An explainable and transparent machine learning approach for predicting dental caries: a cross-national validation...

Area of Science:

• Dentistry
• Artificial Intelligence
• Machine Learning
• Public Health

Background:

• Artificial intelligence (AI) is increasingly used in dentistry, but inadequate validation leads to overestimated machine learning (ML) model performance.
• External validation with independent datasets is essential to confirm ML model generalizability and real-world applicability.

Purpose of the Study:

• To develop and validate Extreme Gradient Boosting (XGBoost) models for dental caries detection using questionnaire data.
• To assess the generalizability of ML models through external validation and enhance interpretability using explainable AI (XAI).

Main Methods:

• Developed XGBoost models using NHANES datasets (n=6070) with nested cross-validation and a holdout test set.
• Externally validated model performance on independent Northern Finland Birth Cohort datasets (NFBC1966 and NFBC1986; n=3616).
• Utilized beeswarm plots for variable importance analysis to improve model interpretability.

Main Results:

• The ML model achieved an AUC of 0.785 internally but showed poor sensitivity (0.391) despite high specificity (0.919).
• External validation revealed a significant performance decline, with AUC dropping to 0.550, sensitivity to 0.053, and specificity slightly increasing to 0.974.
• Key predictors included self-rated oral health, missing teeth, financial status, and time since last dental visit.

Conclusions:

• External validation demonstrated a notable degradation in ML model performance compared to internal validation.
• Explainable AI (XAI) methods show potential for future individualized dental caries risk assessment.