Assessing Screening Methods and Machine Learning for Predicting Childhood Overweight and Obesity: A Population-Based Study

Background: This study aimed to assess the effectiveness of different screening methods, including Large for Gestational Age (LGA), macrosomia, and the WHO weight-for-length growth charts, in predicting childhood overweight and obesity. Additionally, we sought to develop a machine learning model utilizing various demographic, birth-related, maternal, and child growth data to evaluate its potential for improving predictive accuracy. Methods: We conducted this study using data from the Tipat Halav Israeli Screening (THIS) program, covering approximately 70% of Israeli children. The study included all children born between January 2014 and June 2016, with a minimum follow-up of 18 months, while excluding preterm births, multiple pregnancies, and those lacking growth measurements. Childhood overweight was defined based on WHO recommendations, and we evaluated the performance of existing models (LGA, macrosomia, and WHO percentiles) using different cutoffs. We also developed a machine learning model employing Random Forest and XGBoost algorithms. Results: The evaluation of prediction models yielded modest Area Under the Curve (AUC) values, ranging from 0.588 to 0.653. However, these models displayed significant improvement over random sampling. Notably, children selected by these models for post-birth intervention demonstrated up to a 17% likelihood of eventually becoming overweight, with 50% considered at risk. Conclusions: Our study underscores the importance of early intervention in addressing childhood overweight and obesity. Screening models, particularly LGA and macrosomia, exhibit promise in identifying newborns at risk. Although our machine learning model did not substantially enhance prediction, future research should explore the incorporation of additional relevant parameters to refine predictive accuracy and facilitate more effective early intervention efforts.