Scholarlink

Objectives <p>This study integrated cone-beam computed tomography(CBCT) imaging, cephalometric measurements, and clinical characteristics to construct a risk prediction model for proximal caries.</p> Methods <p>Clinical and radiographic data were collected from 140 orthodontic patients. Three machine learning algorithms (least absolute shrinkage and selection operator(LASSO) regression, random forest, and support vector machine recursive feature elimination(SVM-RFE) were employed for variable selection. The variables that were jointly identified by all three algorithms were extracted via Venn diagram intersection analysis. The final selected variables were entered into a multivariable logistic regression model. Overfitting was formally assessed using 5-fold stratified cross-validation with continuous predicted probabilities, quantified by the training–CV AUC gap. Multicollinearity was diagnosed using the variance inflation factor(VIF), and backward elimination guided by AIC and Occam’s razor was performed to identify the most parsimonious model. Model discrimination was evaluated by the area under the ROC curve(AUC), calibration by the Hosmer–Lemeshow test, and clinical utility by decision curve analysis(DCA).</p> Results <p>Four variables(ANB angle, SNB angle, APDI, and age) were initially identified as shared predictors across all three machine learning algorithms. The full four-variable logistic regression model. Among the four candidate models, the full four-variable logistic regression demonstrated the smallest overfitting gap(0.052), which narrowed to 0.026 in the final two-variable model. Backward elimination yielded a final two-variable model comprising age and SNB angle. Both predictors were statistically significant(age: OR = 1.577, 95% CI = 1.098–2.265, <i>P</i> = 0.014; SNB angle: OR = 0.663, 95% CI = 0.458–0.960, <i>P</i> = 0.030).</p> Conclusion <p>A parsimonious two-variable clinical model integrating age and SNB angle was developed for proximal caries risk prediction. The consensus-based variable selection strategy and formal overfitting assessment framework enhance the generalizability and clinical applicability of the model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Construction of a clinical model for proximal caries via machine learning algorithms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月25日

查看全文
  • Juan Liu,
  • Yilin Wang,
  • Xiaoli Hu,
  • Mo Xiao,
  • Zhigang Zuo,
  • Yuanyuan Wang,
  • Bin Yin,
  • Ligeng Wu

摘要

Objectives

This study integrated cone-beam computed tomography(CBCT) imaging, cephalometric measurements, and clinical characteristics to construct a risk prediction model for proximal caries.

Methods

Clinical and radiographic data were collected from 140 orthodontic patients. Three machine learning algorithms (least absolute shrinkage and selection operator(LASSO) regression, random forest, and support vector machine recursive feature elimination(SVM-RFE) were employed for variable selection. The variables that were jointly identified by all three algorithms were extracted via Venn diagram intersection analysis. The final selected variables were entered into a multivariable logistic regression model. Overfitting was formally assessed using 5-fold stratified cross-validation with continuous predicted probabilities, quantified by the training–CV AUC gap. Multicollinearity was diagnosed using the variance inflation factor(VIF), and backward elimination guided by AIC and Occam’s razor was performed to identify the most parsimonious model. Model discrimination was evaluated by the area under the ROC curve(AUC), calibration by the Hosmer–Lemeshow test, and clinical utility by decision curve analysis(DCA).

Results

Four variables(ANB angle, SNB angle, APDI, and age) were initially identified as shared predictors across all three machine learning algorithms. The full four-variable logistic regression model. Among the four candidate models, the full four-variable logistic regression demonstrated the smallest overfitting gap(0.052), which narrowed to 0.026 in the final two-variable model. Backward elimination yielded a final two-variable model comprising age and SNB angle. Both predictors were statistically significant(age: OR = 1.577, 95% CI = 1.098–2.265, P = 0.014; SNB angle: OR = 0.663, 95% CI = 0.458–0.960, P = 0.030).

Conclusion

A parsimonious two-variable clinical model integrating age and SNB angle was developed for proximal caries risk prediction. The consensus-based variable selection strategy and formal overfitting assessment framework enhance the generalizability and clinical applicability of the model.