Skip to main content
Karger Author's Choice logoLink to Karger Author's Choice
. 2026 Jul 16. Online ahead of print. doi: 10.1159/000553309

Prognostic Clinical Predictive Models for Dental Caries Using Artificial Intelligence: Methodological Considerations

Sergio E Uribe a,b,c,d,✉, Alonso Carrasco-Labra e, Falk Schwendicke c, Ilze Maldupa a,b
PMCID: PMC13485288  PMID: 42461868

Abstract

Background

Artificial intelligence (AI) is currently used to develop clinical predictive models for dental caries. However, most prognostic models lack key methodological components. Few reach clinical practice, and fewer demonstrate clinical benefit.

Summary

This narrative methodological review examines considerations for developing, validating, and implementing AI-based caries prognostic models. Based on established prediction model frameworks (TRIPOD+AI, PROBAST+AI, and PROGRESS), we contextualize eight critical phases for caries-specific application. Each phase addresses challenges specific to caries, e.g., clustering of teeth within patients, dominance of baseline caries experience as a predictor, and frequent absence of external validation. Addressing these phases reduces bias, improves reproducibility, and supports meaningful evaluation of clinical impact. AI-based prognostic models should serve as decision-support tools that inform clinician and patient choices rather than substitutes for clinical judgment.

Key Messages

(i) Most AI-based caries prognostic models are never implemented and fewer of those implemented demonstrate clinical benefit. Methodological deficiencies explain this gap. (ii) Caries prognostic research faces field-specific challenges such as retrospective design, clustering of teeth within patients, and dominance of baseline caries experience. Valid caries prognostic models require attention to eight phases, from problem selection through deployment and ongoing monitoring. (iii) Better discrimination (higher AUC) does not indicate clinical usefulness. Calibration and decision curve analysis are essential yet frequently absent from published studies. (iv) Researchers should report regulatory considerations and plan prospective impact assessment from study inception. Models generating risk estimates that influence clinical decisions qualify as software as a medical device, requiring regulatory compliance and demonstrated clinical impact before deployment.

Keywords: Caries prediction, Machine learning, Artificial intelligence, Study design

Introduction

Caries diagnosis and management strategies remain largely reactive [1], detecting lesions only after tissue damage has occurred [2]. This approach has failed to reduce the overall disease burden [1, 3]. Untreated caries lesions affect 2.3 billion people [4] and cost more than USD 540 billion annually in direct treatment costs and productivity losses [5]. The global prevalence of early childhood caries has remained unchanged for decades [3]. Even in Europe, caries prevalence is the highest globally despite high dentist availability [6], showing that current dental care models are insufficient [1]. If clinicians can identify high-risk individuals before caries reaches clinical detection, targeted personalized prevention might succeed [1, 7]. Predictive modeling generate a mathematical description of a process to make accurate, data-driven predictions about future outcomes. However, even with many predictive models developed, methodological limitations rather than statistical sophistication explain why few of them have reached clinical implementation [8] and even fewer demonstrate clinical benefits [9, 10]. Nowadays, machine learning (ML) is increasingly used to develop clinical predictive models (CPMs) for caries [11–13]. Caries prognostic research faces several unique challenges. First, the hierarchical structure of dental data creates clustering: teeth and surfaces cluster within patients, violating independence assumptions in most statistical methods. Using surfaces as the unit of analysis artificially inflates sample sizes and biases standard errors. Second, baseline caries experience dominates prediction in most models, meaning models often identify risk after disease onset rather than before, limiting their preventive value. Third, caries occurs in high-prevalence settings where treating all individuals may be more efficient than risk stratification, requiring careful consideration of when prognostic models add clinical value. Fourth, longitudinal caries data require accounting for censoring (loss to follow-up) and competing events (tooth extraction and restoration). Although TRIPOD+AI [14] and PROBAST+AI [15] provide reporting and risk-of-bias guidance, caries research generally requires additional domain-specific methodological considerations. For instance, systematic reviews using these frameworks identify some of the issues in caries predictive research, such as single data splits without external validation, reporting only discrimination without calibration assessment, inadequate sample sizes for the number of predictors, and failure to account for clustering [11–13]. These field-specific failures justify caries-focused methodological guidance that contextualizes general frameworks to caries research realities (Table 1).

Table 1.

Common methodological issues in published caries AI prediction studies

Methodological item Criterion Reported, % (n = 26) Methodological issue
Data quality Sample size calculation 8 Yes
Study design Preregistration 8 Yes
Model testing Internal validation 85 ​
External validation 15 Yes
Performance (any metric) Discrimination 92 ​
Calibration 15 Yes
Clinical utility 8 Yes
Reporting TRIPOD/TRIPOD+AI followed 19 Yes
Full model equation released 8 Yes
Data/code/calculator available 19 Yes
Deployment Clinical workflow integration 8 Yes
Impact assessment 4 Yes
Update/drift plan 0 Yes

Evaluation of 26 caries prediction models using AI published between 2020 and 2026 (Uribe and Maldupa [16]).

Previous reviews [11–13] have identified recurring methodological issues in caries prediction studies using TRIPOD+AI and PROBAST+AI, among others. This narrative review extends that work by adding the PROGRESS framework [17] to contextualize the different types of caries prognostic research, identify where ML may support further development, and address caries-specific methodological considerations. Its aim is to provide field guidance for researchers developing caries prognostic models, clinicians evaluating artificial intelligence (AI) tools for caries prediction and implementation, and editors reviewing manuscripts on caries prediction models.

CPMs: Definition and Evolution

CPMs estimate the probability of an individual having a condition (diagnostic models) or developing a future outcome (prognostic models) [18] to answer questions such as “Will this patient have caries in the next year?” (binary outcome) or “How many new caries lesions will this patient develop in the next 2 years?” (continuous outcome). The prognosis research strategy (PROGRESS) framework divides prognosis research into four stages [17]: (1) describing the natural history and outcomes of a condition, (2) identifying specific predictors that alter the disease course, (3) creating models to calculate the probability of outcomes for individuals, and (4) using these predictions to match patients with appropriate treatments (Table 2). Most caries prognostic research is type 2: identifying risk factor associations, ranking predictor importance, or reporting odds/relative risk, without generating individual prediction tools [7]. Type 3 prognostic model research remains rare in cariology [12] and the available evidence reveals several methodological flaws [11, 13]. This review focuses exclusively on type 3: the methodological requirements for developing and validating CPMs that reliably predict individual caries risk before disease onset.

Table 2.

