primary clinical domains, preserving heterogeneous performance measures at the article level, and examining validation characteristics alongside design‐appropriate risk‐of‐bias assessments. This approach shifts the emphasis from whether individual algorithms report favorable performance to where the evidence is concentrated, how rigorously those models have been evaluated, and which methodological gaps currently limit pediatric‐specific clinical translation.

Decision‐support applications extended beyond image recognition to orthodontic extraction and treatment‐planning tasks. However, these models generally addressed narrowly defined diagnostic or decision targets and were predominantly evaluated using internally assessed datasets. In some studies, the reference outcome represented a specific radiographic or expert‐defined decision rule rather than the full set of clinical factors used in routine treatment planning. Accordingly, these systems should be interpreted as task‐specific decision‐support models rather than evidence of autonomous clinical decision‐making.

4.2. Methodological Fragility of the Current Evidence

Methodological maturity was constrained principally by the validation structure of the available evidence. Of the 35 included studies, 26 (74.3%) relied on internal or test‐set evaluation, five (14.3%) used observer‐ or manual‐comparative evaluation, and four (11.4%) did not clearly specify the validation approach. Independent external validation was absent across the entire evidence base (0/35, 0%), limiting direct assessment of model transportability across institutions, populations, and acquisition settings.

A central issue is that performance metrics were not directly comparable across tasks. Kim et al. (2022) reported conventional classification metrics for mesiodens diagnosis, Kaya et al. (2022) reported average precision and F1‐score for tooth germ detection, You et al. (2020) reported mIoU for plaque detection, and Zaborowicz et al. (2022) reported error‐based age‐estimation metrics such as MAE, RMSE, and R ². These differences justify the scoping‐review approach and support the decision not to pool performance estimates. Reported accuracy or AUC values should therefore be interpreted as article‐level indicators of feasibility rather than as evidence of comparative clinical effectiveness [12, 13, 18, 19].

Risk‐of‐bias assessment further supports this cautious interpretation. Image‐based diagnostic and classification studies were mainly vulnerable to concerns in patient selection, index‐test conduct, reference‐standard definition, and unclear separation of training, validation, and test data. Prediction‐modeling studies were particularly limited by small effective sample sizes, incomplete calibration reporting, internal‐only validation, and potential overfitting. These limitations are not minor reporting issues; they directly affect whether model performance can be expected to persist across different clinics, imaging devices, operators, populations, and disease distributions.

4.3. Clinical Translation Gap

The current evidence suggests potential clinical utility, but it does not yet establish real‐world clinical readiness. Most included articles evaluated models in controlled datasets rather than live pediatric dental workflows. For example, Kim et al. (2022), Kaya et al. (2022), Zhou et al. (2022), You et al. (2020), and Çetin et al. (2026b) provide useful evidence for specific diagnostic or decision‐support tasks, but they do not yet demonstrate how model outputs affect chairside decisions, referral behavior, treatment timing, diagnostic confidence, patient outcomes, or clinician workload in routine practice [6, 12, 13, 18, 21].

The included studies did not directly evaluate the downstream patient‐level consequences of false‐positive or false‐negative AI outputs. Nevertheless, such errors may have different clinical implications depending on the task and should be considered when evaluating model safety. For example, in extraction‐related decision support, a false‐positive output could theoretically contribute to unnecessary intervention if acted upon without appropriate clinical review, whereas a false‐negative output could theoretically delay recognition of a treatment indication during a time‐sensitive developmental period. These examples represent plausible clinical considerations rather than outcomes demonstrated by the studies included in this review. Future investigations should therefore report class‐specific error rates, confusion matrices, calibration, subgroup performance, and, where appropriate, decision‐curve or workflow‐based analyses to characterize the clinical consequences of model errors more directly.

4.4. Pediatric‐Specific Concerns

Pediatric dental AI cannot simply be extrapolated from adult dental AI. Children present unique sources of variability, including growth, mixed dentition, eruption stage, cooperation during imaging, smaller anatomical structures, changing occlusion, and age‐dependent disease patterns. These factors influence both image acquisition and model interpretation. A model trained on permanent dentition, adult cephalometric measurements, or broad mixed‐age radiographs may not generalize to early mixed dentition unless pediatric subgroup performance is explicitly reported.

