A prediction model for childhood obesity risk based on maternal thyroid status and related parameters using machine learning: a mother-newborn-offspring study in a mild-to-moderate iodine deficiency area

Abstract

Background: Childhood obesity and iodine deficiency are prevalent in developed countries and are linked to adverse health outcomes in adulthood. Mild-to-moderate iodine deficiency and insufficient maternal iodine intake during pregnancy may increase the risk of large-for-gestational-age newborns, which are associated with childhood obesity. Despite this, predicting childhood obesity during pregnancy remains a challenge. We assessed and evaluated machine learning algorithms predicting childhood obesity risk using maternal anthropometrics, thyroid function and iodine intake; and identified key prenatal factors contributing to childhood obesity.

Methods: A diagnostic accuracy study was conducted based on 87 parameters collected from a mother-newborn-offspring prospective cohort (N = 191) in a mild-to-moderate iodine deficiency region. Maternal iodine status and thyroid function, including serum free tri-iodo-thyronine (FT3) concentrations, were assessed during the second half of pregnancy. Iodine intake was evaluated using a semi-quantitative food frequency questionnaire. Anthropometric measurements were obtained from mothers during pregnancy, from newborns at birth, and from children at 2 years of age. An outcome of overweight at 2 years was defined as a gender-adjusted weight percentile >85%. The dataset was split into training (80%) and test (20%) sets. Synthetic datasets were created to evaluate the performance of six machine learning models, including artificial neural networks (Nnet) that trained and evaluated the model using 5-fold cross-validation.

Results: The best-performing model was Nnet, which achieved the highest accuracy (1500 instances with a balanced predicted outcome). On the unseen test data, accuracy, Kappa, outcome F1-score and weighted F1 were 0.743, 0.347, 0.500 and 0.769 (respectively). Significant predictors included gravidity, maternal-newborn anthropometrics (height and head circumference, respectively), maternal consumption and dietary intake of iodine-rich foods (popsicle, selected fish, and yogurt) and FT3.

Conclusions: Machine learning approaches show promise in predicting childhood obesity risk using maternal and dietary factors during pregnancy. If validated, these findings could support interventions to reduce childhood obesity rates.