Examples of dental caries clinical predictive model (CPM) studies classified according to the PROGRESS framework

Progress category Research question and examples Main findings
1. Fundamental prognosis or baseline risk What is the average risk/outcome in a population? Describes the average course of a disease and health outcomes under current care standards
What is the global prevalence of ECC? Pooled prevalence 48% [95% CI: 43–53]. Africa 30%, Americas 48%, Asia 52%, Europe 43% (Uribe et al. [3])
What is the caries progression rate in children? Pooled caries incidence rate: 0.11 [0.09–0.13] per person-year at risk. Annual DMFS increment ranged 0.07–1.77 (Hummel et al. [19])
2. Risk factors Which factors are known to be associated with the outcome? Finding specific predictors that alter the course of the disease
Is past caries associated with future caries? Past caries experience is the strongest single predictor. Moderate/good accuracy in preschoolers, limited in schoolchildren (Tellez et al. [20])
Is lesion activity associated with progression? Active lesions are more likely to progress than inactive ones. Severity at baseline is associated with progression (Abdalla et al. [21])
Are genetic factors associated with caries? Maybe. Single-nucleotide polymorphisms (SNPs) identified through genome-wide association studies (GWASs), although the evidence is inconclusive and varies across studies, with no common SNPs found (Zhou et al. [22])
3. Prognostic model Can we predict individual risk using multiple factors? Creating models to calculate the probability of outcomes for individuals
Can Cariogram predict caries increment? A strong correlation between caries risk profiles at baseline and caries incidence in the permanent teeth after 2 years was found (Campus et al. [23])
Can ML predict caries? Maybe, but the evidence is still limited, as most studies reviewed showed a significant risk of bias, and the machine learning models were primarily developed outside of real clinical settings (Reyes et al. [12])
4. Predictors of treatment effect Who benefits from treatment? Using risk estimates to guide clinical management and treatment selection
Does risk-stratified caries management improve treatment outcomes compared with simplified risk assessment? An ongoing randomized trial will test whether caries management guided by a detailed, multivariate risk assessment (CariesCare International 4D) results in different caries-related treatment outcomes compared with a simplified DMFT-based approach, evaluating risk stratification as a predictor of treatment benefit (Menegaz et al. [24])

In caries research, prognostic CPMs differ in their presentation of risk estimates. At the simplest level, some tools use additive or rule-based approaches to stratify individuals into ordinal risk categories (low/moderate/high) without estimating the absolute future risk. The ADA Caries Risk Assessment exemplifies this approach by classifying caries risk without a defined prediction horizon [20]. More structured approaches apply weighted scoring systems to produce refined risk strata linked to management strategies. CAMBRA represents this intermediate complexity by formalizing the combination of disease indicators, risk factors, and protective factors without providing an individual-level probability of future caries development [25]. Although risk scores may improve clinical usability, simplification often comes at the cost of predictive accuracy and calibration [26]. At the highest complexity level, probability estimation tools use statistical models to generate absolute risk estimates over defined time horizons. In the 1980s, de Dombal et al. [27] demonstrated that computer-aided multivariable models could generate calibrated, prospectively validated individual prognostic estimates to support clinical decision-making. Later, the Cariogram [28] integrated multiple predictors to estimate the 1-year probability of avoiding new caries [28]. Overall, these caries risk tools show modest predictive performance [20, 29]. Since the prediction is largely driven by baseline caries experience [30, 31], they often identify risk after disease onset.

AI for Caries Prognostic Models

Recently, AI and ML have been used to develop prognostic CPMs for caries research [12, 32]. A new CPM for medical conditions using AI is now developed approximately every 1.5 h worldwide [33], but few models reach clinical implementation [8], and even fewer demonstrate clinical benefits [9]. The increasing use of AI/ML for caries prognosis is driven by three factors. First, larger datasets, including electronic health records, insurance claims, and longitudinal cohorts, now provide access to hundreds of thousands of patient records [34]. Second, richer data sources beyond traditional clinical risk factors, such as salivary biomarkers, oral microbiome profiles, genomic variants, dietary tracking, and imaging, have become available [35]. Third, novel statistical methods and ML algorithms can handle high-dimensional data without requiring researchers to prespecify predictors or interactions [36].

However, systematic reviews comparing ML algorithms to traditional logistic regression for clinical prediction found no performance benefit [37]. Traditional regression methods can model nonlinearities effectively using splines without requiring massive datasets [38]. Furthermore, flexible ML algorithms frequently suffer from poor calibration [39, 40], and complex ML algorithms demand enormous amounts of data unless the signal-to-noise ratio is high [41]. Available studies on AI models for predicting dental caries demonstrated significant risks of bias and methodological deficiencies. Most models were developed outside real clinical settings, and many exhibited unclear or high risks of bias [12]. These methodological deficiencies limit the clinical applicability and effectiveness of the ML approaches and suggest that specific field guidance covering both reporting and the development of caries prediction models could be helpful. This narrative review focuses exclusively on AI methods based on supervised ML for caries prognosis using structured clinical data.

Methodological Considerations for Predictive AI Model Development

We illustrate the methodological requirements using a case scenario: developing a model to predict which caries-free children will develop new cavitated lesions. This scenario represents type 3 prognostic research under the PROGRESS framework. The child had no caries at baseline. The model predicts future disease occurrence before the appearance of clinical signs. Figure 1 summarizes the development and testing phases.

Fig. 1.

This diagram shows the four types of prognostic research, known as PROGRESS. There is also a subdiagram detailing PROGRESS 3, which focuses on prognostic models.

Methodological workflow for prognostic predictive model research in dental caries.

Problem Selection

van Smeden [18] proposed general preliminary questions before developing a prognostic model. In this review, we apply and adapt them to caries prediction. Is a model needed for the intended clinical decision? Is the target population clearly defined? Are sufficiently high-quality data and enough outcome events available? Does an existing model already address the same question in a similar setting? Are the candidate predictors routinely available in practice? Here, outcome events are defined at the patient level because caries risk factors, such as sugar intake and fluoride exposure, primarily operate at the individual level. If not, model development should generally not proceed. Before developing a new model, search for, critically appraise, and validate existing models for the same outcome and population. For caries, systematic reviews have identified multiple existing AI models with a high risk of bias and poor reporting [11–13]. Few have been externally validated and none implemented. It is recommended to develop a new caries predictive model when existing models use impractical predictors, predict different outcomes, lack sufficient reporting, or show poor performance that updating cannot fix (Table 3). Answering these preliminary questions requires patient and end-user involvement from the start. Clinicians are encouraged to identify workflow barriers and data availability constraints. Patients clarify which outcomes matter the most and what risk information influences decisions. For instance, parents may prioritize preventing pain over preventing noncavitated lesions. Patient and clinician collaboration ensures that the model addresses real clinical needs and produces precise clinical questions. Is it helpful to compare vague formulations like “Predict caries risk in children” with specific ones: “Among caries-free children aged 2–3 years attending primary care clinics, what is the probability of developing at least one cavitated lesion in permanent first molars by age 7 years?” defining the population, outcome, time horizon, and setting explicitly.

Table 3.

When to validate existing models vs. develop new models

Situation Action Rationale
Existing model for same outcome and population, well-reported Validate and update if needed Validation is faster and builds on prior work
Existing model uses impractical predictors (microbiome, genetics) for primary care setting Develop new model Implementation barriers cannot be overcome through updating
Existing model shows poor performance but uses appropriate predictors Validate and attempt updating first Updating may improve performance without new development
No existing model for the specific outcome Develop new model Clinical gap requires new development
Existing model insufficiently reported (missing coefficients or baseline hazard) Develop new model Cannot validate or implement models lacking essential information
Existing model predicts different outcome (white spot vs. cavitation) Develop new model Different outcomes serve different clinical decisions
Multiple existing models, none externally validated Validate best-performing model in new setting External validation reveals which models generalize

Beyond precisely defining questions, three implementation criteria determine whether prediction offers clinical value. First, does prediction change clinical decisions? Prediction windows must precede the current detection and allow intervention time. Predicting permanent molar caries at age 2–3 for outcomes at age 7 creates a 4-year prevention window. Second, are predictors routinely available? Models requiring sophisticated analysis, such as microbiome analysis, currently fail in routine implementation and models using salivary flow and fluoride exposure may succeed. Third, does prediction enable differential treatment? Without differential treatment strategies, prediction rarely offers clinical benefits.