Radiation exposure also has a distinct pediatric dimension. AI tools using radiographic data should be evaluated in relation to clinically justified imaging, not as a reason to increase radiographic acquisition. The value of AI in pediatric imaging should therefore be framed around extracting more reliable information from already indicated images, reducing observer variability, and supporting timely decision‐making, rather than encouraging additional imaging.

Fairness and generalizability are also critical. Age, sex, dentition stage, socioeconomic context, ethnicity, imaging device, and clinical setting may all influence model performance. Yet demographic and device‐level reporting remains inconsistent across the included literature. Without transparent subgroup analysis, AI systems may appear accurate overall while performing less reliably in younger children, mixed‐dentition cases, low‐quality images, or underrepresented populations. This is especially important for tools proposed for underserved settings, where data quality, device availability, and clinical workflows may differ from the development environment.

4.5. Standards for Future Research

The next phase of pediatric dental AI research should move from isolated proof‐of‐concept models toward clinically validated, transparent, and auditable decision‐support systems. Future studies should prioritize multicenter external validation, clear training‐validation‐test separation, standardized annotation protocols, calibration assessment, subgroup performance analysis, and complete reporting of missing data, class imbalance, and model failure cases. Reporting should also include dataset composition, imaging device details, preprocessing steps, code or model availability where possible, and enough methodological information to allow independent replication.

For prediction and decision‐support models, authors should adopt reporting and evaluation standards aligned with AI‐specific guidance such as CLAIM, TRIPOD + AI, DECIDE‐AI, CONSORT‐AI/SPIRIT‐AI where applicable, and emerging trustworthy‐AI frameworks. Model cards or structured reporting summaries may also improve transparency by documenting intended use, input requirements, target population, contraindications, known limitations, and expected failure modes. In pediatric dentistry specifically, future work should report whether mixed‐age datasets include separable pediatric subgroups and whether performance differs by age group, dentition stage, imaging device, or clinical site.

Overall, the current evidence supports cautious optimism. AI has credible potential to assist pediatric dentists in image interpretation, caries‐risk assessment, growth evaluation, and treatment‐timing decisions. However, the field remains methodologically uneven and clinically immature. The central challenge is no longer whether AI can achieve promising performance in selected pediatric dental tasks, but whether these systems can be externally validated, transparently reported, fairly evaluated, and safely integrated into pediatric dental workflows without weakening clinician accountability.

5. Limitations

This systematic scoping review has several limitations. First, because the aim was evidence mapping rather than effect estimation, no meta‐analysis was performed. The included articles differed substantially in clinical task, data type, model architecture, reference standard, validation design, and reported performance metric. Therefore, reported values such as accuracy, AUC, F1‐score, mIoU, Dice coefficient, MAE, RMSE, and R ² were summarized descriptively and should not be interpreted as pooled estimates of diagnostic or predictive performance.

Second, the pediatric focus of the review was limited by inconsistent age reporting across the literature. Pediatric‐only cohorts were available in articles such as Kaya et al. (2022), Kim et al. (2022), Zhou et al. (2022), You et al. (2020), Park et al. (2021), Qu et al. (2022), Karhade et al. (2021), Koopaie et al. (2021), de Jesus et al. (2020), and Çetin et al. (2026b). However, several dental or orthodontic AI articles used broader or mixed‐age datasets, including Lin et al. (2021), Panesar et al. (2023), Nauwelaers et al. (2021), Prasad et al. (2023), Guo et al. (2021), Jung and Kim (2016), and Lee et al. (2022). For these articles, only pediatric‐relevant outcomes, developmental dental tasks, or clearly extractable pediatric components were considered. Nevertheless, incomplete age‐stratified reporting limited the granularity of synthesis and may reduce certainty regarding pediatric‐specific generalizability.

Third, the methodological quality of the evidence was uneven. Many image‐based articles, including Kaya et al. (2022), Kim et al. (2022), Zhou et al. (2022), You et al. (2020), Seo et al. (2021), and Çetin et al. (2026b), were based on retrospective or curated datasets. Prediction‐oriented articles such as Park et al. (2021), Qu et al. (2022), Karhade et al. (2021), Koopaie et al. (2021), Wu et al. (2022), Heimisdottir et al. (2021), and de Jesus et al. (2020) also varied in predictor selection, calibration reporting, missing‐data handling, and validation strategy. External validation, prospective testing, subgroup analysis, and explicit reporting of model failure modes were uncommon, limiting conclusions about clinical readiness.

