Machine Learning versus Traditional Regression for Early Childhood Caries Risk Factor Identification: A Methodological Comparison Study

Zenodo repository description: Title: Machine Learning versus Traditional Regression for Early Childhood Caries Risk Factor Identification: A Methodological Comparison Study Description: This preprint compares traditional regression and machine learning (ML) approaches for identifying risk factors associated with early childhood caries (ECC) in Latvian preschool children. We analysed data from 209 children aged 2 to 5 years (102 with ECC, 107 caries-free) collected in Latvia between 2020 and 2021. Traditional models included logistic regression and spline regression. ML models included random forest, XGBoost, elastic net, and k-nearest neighbours. Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC). Logistic regression achieved an AUC of 0.92, equivalent to or higher than all ML models tested. Random forest achieved 0.92, XGBoost 0.91, elastic net 0.89, and KNN 0.86. Logistic regression identified visible plaque (OR = 9.14), infrequent parental brushing (OR = 5.13), and reduced toothbrushing frequency (OR = 3.49) as the strongest risk indicators. No ML model showed advantage over logistic regression. ML did not outperform traditional regression for identifying ECC risk factors in this sample. Traditional regression methods remain valid and interpretable for ECC risk modelling in dental public health research. Keywords: early childhood caries, machine learning, logistic regression, risk factors, predictive modelling, oral epidemiology, Latvia Related project: EpiDentLatvia: Mapping the Epidemiological Profile of Oral Health in Latvian Children and Adolescents Funding: Scientist Grant "EpiDentLatvia" (RSU-ZG-2024/1-0044), project "RSU Internal and RSU with LASE External Consolidation" (No. 5.2.1.1.i.0/2/24/I/CFLA/005), funded by the European Union Recovery and Resilience Facility and the Republic of Latvia. Dataset: https://doi.org/10.48510/FK2/PZL9DW