Data Quality and Preparation

Data quality determines model validity more than algorithm choice, resting on three pillars: adequate sample size, complete predictor measurement, and standardized outcome definition. Traditional rules suggesting 10–20 events per predictor may often not apply [42], and stable predictions commonly require hundreds or thousands of outcome events. The required total sample size depends on the model complexity and caries baseline prevalence [42, 43]. Teeth and surfaces cluster within patients, creating correlated observations. In association studies, ignoring this clustering biases standard errors; in prognostic models, it influences sample-size calculations, model development, and performance estimation. Because incident caries events generally occur at the patient level, the patient is the natural unit of analysis unless strong justification exists for tooth-level prediction, in which case mixed-effects models or generalized estimating equations can account for within-patient correlation. Sample size calculations are based on patients (independent units), not teeth or surfaces. A study with 100 patients and 2,800 teeth still has only 100 independent observations. Treating teeth as independent units inflates effective sample size and biases standard errors [44]. If 30% of children develop cavitated lesions over 4 years, a cohort of 1,000 children yields 300 events (37.5 events per predictor for eight predictors), meeting minimum criteria. Formal sample size calculation before data collection is advisable [42, 43]. We provide worked examples with R code in https://github.com/sergiouribe/sample-size-prediction-models.

But even with sufficient volume, data completeness poses major challenges. Electronic health records provide large samples but suffer severe completeness problems. Analysis of 543,363 dental visits from the BigMouth dataset revealed pain ratings missing in 98–99% of records, alcohol use in more than half, dietary sugar in 40–60%, and drug use in >60% [45]. Missing data require handling based on the mechanism: missing completely at random, missing at random, or missing not at random [46]. Complete case analysis introduces bias unless data are missing completely at random, which is rare, and reduces effective sample size [26]. Multiple imputation is preferred, preserving sample size and reducing bias when data are missing at random [47]. However, imputation assumes that missingness can be predicted from observed data. Children missing dietary questionnaires may differ systematically from completers (missing not at random), making imputation unreliable [26]. Here, it is recommended to report missingness for each predictor and assess whether it relates to observed characteristics. Behavioral predictors with >50% missingness should generally be excluded or prospectively collected. For deployment, specify whether missing predictors will be excluded or handled using an imputation strategy aligned with model development. Standardizing measurements supports replication and valid prediction. Predictors should be measured consistently across participants using validated methods, as inconsistent measurements introduce noise that algorithms cannot overcome. Each predictor should be defined in detail: measurement tools, units, timing, and handling of abnormal values. For example, dietary sugar intake is best assessed with validated questionnaires rather than informal recall, and fluoride exposure benefits from clear documentation of water fluoridation, varnish use, toothpaste concentration, and brushing frequency. Similar precision applies to outcomes. Outcome definitions should be explicit, clinically meaningful, and consistent. Composite outcomes combining events of different clinical relevance can obscure prediction targets. Defining a new lesion, for instance, benefits from specifying diagnostic criteria such as an ICDAS code threshold. Transparent reporting of timing, detection methods, and examiner training strengthens outcome measurement. When using routine clinical data, describing existing documentation practices and considering heterogeneity in outcome assessment as a source of uncertainty is advisable. Reliance on routine care visits alone can bias outcome detection, as high-risk individuals may be examined more frequently.

Study Design

Cohort designs provide valid prognostic inferences [48]. Cross-sectional designs cannot establish temporal sequence between predictors and outcomes. Case-control designs introduce bias through retrospective predictor measurements. Prospective cohorts collect predictor data before outcomes occur and follow participants forward until the outcomes develop. Retrospective cohorts were used to extract predictor and outcome data from the existing records. Both can produce valid models if data quality requirements are met; however, retrospective designs face greater risks of incomplete predictor measurements and outcome misclassification. Cohort studies usually involve some loss to follow-up. Loss to follow-up exceeding 20% threatens validity. Regardless of the timing, preregistration of the study protocol before data collection is critical. Prospective registration documents predictor selection, outcome definitions, analysis plans, and sample size calculations before examining the data. This prevents data-driven decisions that inflate overfitting risk and reduce publication bias. Preregistration also increases transparency and allows readers to identify deviations from the planned analyses [49].

Also, enrollment criteria determine model validity. Inception cohorts enroll participants at a uniform point in their disease trajectory: the moment prediction would occur clinically [17] for example, children must be caries-free at enrollment (age 2–3 years). Including children with varying baseline caries experiences violates the inception cohort principles and confounds incident disease prediction with disease progression.

Finally, competing events require explicit consideration. Competing risks are events that prevent an outcome from occurring. For caries prediction, tooth extraction before the end of the prediction window represents a competing risk. Participants experiencing competing risks should preferably not be excluded without appropriate statistical handling [26]. If permanent first molars are extracted before age 7 years due to severe caries, excluding these children underestimates the true caries risk. Competing risk models or composite outcomes (cavitated lesion or teeth extracted due to caries) may be needed.

Model Development

Model development requires prespecified predictor selection, appropriate algorithm choice, and explicit overfitting control. Predictor selection in most cases is guided by prior prognostic evidence, clinical knowledge, and biological plausibility before examining the data [26]. Data-driven selection based on univariate p values increases overfitting and produces unstable models. Candidate predictors ideally have strong prior evidence and be routinely available in practice; each additional parameter increases sample size requirements. Continuous predictors should remain continuous unless clinically justified thresholds exist, as categorization wastes information and degrades performance [48, 50]. Nonlinear effects are recommended to be modeled using restricted cubic splines [51].

A distinction exists between association studies (PROGRESS type 2) and prognostic models (PROGRESS type 3). Association studies identify risk factors and estimate effects, requiring biological plausibility and causal interpretability. Prognostic models aim to predict individual outcomes accurately. Shmueli (2010) distinguished explanatory modeling (understanding mechanisms) from predictive modeling (forecasting outcomes), arguing each requires different variable-selection strategies. In predictive frameworks with large datasets, data-driven selection can identify patterns that improve accuracy even without clear biological mechanisms. However, for clinical prediction models intended for implementation, it is recommended to start with biologically plausible predictors. First, small dental datasets (typical N = 200–1,000) make overfitting a major risk when testing many variables. Second, biological plausibility improves transportability across populations. Third, clinicians trust models with mechanistically grounded predictors. Fourth, pre-specification reduces researcher degrees of freedom (Steyerberg et al. [38]). Data-driven selection can be appropriate when sample sizes exceed 5,000, prediction accuracy outweighs interpretability, the goal is hypothesis generation, or the model will undergo rigorous external validation before deployment.

Algorithm choice depends on data structure and sample size. Logistic regression suits binary outcomes with modest samples and provides interpretable coefficients. Random forests and gradient boosting handle nonlinear interactions but require larger samples and reduce interpretability. Deep learning demands tens of thousands of outcome events, rarely available in dental research [45]. Online supplementary Table 1 (for all online suppl. material, see https://doi.org/10.1159/000553309) summarizes algorithm characteristics. If logistic regression achieves adequate discrimination and calibration, more complex algorithms offer no advantage [50].

