BACKGROUND:
Early childhood overweight and obesity remain pressing public health concerns, with increasing global prevalence and established links to adverse health trajectories. The early identification of at-risk-children through predictive modelling offers opportunities for targeted intervention. However, models specifically calibrated for children under three years are scarce and have not been developed in a Belgian context.

OBJECTIVES:
To develop and validate a prediction model for overweight and obesity at 24 months using routinely collected population-level data from perinatal and early childhood registries in Flanders, Belgium.

METHODS:
Data from 340,502 mother-child dyads (2009–2018) were derived from two routinely collected datasets: the Study Centre for Perinatal Epidemiology and Opgroeien (formerly Kind&Gezin). A prediction model using logistic regression was trained on 70% of 2009–2016 data (n≈195,000) and tested on the remaining 30%, with additional validation using 2017–2018 birth cohorts. Predictors included maternal, perinatal, sociodemographic, and early growth characteristics. Model performance was evaluated using AUC, sensitivity, specificity, PPV, NPV, diagnostic odds ratio, and F1-score.

RESULTS:
The base model, incorporating only predictors available at birth and infancy, achieved moderate discriminatory capacity (AUC:0.670–0.675). Key predictors included infant birth weight, language spoken at home, and household urbanicity. Inclusion of infant BMI at 12 months substantially improved model performance (AUC:0.861–0.883). At an optimal cutoff (0.08), the final model achieved a sensitivity of 78.0%, specificity of 81.9%, PPV of 24.1%, NPV of 98.1%, and an F1-score of 18.4%.

CONCLUSIONS:
A predictive model for overweight/obesity at age 24 months was successfully developed using linked Belgian registry data. While the model demonstrated strong discriminatory performance, predictive accuracy was limited, particularly regarding PPV. These findings underscore the importance of incorporating broader behavioural and environmental data in future modelling efforts to enhance clinical relevance and utility.

KEY MESSAGE:
Predictive models based on early-life registry data show promise but must evolve to incorporate contextual psychosocial, behavioural, and environmental factors. Poster session 4 (Group B)