Fourth, the review was constrained by the quality and completeness of reporting in the original articles. Details on imaging devices, acquisition protocols, annotation procedures, class balance, preprocessing pipelines, code availability, and data sharing were inconsistently reported. These gaps limited the ability to compare studies directly or assess reproducibility. In addition, some studies reported only selected performance metrics, making it difficult to evaluate calibration, clinical utility, or error consequences across comparable pediatric subgroups.

Finally, In addition, because Google Scholar results are dynamically relevance‐ranked and screening was restricted to the first 1000 displayed records, eligible studies appearing below this threshold may have been missed. Google Scholar was therefore used only as a supplementary source alongside three structured bibliographic databases. The review was also limited to English‐language full‐text articles, which may have excluded relevant studies published in other languages. Despite these limitations, the review provides a structured map of the current pediatric dental AI evidence base and identifies the methodological, reporting, and validation priorities that should guide future research.

6. Conclusion

This systematic scoping review maps the current evidence on AI applications in pediatric dentistry and highlights the growing concentration of work in diagnostic imaging, caries and plaque detection, orthodontic decision support, dental age estimation, and risk prediction. The visual evidence map shows that articles such as Kim et al. (2022), Kaya et al. (2022), Zhou et al. (2022), and You et al. (2020) reported promising task‐specific performance, particularly for image‐based applications. However, because the evidence base is heterogeneous in design, validation strategy, reference standards, and reported metrics, these findings should be interpreted as mapped evidence rather than pooled estimates of clinical effectiveness. Future pediatric dental AI research should prioritize external validation, transparent reporting, demographic and device‐level generalizability, and clinically meaningful evaluation before routine implementation.

Author Contributions

Aref Zarei: writing – review and editing, validation, writing – original draft. Amirhossein Ehsani: writing – original draft, writing – review and editing, validation. Amir Ebrahimzedeh Pasha: writing – review and editing, writing – original draft, validation. Mansour Gorjizad: validation, writing – original draft, writing – review and editing, investigation. Majid Aryannejad: writing – review and editing, writing – original draft. Seyedeh sara Rezazadeh Shojaee: writing – original draft, writing – review and editing. Maedeh Barkhordari Dashtkhaki: writing – review and editing, writing – original draft. Sharif Ebrahimi: writing – original draft, writing – review and editing. Mohamad Mahdi Aghajany: writing – review and editing, writing – original draft. Hamidreza Mousavi: writing – original draft, writing – review and editing. Masoud Kiany: writing – review and editing, writing – original draft. Mahsa Asadi Anar: writing – original draft, writing – review and editing, supervision, data curation, software, methodology, validation, conceptualization, investigation. Farzaneh Sadeghi Mahounak: writing – review and editing, writing – original draft, validation. Farbod Khosravi: writing – review and editing, validation, writing – original draft. Sara Zamani: conceptualization, investigation, writing – review and editing, writing – original draft. Sepehr Ramezanipour: writing – review and editing, writing – original draft. Nima Dehghan Banadaki: writing – original draft, writing – review and editing, validation, investigation, conceptualization.

Funding

The authors have nothing to report.

Disclosure

The lead/corresponding author (MAA) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Ethics Statement

The authors have nothing to report.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Policy on Using ChatGPT and Similar AI Tools

Artificial intelligence‐assisted language tools were used only to support language editing, formatting, and readability improvement. All study design decisions, screening, data extraction, interpretation, and final approval were performed by the authors, who reviewed and verified all AI‐assisted text and take full responsibility for the final manuscript.

Zarei A., Ehsani A., Ebrahimzedeh Pasha A., et al., “Artificial Intelligence Applications in Pediatric Dentistry: A Systematic Scoping Review With Visual Evidence Mapping,” Health Science Reports 9 (2026): e73490. 10.1002/hsr2.73490.

Data Availability Statement

Data sharing is not applicable to this article, as no new data were created or analyzed in this study.