Outcome events should be defined at the patient level, as most caries risk factors act at the individual level. Sample size calculations must be based on patients experiencing the outcome [42, 43]. If prediction targets are tooth- or surface-level, clustering must be accounted for using the design effect, which depends on the number of units per patient and the intraclass correlation (typically 0.1–0.4 for dental caries).

When outcomes are imbalanced, for instance when few children develop new dentine caries over follow-up, then model stability and calibration suffer [39, 40]. Synthetic oversampling techniques such as SMOTE can distort outcome prevalence and degrade calibration, so they are best avoided [52]. Penalized regression methods (ridge or elastic net) are preferable under imbalance [38], combined with appropriate metrics including calibration plots and decision curve analysis.

Overfitting prevention requires penalization and adequate sample size. Hyperparameters must be optimized using resampling within the development dataset only, never using validation data (de Hond et al. [53]). The traditional rule of 10 events per parameter is context-dependent and often insufficient [54, 55]. We provide worked examples at https://github.com/sergiouribe/sample-size-prediction-models.

Model Validation

Validation assesses whether models perform accurately beyond the development dataset. There are two types of validation: internal validation using the development dataset and external validation using new patients from different settings. We acknowledge that “validation” is problematic terminology because models require continuous performance assessment rather than definitive validation [56]. We use “validation” for consistency with existing literature, though “testing” is preferable.

Internal validation detects overfitting by evaluating the performance on data not used for development. Single random data splitting should be avoided [50]. Splitting data once into development and testing sets wastes information and produces unstable estimates; any performance difference may reflect a random split rather than true overfitting [50]. Using the full dataset for training with bootstrap resampling (e.g., 500 iterations) for performance evaluation is preferable because it provides stable discrimination and calibration estimates while maximizing development data [54]. Internal validation/testing alone is insufficient and cannot assess model transportability to different populations.

External validation tests model performance in new patients not used for development. Table 4 presents the validation types assessing different aspects of transportability. External validation often reveals poor performance due to different outcome prevalence, predictor distributions, measurement methods, or model overfitting [56]. For instance, a model developed in dental school clinics, where 45% of children develop caries, may perform poorly in private practice, where only 15% develop caries. The sample size for external validation differs from development. The aim of the development is to build a stable model that does not overfit. For external validation, the aim is to precisely measure the model performance. Validation sample size must ensure precise estimation of performance measures, including calibration slope, C-statistic, and observed/expected ratio [57]. The required sample size depends on the target precision for confidence intervals (CIs; e.g., ≤0.1 for C-statistic, ≤0.3 for calibration slope) and the expected model performance in the validation population. Performance may vary across patient subgroups defined by sex, age, socioeconomic status, or baseline risk. Assess discrimination and calibration separately in relevant subgroups and transparently report performance heterogeneity [26]. Poor external validation does not necessarily indicate model failure. When external validation reveals miscalibration, updating the model for the new setting should be considered rather than discarding it [46]. Begin with recalibration of the intercept, then re-estimate the coefficients if needed [26]. Updating often improves performance more efficiently than developing new models. A model is never “validated” permanently; continuous assessment across settings and time establishes trustworthiness. Also, federated learning offers a potential approach for multicenter model development and validation while preserving data privacy [58]. Models are trained across institutions without sharing patient-level data: the algorithm moves to the data rather than data to a central repository. This approach could facilitate external validation across diverse caries populations while maintaining compliance with data-protection regulations. However, federated learning for caries prognostic models remains largely unexplored and requires infrastructure investment before adoption.

Table 4.

External validation types and purposes

Validation type Purpose Example
Temporal Assesses stability over time as populations and practices evolve Model developed 2015–2019, validated 2020–2024 in same clinics
Geographic Assesses transportability across regions Model developed in country A clinics, validated in country B clinics
Setting Assesses generalizability to new contexts Model developed in dental school clinics, validated in dental private practices
Population Assesses performance across patient subgroups Model developed in urban children, validated in rural children

Performance Metrics

Predictive model performance reports ideally must include (1) discrimination, (2) calibration, and (3) clinical utility. (1) Discrimination assesses whether individuals who develop caries are at a higher risk than those who do not develop caries; (2) calibration assesses whether the predicted risks correspond to the observed outcome frequencies; and (3) clinical utility evaluates whether using the model leads to better clinical decisions than using alternative strategies. All three are required before implementation can be decided, and discrimination alone is insufficient [59]. Table 5 explains these concepts for clinical readers (adapted from [59, 60]).

Table 5.

Performance metrics in prognostic models

Concept Question Usual metrics Interpretation
Discrimination Does the model correctly rank patients from lower to higher future caries risk? C-statistic (AUC), concordance index AUC = 0.75
Children who later develop new caries usually receive higher predicted risk than those who remain caries-free
Calibration Do predicted risks match what actually happens? Calibration plot, calibration slope, calibration-in-the-large Calibration slope = 0.60; predicted risk = 40%, observed risk = 20%
The model overestimates caries risk, which may lead to unnecessary intensive prevention despite acceptable discrimination
Clinical utility Does using the model lead to better preventive decisions than treating all or no patients? Decision curve analysis, net benefit Net benefit >0 at a 40% risk threshold. Using the model to intensify prevention above this threshold results in fewer unnecessary treatments than treating all children intensively

(1) Discrimination is commonly quantified using the C-statistic (AUC), ranging from 0.5 (no discrimination) to 1.0 (perfect), with 0.70–0.80 considered acceptable [26]. A C-statistic of 0.75 means the model assigns higher risk to the affected child 75% of the time when comparing one who developed caries with one who did not. However, CIs frequently span 0.10–0.15 units, so small differences between models are clinically meaningless [61]. For example, a model with C-statistic 0.75 (95% CI: 0.70–0.80) in development and 0.72 (95% CI: 0.65–0.79) in validation shows overlapping intervals; the apparent decline reflects random variation, not true performance loss. CIs must always be reported. More importantly, good discrimination does not mean predicted risks are accurate. A model may correctly rank children by risk while systematically over- or underestimating absolute probabilities (poor calibration), leading to inappropriate preventive decisions. This is where (2) calibration becomes essential. Calibration plots compare predicted risks against observed outcome frequencies; perfect calibration follows a 45° line. The calibration slope quantifies this relationship: 1.0 indicates perfect calibration, <1.0 indicates overfitting, and >1.0 indicates underfitting. Calibration-in-the-large assesses whether the average predicted risk matches the observed prevalence. Calibration is the “Achilles heel” of predictive models [39, 40]. A model with C-statistic 0.80 but calibration slope 0.60 predicts 60% caries risk for children whose observed risk is 40%, prompting unnecessary treatment and potentially harming low-risk children through overtreatment. Beyond discrimination and calibration, models must demonstrate (3) clinical utility. Decision curve analysis estimates net benefit by weighing true positives against false positives according to clinical consequences [62]. For instance, if treatment intensity increases when predicted risk exceeds 40%, the analysis compares treating all children intensively, treating none, or using model-based stratification. A useful model shows higher net benefit across relevant thresholds. However, decision curve analysis on development or validation data only estimates potential utility. Demonstrating actual outcome improvement requires prospective impact studies comparing model-guided care with usual care, ideally through pragmatic randomized trials [53].

