An Automated Machine Learning Classifier for Early Childhood Caries

Area of Science:

• Pediatric Dentistry
• Machine Learning in Healthcare
• Public Health

Background:

• Early childhood caries (ECC) poses a significant public health challenge.
• Accurate and efficient screening tools are crucial for early intervention.
• Existing methods for ECC assessment can be resource-intensive.

Purpose of the Study:

• To develop and evaluate an automated machine learning (AutoML) algorithm for classifying early childhood caries (ECC) status in children.
• To assess the performance of different predictor models for ECC classification.
• To determine the utility of AutoML in pediatric oral health screening.

Main Methods:

• Utilized data from 6,404 children aged 3-5 years in North Carolina.
• Employed an AutoML approach on Google Cloud for ECC classification.
• Evaluated ten sets of ECC predictors for accuracy using AUC, Se, and PPV.
• Validated models internally and externally using the NHANES dataset.

Main Results:

• A parsimonious model with children's age and parent-reported oral health status achieved the highest accuracy (AUC=0.74, Se=0.67, PPV=0.64).
• This model showed comparable performance on an external NHANES dataset (AUC=0.80, Se=0.73, PPV=0.49).
• A comprehensive model with 12 variables performed worse (AUC=0.66, Se=0.54, PPV=0.61).

Conclusions:

• Simple, parsimonious AutoML classifiers, including single-item self-reports, are valuable for ECC screening.
• These algorithms can be enhanced with biological information for improved future performance.
• AutoML offers a promising avenue for efficient and accurate pediatric caries assessment.