Construction of a dental caries prediction model by data mining

Area of Science:

• Oral Health Research
• Biostatistics
• Predictive Modeling

Background:

• Dental caries distribution is often skewed, hindering precise prediction.
• Conventional models struggle with imbalanced datasets, impacting accuracy.
• Accurate prediction models are vital for effective caries prevention strategies.

Purpose of the Study:

• To evaluate a balancing technique for improving dental caries prediction models.
• To compare the accuracy of models built with the balancing technique against conventional methods.
• To assess the robustness of prediction models using skewed data.

Main Methods:

• Utilized three-year follow-up data from 500 Japanese schoolchildren (ages 5-8).
• Included variables: salivary mutans streptococci/lactobacilli, saliva volume, pH, fluoride use, and sugar intake.
• Constructed models using logistic regression, neural networks, and decision analysis, with and without a balancing technique involving random sampling of subjects with/without new caries.

Main Results:

• The balancing technique produced the most robust model.
• C 5.0 analysis with the balancing technique achieved 0.73 sensitivity and 0.77 specificity.
• This approach demonstrated superior performance compared to conventional models on skewed data.

Conclusions:

• The balancing technique is a valuable method for developing suitable and robust dental caries prediction models.
• This technique addresses challenges posed by skewed data distributions in epidemiological studies.
• Improved prediction accuracy can lead to more targeted and effective public health interventions for dental caries.