Reporting

Transparent reporting enables readers to assess validity, reproduce findings, and implement models. TRIPOD+AI provides comprehensive guidelines for predictive model studies using regression or AI/ML methods [14]. Its 27-item checklist includes AI-specific requirements such as hyperparameter tuning, model architecture, computing environment, and code availability, not just “we used a random forest” but specifying the number of trees, maximum depth, minimum samples per leaf, software version, and random seed. Despite this guidance, reporting quality remains poor [12], and current omissions render most published caries AI models unusable. Table 6 summarizes essential reporting elements with complete and incomplete examples. Risk predictions carry inherent uncertainty from three sources (Barreñada et al. [63]). Estimation uncertainty arises from finite sample sizes and can be quantified through resampling or Bayesian methods. Model uncertainty reflects variability from modeling choices, including predictor selection and algorithm. Applicability uncertainty arises from differences between development and deployment settings in populations, measurements, and workflows. Model and applicability uncertainty often dominate estimation uncertainty but remain invisible without external validation. Reporting ideally includes uncertainty around individual risk estimates, not only population-level performance, and full model publication [63]. Beyond reporting, PROBAST+AI provides quality assessment for critical appraisal in systematic reviews and self-assessment during study design [64], evaluating risk of bias across four domains: participants, predictors, outcome, and analysis.

Table 6.

Essential reporting elements for caries prognostic models

Element What to report Complete reporting example Incomplete reporting example
Model specification Complete equation or algorithm with all coefficients, intercept/baseline hazard, predictor definitions, coding schemes; for ML: algorithm type, hyperparameters, software versions, random seeds Logistic regression: log(odds) = −2.43 + 0.85 × DMFT + 0.12 × age_months (centered at 30) - 0.65 × fluoride (1 = yes, 0 = no). Software: R 4.3.0, glm function We used XGBoost to predict caries risk
Predictor definitions Measurement methods, timing, units, missing data handling for each predictor Dietary sugar frequency: number of sugar-containing foods/beverages per day, assessed by validated 24-h self-reported, collected at baseline by trained assistants. Missing: 8.3%, imputed using multiple imputation Diet was assessed
Outcome definitions Diagnostic criteria, detection methods, timing, examiner reliability New cavitated caries: ICDAS 5–6 detected by calibrated examiners (kappa = 0.85) using visual-tactile examination at 84-month follow-up Caries was assessed at follow-up
Performance metrics Discrimination (C-statistic with 95% CI), calibration (plot, slope, calibration-in-the-large), clinical utility (decision curve analysis) for development and validation Development: C-statistic 0.76 (95% CI: 0.72–0.80), calibration slope 0.95, calibration-in-the-large 0.02. Validation: C-statistic 0.73 (95% CI: 0.67–0.79), calibration slope 0.88 The model showed good discrimination (AUC = 0.76)
Sample size for development Sample size calculation targeting adequate precision for predictor effects and minimal overfitting; number of candidate parameters; outcome prevalence; anticipated performance; target shrinkage; participant flow; missing data Sample size calculated using pmsampsize (indicate package) with 16 candidate parameters, prevalence 0.35, anticipated C-statistic 0.70, targeting shrinkage ≥0.90. Minimum required: 784 participants (274 events). Actual: 967 analyzed (338 events). DMFT missing: 2.1%; dietary data: 8.3%; multiple imputation with 50 datasets 1,000 children were included in the study
Sample size for validation Sample size calculation targeting precise estimation of model performance; target CI widths for C-statistic and calibration slope; expected performance in validation population; participant flow; missing data Sample size calculated using pmvalsampsize targeting CI width ≤0.1 for C-statistic and ≤0.3 for calibration slope. With prevalence 0.40 and expected C-statistic 0.73, minimum required: 542 participants (217 events). Actual: 568 analyzed (227 events) from 12 geographically distinct clinics The model was validated in 200 children/Model was validated in 30% of the original dataset
Missing data Frequency per predictor; missingness pattern (MCAR/MAR/MNAR); handling method with justification; number of imputations and software; sensitivity analysis comparing approaches Predictor missingness: DMFT 2.1%, dietary data 8.3%, fluoride 0.5%. MAR assumed (missingness associated with lower SES). Multiple imputation (50 datasets) including all predictors, outcome, and auxiliary predictors; mice package v3.16. Complete case sensitivity analysis: C-statistic 0.74 vs. 0.76 with imputation Missing data were imputed
Code and model access Public repository links, calculator URLs, executable code with persistent identifiers Code: github.com/user/caries-model (DOI: 10.12345/repository.123456). Calculator: URL. R function provided in online supplementary materials Code/Data available upon reasonable request

Deployment and Maintenance

Most prognostic models never reach clinical implementation. Few progress beyond external validation to deployment, and fewer still demonstrate improved decision-making or patient outcomes [9]. This represents substantial research waste: models developed, validated, and published but never used. Table 7 summarizes implementation barriers, strategies, monitoring approaches, and updating methods. Models requiring data not collected in routine workflows face immediate implementation failure regardless of statistical performance [53]. Planning for implementation from problem selection onward, with multidisciplinary teams of clinicians, data scientists, and IT specialists, increases the chances of success. Models integrated into existing patient management systems achieve higher adoption than standalone applications [9]. Regulatory considerations also apply. In the European Union, prognostic CPMs are generally classified as Class IIa medical devices (MDR 2017/745, Rule 11). UK guidance similarly regulates software as a device when its internal calculations cannot be easily verified by the user, as is typical of complex AI models. In the USA, software calculating disease risk probabilities is typically ineligible for exemption as clinical decision support and remains subject to FDA regulation [65].

Table 7.

Deployment and maintenance framework for caries prognostic models

Phase Challenge Solution/strategy Example for caries model
Implementation barriers Data unavailable in routine workflow Select only predictors collected routinely Use fluoride exposure from standard exams; avoid microbiome analysis
Complex models (>10 predictors) Limit predictor number, prioritize parsimony 6–8 predictors maximum for clinical usability
Black-box algorithms Use interpretable models or provide explanations Logistic regression with coefficients; decision curve analysis showing clinical benefit
Dental patient management system integration Establish multidisciplinary teams early Include clinicians, data scientists, IT specialists from problem selection phase
Implementation strategies Dental patient management system integration Embed predictions in clinical workflow Display risk scores during patient check-in or examination documentation
Web-based calculator Provide standalone access Online tool accepting manual predictor entry for settings without EHR integration
Mobile application Enable point-of-care use Smartphone app for community health workers or mobile dental clinics
Code repository Ensure reproducibility Deposit R/Python code in public repository with persistent DOI
Impact assessment Unknown clinical benefit Conduct pragmatic trials Cluster-randomized trial: clinics using model vs. usual care, outcome = caries at 4 years
Potential harm from predictions Evaluate false positive burden Assess unnecessary intensive fluoride in low-risk children incorrectly classified
Lack of evidence Require impact assessment before widespread adoption Document improved prevention allocation or reduced caries incidence
Monitoring triggers Performance drift detection Define reassessment intervals Annual calibration assessment in new patient cohorts
Calibration degradation Set calibration thresholds Update if calibration slope <0.9 or calibration-in-the-large >0.1
Discrimination decline Set discrimination thresholds Update if C-statistic drops >0.05 from initial performance
Equity concerns Monitor subgroup performance Quarterly assessment across socioeconomic groups, baseline risk levels
Updating strategies Baseline risk shift Recalibration (intercept adjustment) Fluoride policy reduces caries prevalence 30%→15%; adjust baseline prediction
Coefficient drift Re-estimation (all coefficients) Dietary patterns change; re-estimate predictor effects in new data
New predictors available Model extension Dental sealant coverage becomes routinely documented; add as predictor
Fundamental model failure Complete redevelopment Outcome definition changes (ICDAS criteria updated); develop new model
Governance Unclear responsibility Define roles before deployment Specify: who monitors (data scientist), who approves updates (clinical lead), how often (annually)
Version control Document all changes Log every model version, performance metrics, and modification date in registry
Accountability Establish oversight committee Multidisciplinary committee reviews monitoring data quarterly, approves updates

