Development of a Framework for Predictive Dental Caries Risk Modelling Using Integrated Patient Data

Thesis Abstract

The study addresses the persistent challenge of accurately predicting individual dental caries risk in diverse populations by leveraging integrated patient data from clinical records, epidemiological surveys, and lifestyle, genetic, and microbiome indicators. Despite advances in caries epidemiology, existing risk models often rely on limited data sources and fail to generalize across settings, hindering personalized prevention strategies and efficient allocation of preventive resources. The aim is to develop a robust framework for predictive dental caries risk modelling that synthesizes multi-source patient data into a parsimonious, clinically actionable model. Specific objectives include (1) identifying and harmonizing relevant data modalities across clinical, behavioral, genetic, and microbial domains; (2) evaluating the predictive performance of machine learning and traditional statistical approaches for caries risk estimation; (3) developing a modular, interoperable framework that supports real-time risk scoring within routine dental care; (4) validating the framework in two demographically distinct cohorts to assess generalizability; and (5) outlining governance, ethical, and implementation considerations for integration into dental practice. A mixed-methods, multisite study design will be employed. The population comprises patients aged 6–18 and 19–65 from three tertiary dental centers and two community dental clinics, with an anticipated total sample of 6,000 participants for model development and 2,000 for external validation. Data collection will integrate electronic health records (demographics, past caries experience, fluoride exposure, treatment history), clinical examination findings (dmft/DMFT, lesion activity, occlusal pattern), salivary biomarkers (pH, flow rate, mutans streptococci quantified by qPCR), genomic variants linked to caries susceptibility, and lifestyle factors (dietary sugar intake, oral hygiene practices, socioeconomic status). Standardized data harmonization protocols and privacy-preserving techniques will be implemented. Instruments include calibrated caries assessment tools, validated dietary and behavior questionnaires, saliva assays, and genotyping arrays targeting known caries-associated loci. Validity and reliability will be established through pilot testing, inter-examiner calibration (kappa >0.80 for lesion scoring), and reliability testing for laboratory assays (intra-assay CV <10%). Analytical methods will commence with data preprocessing, missing data imputation (multiple imputation by chained equations), and feature engineering to derive composite risk indicators. Predictive modelling will compare logistic regression, random forests, gradient boosting (XGBoost), and deep learning architectures capable of handling heterogeneous data. Model performance will be evaluated using discrimination (AUC/ROC), calibration plots, Brier score, and decision-curve analysis to determine clinical usefulness. Nested cross-validation will optimize hyperparameters, while feature importance and SHAP (SHapley Additive exPlanations) values will be employed to interpret model decisions. The theoretical underpinning will draw on the Health Belief Model to contextualize behavior-related predictors and the Ecological Systems Theory to justify multi-level data integration. A conceptual framework will illustrate data flow from acquisition to real-time risk scoring, with modular components enabling plug-and-play substitutions as new biomarkers emerge. Expected findings include (a) identification of a minimal, high-yield feature set achieving AUCs ?0.80 in internal validation and ?0.75 in external validation; (b) demonstration that integrated data (clinical + behavioral + biological) outperforms models restricted to traditional clinical indicators; (c) evidence that model calibration remains robust across age groups and settings, with systematic biases identified and corrected; and (d) a practical risk-scoring prototype with a user-centric interface suitable for incorporation into electronic dental records. The study will contribute to knowledge by operationalizing a scalable framework for multimodal data fusion in dental caries prediction, informing precision prevention strategies, and offering methodological guidance for similar computational phenotyping in oral health. The main conclusion will propose that predictive caries risk modelling based on integrated patient data is feasible, generalizable, and adds substantial predictive value over conventional risk assessments, with demonstrated readiness for pilot implementation in routine care. Recommendations include integrating the framework into dental education and continuing professional development, conducting longitudinal implementation trials to assess impact on caries incidence, and establishing governance policies for data sharing, privacy, and ethical use of genomic information in clinical decision-making.