Interpretable machine learning modeling of treatment outcomes for silver and fluoride therapy in early childhood caries

Abstract

A Commentary on

Wu Y, Jia M, Fang Y, Duangthip D, Chu C H, Gao S S.

Use machine learning to predict the treatment outcome of early childhood caries. BMC Oral Health 2025; 25: 389.

Design

This was a retrospective secondary analysis of data from a randomized controlled trial published in 2020. No new participants or patient involvement were included.

Case selection

This study used the available data from 880 children, contributing 4157 carious tooth surfaces assessed at 30 months. These data were drawn from an original community-based RCT in Hong Kong (baseline N = 1070) that evaluated fluoride and silver interventions for arresting ECC.

Data analysis

Six algorithms were used and compared, including logistic regression (LR), naive Bayes (NB), support-vector machine (SVM), decision tree (DT), random forest (RF), and XGBoost. A surface-level 70:30 train: test split was used. SMOTE was applied to address class imbalance. Models were tuned and evaluated using 1000 bootstrap resamples. Performance metrics were reported as accuracy, recall, precision, F1 score, AUROC, and Brier score. Model interpretability was explored through SHAP analysis to rank variable importance and visualize feature effects.

Results

All models achieved acceptable discrimination. Metrics ranged as follows: accuracy 0.674–0.740; recall 0.731–0.809; precision 0.762–0.802; F1 0.741–0.804; AUROC 0.771–0.859 (RF and XGBoost ≈ 0.86); Brier 0.134–0.227. SHAP interpretation indicated that tooth and surface location of caries, newly developed dmfs, baseline dmfs, caregiver-assisted brushing, and visible plaque index contributed most to the model’s predictions, while the fluoride and silver intervention ranked mid-to-low in importance. The original study, however, highlighted tooth and surface location of caries, newly developed dmfs, and daily snack intake as key effect modifiers influencing caries-arrest outcomes.

Conclusions

Machine learning models, specifically ensemble tree-based methods, can discriminate which surfaces arrest by 30 months using multidomain inputs spanning clinical, behavioral, and socioeconomic features. However, participant-level splitting was not performed, allowing potential clustering leakage. External validation and detailed calibration analyses were also lacking. These methodological constraints likely inflated model performance and currently limit its generalizability and readiness for clinical implementation.