Demonstrating clinical value calls for formal impact assessment, evaluating whether models improve outcomes compared with current practice. Without this step, models risk causing harm through overtreatment of false positives or missed prevention from false negatives. Even successfully implemented models require ongoing maintenance. Populations change, disease epidemiology shifts, and clinical practice evolves, making performance drift inevitable [56]. A model developed when fluoride varnish was uncommon may miscalibrate where varnish is now routine. Yet only 13% of published medical AI-CPMs include updating strategies [9]. Monitoring protocols benefit from specifying reassessment intervals (e.g., annually), performance thresholds triggering updates (calibration slope <0.9, C-statistic drop >0.05), and subgroup analyses examining equity across populations. When drift is detected, updating options include neural-network-specific approaches (adjusting retrained layers), ensemble methods (reweighting members), and model-agnostic strategies [66]. Updating an existing model is generally more efficient than developing a new one, preserving accumulated knowledge while adapting to new contexts. Governance structures, clarifying who monitors performance, initiates updates, and approves changes, help ensure systematic maintenance. Documenting all model versions, changes, and performance metrics over time is equally important; models without updating strategies face eventual degradation and potential patient harm. Figure 2 shows an example of the workflow of a hypothetical caries prediction model that uses AI during the development and then triggering automatically the necessity of recalibration.

Fig. 2.

The figure illustrates the eight steps involved in the development and validation or testing of a clinical prognostic model for dental caries.

An example of the workflow of a caries prognostic model.

Conclusion

AI methods offer opportunities to predict which patients will develop caries or experience lesion progression, but current evidence shows that most caries prognostic models remain early stage, overoptimistic, inconsistently reported, and rarely linked to interventions proven to modify the disease course. Most models prioritize discrimination metrics while neglecting calibration, uncertainty, and clinical impact. Better predictions do not automatically translate into better outcomes. Experience with diagnostic AI shows that earlier detection can increase intervention without improving health if predictions trigger premature or unnecessary treatment [67]. Individual risk predictions remain inherently uncertain due to estimation, model, and applicability uncertainty, even when population-level performance appears strong [63]. Transparent reporting, external validation, and explicit communication regarding uncertainty are essential. Conceptual frameworks for AI-driven precision dentistry propose linking caries risk estimates to personalized preventive strategies [7]; however, such approaches have rarely been evaluated in pragmatic trials with caries endpoints. Many AI-based prognostic models qualify as software as a medical device when they provide risk estimates that influence decisions, requiring a clear definition of intended use, rigorous validation, and post-deployment monitoring. Academic incentives prioritize novelty over implementation: publishing another model advances careers more than integrating an established model into clinical workflows. This systemic bias against implementation prevents patient benefits. Successful models share characteristics beyond statistical performance: they use routinely available data, link to clear clinical actions (What specific decision does this prediction inform?), and function outside academic centers of excellence. Implementation planning preferably begins during protocol development, not after publication. Every prediction model manuscript should include an implementation section: What needs to happen for dentists to use this tool? What validations are needed? If the model serves hypothesis generation rather than patient care, state this explicitly [68]. Journals should require evidence of clinical utility or concrete implementation plans before accepting prediction model manuscripts. Funders should prioritize prospective validation and implementation of existing models over development of new models addressing the same questions. Researchers should begin with the clinical problem and implementation barriers, not with available data and algorithms.

Conflict of Interest Statement

S.E.U. and F.S. are members of the Board of the ITU/WHO/WIPO Global Initiative on Artificial Intelligence for Health – Dental Diagnostics and Digital Dentistry. This role is unpaid and did not influence the conception, writing, or conclusions of this manuscript. All other authors declare that they have no conflicts of interest to disclose.

Funding Sources

This work was funded by the Scientist Grant “EpiDentLatvia” (RSU-ZG-2024/1-0044), as part of the project “RSU Internal and RSU with LASE External Consolidation” (Project No. 5.2.1.1.i.0/2/24/I/CFLA/005), funded by the European Union Recovery and Resilience Facility and the state budget of the Republic of Latvia. S.E.U. and I.M. acknowledge financial support from the Latvian Council of Science (project No. lzp-2022/1-0047, “IEVA Project”). S.E.U. also acknowledges support from the European Union’s Horizon 2020 Research and Innovation Programme (grant agreement No. 857287) for the Baltic Biomaterials Centre of Excellence. The funders had no role in the study design, data collection, data analysis, interpretation, manuscript preparation, or decision to publish.

Author Contributions

S.E.U. conceived the manuscript, led the conceptual framework, drafted the initial version, and coordinated revisions; A.C.-L., F.S., and I.M. contributed to the methodological interpretation and manuscript revision. Paperpal and Grammarly were used for grammar and style improvements and to ensure adherence to journal style guidelines. AI tool Grammarly was used to improve grammar, spelling, and clarity. It did not contribute to the scientific content, interpretation, or conclusions. All authors critically reviewed the manuscript, approved the final version, and agreed to be accountable for all aspects of the work.

Funding Statement

This work was funded by the Scientist Grant “EpiDentLatvia” (RSU-ZG-2024/1-0044), as part of the project “RSU Internal and RSU with LASE External Consolidation” (Project No. 5.2.1.1.i.0/2/24/I/CFLA/005), funded by the European Union Recovery and Resilience Facility and the state budget of the Republic of Latvia. S.E.U. and I.M. acknowledge financial support from the Latvian Council of Science (project No. lzp-2022/1-0047, “IEVA Project”). S.E.U. also acknowledges support from the European Union’s Horizon 2020 Research and Innovation Programme (grant agreement No. 857287) for the Baltic Biomaterials Centre of Excellence. The funders had no role in the study design, data collection, data analysis, interpretation, manuscript preparation, or decision to publish.

Supplementary Material.

References

  • 1. Innes NPT, Chu CH, Fontana M, Lo ECM, Thomson WM, Uribe S, et al. A century of change towards prevention and minimal intervention in cariology. J Dent Res. 2019;98(6):611–7. [DOI] [PubMed] [Google Scholar]
  • 2. Lundbeck HJ, Pitchika V, Wilson P, Raggio DP, Galloway J, Al-Yaseen W, et al. Dental practitioners’ thresholds for restorative intervention in carious lesions: a survey based systematic review update. Caries Res. 2026;60(1):65–79. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Uribe SE, Innes N, Maldupa I. The global prevalence of early childhood caries: a systematic review with meta-analysis using the WHO diagnostic criteria. Int J Paediatr Dent. 2021;31(6):817–30. [DOI] [PubMed] [Google Scholar]
  • 4. GBD 2017 Oral Disorders Collaborators; Bernabe E, Marcenes W, Hernandez CR, Bailey J, Abreu LG, et al. Global, regional, and national levels and trends in burden of oral conditions from 1990 to 2017: a systematic analysis for the global burden of disease 2017 study. J Dent Res. 2020;99(4):362–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Righolt AJ, Jevdjevic M, Marcenes W, Listl S. Global-regional-and country-level economic impacts of dental diseases in 2015. J Dent Res. 2018;97(5):501–7. [DOI] [PubMed] [Google Scholar]
  • 6. WHO . WHO/Europe calls for urgent action on oral disease as highest rates globally are recorded in European Region. WHO/News/Europe; 2023. Available from: https://www.who.int/europe/news/item/20-04-2023-who-europe-calls-for-urgent-action-on-oral-disease-as-highest-rates-globally-are-recorded-in-european-region (accessed 10 May 2023). [Google Scholar]
  • 7. Schwendicke F, Krois J. Precision dentistry-what it is, where it fails (yet), and how to get there. Clin Oral Investig. 2022;26(4):3395–403. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. van Royen FS, Moons KGM, Geersing G-J, van Smeden M. Developing, validating, updating and judging the impact of prognostic models for respiratory diseases. Eur Respir J. 2022;60(3):2200250. [DOI] [PubMed] [Google Scholar]
  • 9. Saelmans A, Seinen T, Pera V, Markus AF, Fridgeirsson E, John LH, et al. Implementation and updating of clinical prediction models: a systematic review. Mayo Clin Proc Digit Health. 2025;3:100228. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Arshi B, Cowley LE, Rijnhart E, Reeve K, Smits LJ, Wynants L. External validation, impact assessment and clinical utilization of clinical prediction models: a prospective cohort study. J Clin Epidemiol. 2025;186:111902. [DOI] [PubMed] [Google Scholar]
  • 11. Havsed K, Hänsel Petersson G, Isberg P-E, Pigg M, Svensäter G; Foresight Research Consortium; et al. Multivariable prediction models of caries increment: a systematic review and critical appraisal. Syst Rev. 2023;12(1):202. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Reyes LT, Knorst JK, Ortiz FR, Ardenghi TM. Machine learning in the diagnosis and prognostic prediction of dental caries: a systematic review. Caries Res. 2022;56(3):161–70. [DOI] [PubMed] [Google Scholar]
  • 13. Wang X, Zhang P, Lu H, Luo D, Yang D, Li K, et al. Risk prediction models for dental caries in children and adolescents: a systematic review and meta-analysis. BMJ Open. 2025;15(3):e088253. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Moons KGM, Damen JAA, Kaul T, Hooft L, Andaur Navarro C, Dhiman P, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Uribe S, Maldupa I. Methodological quality of machine learning-based caries prognostic models: a scoping review. Zenodo; 2026. [Google Scholar]
  • 17. Hemingway H, Croft P, Perel P, Hayden JA, Abrams K, Timmis A, et al. Prognosis research strategy (PROGRESS) 1: a framework for researching clinical outcomes. BMJ. 2013;346:e5595. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. van Smeden M, Reitsma JB, Riley RD, Collins GS, Moons KG. Clinical prediction models: diagnosis versus prognosis. J Clin Epidemiol. 2021;132:142–5. [DOI] [PubMed] [Google Scholar]
  • 19. Hummel R, Akveld NAE, Bruers JJM, van der Sanden WJM, Su N, van der Heijden GJMG. Caries progression rates revisited: a systematic review. J Dent Res. 2019;98(7):746–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Tellez M, Gomez J, Pretty I, Ellwood R, Ismail AI. Evidence on existing caries risk assessment systems: are they predictive of future caries? Community Dent Oral Epidemiol. 2013;41(1):67–78. [DOI] [PubMed] [Google Scholar]
  • 21. Abdalla H, Allison PJ, Madathil SA, Veronneau JE, Pustavoitava N, Tikhonova S. Caries lesions progression in adults: a prospective 2-year cohort study. Community Dent Oral Epidemiol. 2025;53(1):33–41. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Zhou K, Gao C, Wu J, Banerjee A, Ide M, Kang J. Genome-wide association studies on dental caries: a systematic review. Caries Res. 2025:1–20. [DOI] [PubMed] [Google Scholar]
  • 23. Campus G, Cagetti MG, Sale S, Carta G, Lingström P. Cariogram validity in schoolchildren: a two-year follow-up study. Caries Res. 2012;46(1):16–22. [DOI] [PubMed] [Google Scholar]
  • 24. Menegaz AM, Oliveira TTV, Braga MM, Raggio DP, Cenci MS, Mendes FM, et al. Randomized clinical trial to evaluate two methods of caries risk assessment in schoolchildren: the CARDEC-PEL 04 study protocol. BMC Oral Health. 2021;21(1):654. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Young DA, Kutsch VK, Whitehouse J. A clinician’s guide to CAMBRA: a simple approach. Compend Contin Educ Dent. 2009;30(2):92–4. 96, 98, passim. [PubMed] [Google Scholar]
  • 26. Moons KGM, Altman DG, Reitsma JB, Ioannidis JPA, Macaskill P, Steyerberg EW, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. 2015;162(1):W1–73. [DOI] [PubMed] [Google Scholar]
  • 27. de Dombal FT, Clamp SE, Softley A, Unwin BJ, Staniland JR. Prediction of individual patient prognosis: value of computer-aided systems. Med Decis Mak. 1986;6(1):18–22. [DOI] [PubMed] [Google Scholar]
  • 28. Hänsel Petersson G, Twetman S, Bratthall D. Evaluation of a computer program for caries risk assessment in schoolchildren. Caries Res. 2002;36(5):327–40. [DOI] [PubMed] [Google Scholar]
  • 29. Vanobbergen J, Martens L, Lesaffre E, Bogaerts K, Declerck D. The value of a baseline caries risk assessment model in the primary dentition for the prediction of caries incidence in the permanent dentition. Caries Res. 2001;35(6):442–50. [DOI] [PubMed] [Google Scholar]
  • 30. Christian B, Armstrong R, Calache H, Carpenter L, Gibbs L, Gussy M. A systematic review to assess the methodological quality of studies on measurement properties for caries risk assessment tools for young children. Int J Paediatr Dent. 2018. [DOI] [PubMed] [Google Scholar]
  • 31. Feldens CA, Braga VS, Kramer PF, Vítolo MR, Rodrigues PH, de Barros Coelho EMR, et al. Primary dentition caries patterns as predictors of permanent dentition caries: a prospective cohort study. Caries Res. 2023;57(2):167–76. [DOI] [PubMed] [Google Scholar]
  • 32. Hung M, Voss MW, Rosales MN, Li W, Su W, Xu J, et al. Application of machine learning for diagnostic prediction of root caries. Gerodontology. 2019;36(4):395–404. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Arshi B, Smits LJ, Wynants L, Cowley LE, Reeve K, Rijnhart E. Number of publications on new clinical prediction models: a systematic literature search. [DOI] [PMC free article] [PubMed]
  • 34. Walji MF, Spallek H, Kookal KK, Barrow J, Magnuson B, Tiwari T, et al. BigMouth: development and maintenance of a successful dental data repository. J Am Med Inform Assoc. 2022;29(4):701–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Divaris K. Precision dentistry in early childhood: the central role of genomics. Dent Clin North Am. 2017;61(3):619–25. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Bzdok D, Altman N, Krzywinski M. Statistics versus machine learning. Nat Methods. 2018;15(4):233–4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Christodoulou E, Ma J, Collins GS, Steyerberg EW, Verbakel JY, Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. 2019;110:12–22. [DOI] [PubMed] [Google Scholar]
  • 38. Steyerberg EW, Uno H, Ioannidis JPA, van Calster B; Collaborators . Poor performance of clinical prediction models: the harm of commonly applied methods. J Clin Epidemiol. 2018;98:133–43. [DOI] [PubMed] [Google Scholar]
  • 39. Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW; Topic Group “Evaluating diagnostic tests and prediction models” of the STRATOS initiative . Calibration: the Achilles heel of predictive analytics. BMC Med. 2019;17(1):230. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Trottini M, Campus G, Corridore D, Cocco F, Cagetti MG, Vigo MI, et al. Assessing the predictive performance of probabilistic caries risk assessment models: the importance of calibration. Caries Res. 2020;54(3):258–65. [DOI] [PubMed] [Google Scholar]
  • 41. van der Ploeg T, Austin PC, Steyerberg EW. Modern modelling techniques are data hungry: a simulation study for predicting dichotomous endpoints. BMC Med Res Methodol. 2014;14:137. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42. Riley RD, Snell KI, Ensor J, Burke DL, Harrell FE Jr, Moons KG, et al. Minimum sample size for developing a multivariable prediction model: PART II - binary and time-to-event outcomes. Stat Med. 2019;38(7):1276–96. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Riley RD, Snell KIE, Ensor J, Burke DL, Harrell FE Jr, Moons KGM, et al. Minimum sample size for developing a multivariable prediction model: Part I - continuous outcomes. Stat Med. 2019;38(7):1262–75. [DOI] [PubMed] [Google Scholar]
  • 44. Chow DY, Tay JRH, Nascimento GG. Systematic review of prognosis models in predicting tooth loss in periodontitis. J Dent Res. 2024;103(6):596–604. [DOI] [PubMed] [Google Scholar]
  • 45. Tiwari T, Kondratenko M, Nasiha N, Ong T, Chandrasekaran S, Kostbade G, et al. Evaluating the completeness of electronic health records in dental education: a big data study. Front Oral Health. 2025;6:1535164. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Altman DG, Vergouwe Y, Royston P, Moons KGM. Prognosis and prognostic research: validating a prognostic model. BMJ. 2009;338:b605. [DOI] [PubMed] [Google Scholar]
  • 47. van Buuren S. Flexible imputation of missing data. 2nd ed. London, England: Taylor & Francis; 2021. [Google Scholar]
  • 48. Royston P, Moons KGM, Altman DG, Vergouwe Y. Prognosis and prognostic research: developing a prognostic model. BMJ. 2009;338:b604. [DOI] [PubMed] [Google Scholar]
  • 49. Andaur Navarro CL, Damen JAA, Takada T, Nijman SWJ, Dhiman P, Ma J, et al. Risk of bias in studies on prediction models developed using supervised machine learning techniques: systematic review. BMJ. 2021;375:n2281. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50. Steyerberg EW. Validation in prediction research: the waste by data splitting. J Clin Epidemiol. 2018;103:131–3. [DOI] [PubMed] [Google Scholar]
  • 51. Lopez-Ayala P, Riley RD, Collins GS, Zimmermann T. Dealing with continuous variables and modelling non-linear associations in healthcare data: practical guide. BMJ. 2025;390:e082440. [DOI] [PubMed] [Google Scholar]
  • 52. van den Goorbergh R, van Smeden M, Timmerman D, Van Calster B. The harm of class imbalance corrections for risk prediction models: illustration and simulation using logistic regression. J Am Med Inform Assoc. 2022;29(9):1525–34. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53. de Hond AAH, Leeuwenberg AM, Hooft L, Kant IMJ, Nijman SWJ, van Os HJA, et al. Guidelines and quality criteria for artificial intelligence-based prediction models in healthcare: a scoping review. NPJ Digit Med. 2022;5(1):2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54. Riley RD, Ensor J, Snell KIE, Archer L, Whittle R, Dhiman P, et al. Importance of sample size on the quality and utility of AI-based prediction models for healthcare. Lancet Digit Health. 2025;7(6):100857. [DOI] [PubMed] [Google Scholar]
  • 55. van Smeden M, de Groot JAH, Moons KGM, Collins GS, Altman DG, Eijkemans MJC, et al. No rationale for 1 variable per 10 events criterion for binary logistic regression analysis. BMC Med Res Methodol. 2016;16:163. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 56. Van Calster B, Steyerberg EW, Wynants L, van Smeden M. There is no such thing as a validated prediction model. BMC Med. 2023;21:70–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 57. Riley RD, Snell KIE, Archer L, Ensor J, Debray TPA, van Calster B, et al. Evaluation of clinical prediction models (part 3): calculating the sample size required for an external validation study. BMJ. 2024;384:e074821. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 58. Schneider L, Rischke R, Krois J, Krasowski A, Büttner M, Mohammad-Rahimi H, et al. Federated vs local vs central deep learning of tooth segmentation on panoramic radiographs. J Dent. 2023;135:104556. [DOI] [PubMed] [Google Scholar]
  • 59. Van Calster B, Collins GS, Vickers AJ, Wynants L, Kerr KF, Barreñada L, et al. Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: overview and guidance. Lancet Digit Health. 2025;7(12):100916. [DOI] [PubMed] [Google Scholar]
  • 60. Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010;21(1):128–38. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 61. van Leeuwen FD, Steyerberg EW, van Klaveren D, Wessler B, Kent DM, van Zwet EW. Instability of the AUROC of clinical prediction models. Stat Med. 2025;44(5):e70011. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 62. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Mak. 2006;26(6):565–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 63. Barreñada L, Steyerberg EW, Timmerman D, Thomassen D, Wynants L, Calster BV. The fundamental problem of risk prediction for individuals: health AI, uncertainty, and personalized medicine. arXiv. 2025;1. [Google Scholar]
  • 64. Collins GS, Dhiman P, Andaur Navarro CL, Ma J, Hooft L, Reitsma JB, et al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open. 2021;11(7):e048008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 65. Alderman J, Riley R, Parekh D, Summers C, Liu X, Denniston A. Hidden risks of predictive models in healthcare. BMJ Evid Based Med. 2025:bmjebm-2025-113730. [DOI] [PubMed] [Google Scholar]
  • 66. Meijerink LM, Dunias ZS, Leeuwenberg AM, de Hond AAH, Jenkins DA, Martin GP, et al. Updating methods for AI-based clinical prediction models: a scoping review. J Clin Epidemiol. 2024;178:111636. [DOI] [PubMed] [Google Scholar]
  • 67. Schwendicke F, Rossi JG, Göstemeyer G, Elhennawy K, Cantu AG, Gaudin R, et al. Cost-effectiveness of artificial intelligence for proximal caries detection. J Dent Res. 2021;100(4):369–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 68. Markowetz F. All models are wrong and yours are useless: making clinical prediction models impactful for patients. NPJ Precis Oncol. 2024;8(1):54. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials


Articles from Caries Research are provided here courtesy of Karger Publishers

RESOURCES