Research in China identifies age, parental BMI, screen time and birthweight as key predictors of childhood obesity, while highlighting how risk assessments change under Chinese and WHO growth standards.
Childhood obesity has become a major global health problem, with the World Health Organization warning that rates have continued to climb in both adults and children. WHO said that in 2024 about 400 million children and adolescents aged 5 to 19 were overweight, including 170 million living with obesity, while an estimated 35 million children under five were also overweight. A recent meta-analysis published in JAMA Pediatrics put the global prevalence of obesity in children and adolescents at 8.5%, with striking variation between countries, from 0.4% in Vanuatu to 28.4% in Puerto Rico. Against that backdrop, new research from China offers a more detailed look at the mix of family, behavioural and early-life factors that may help explain why some children are at higher risk than others.
The study, published in the Journal of Global Health, examined obesity among children and adolescents aged 3 to 18 years in Beijing and tested whether machine learning could sharpen risk prediction under both Chinese and WHO growth standards. The researchers worked with data from two regional surveys in Pinggu district and then checked their findings against an independent cohort from Tangshan, Hebei province. They say the project is the first to build and validate predictive models for paediatric obesity using both standards, an important point because different diagnostic systems can classify the same child differently, especially near the threshold for obesity.
From an initial pool of 22,555 participants, the researchers excluded children in the overweight category and focused on 18,936 children and adolescents under Chinese criteria and 18,517 under WHO criteria. They examined 37 features spanning demographics, early-life history, daily habits, health conditions and family background. After feature selection, the final model under Chinese standards kept five variables: age, maternal body mass index, paternal body mass index, screen time and birthweight. Under WHO standards, 12 variables remained, including those five along with paternal age at conception, gestational age, outdoor exercise, breastfeeding duration and complementary feeding time.
The best-performing model for Chinese standards was a neural network, while XGBoost performed best under WHO standards. Both delivered similar discrimination, with area under the receiver operating characteristic curve scores of about 0.73. That level suggests the models were better than chance but not strong enough to be used as stand-alone screening tools. The researchers also found that the models were highly specific but less sensitive, meaning they were better at ruling out children without obesity than catching every child who had it. In practical terms, that makes them more suitable as decision-support tools than as broad population screening devices.
Age was the most influential factor in both sets of results, followed by parental body mass index, screen time and birthweight. The study also pointed to different patterns depending on which diagnostic framework was used. Under Chinese standards, maternal and paternal body mass index carried the greatest weight, while under WHO standards, gestational age, breastfeeding duration, paternal age at conception and other early-life factors also contributed meaningfully. The team used SHAP analysis, a method for explaining how individual variables influence model output, to make the findings easier to interpret for clinicians and families.
The researchers argue that the results reinforce a familiar but often difficult message: childhood obesity is rarely driven by a single cause. Rather, it emerges from the interaction of inherited risk, early feeding patterns, parental characteristics and day-to-day behaviour. They also note that a child’s risk profile may look different depending on whether Chinese or WHO growth references are used, which could matter in clinical practice and public health planning. To help translate the work into practice, the team has also developed an online tool that estimates obesity risk and generates visual explanations of the contributing factors.
The authors are careful to note the limits of their work. The study is cross-sectional, which means it cannot prove cause and effect, and many inputs relied on parent-reported questionnaires. Even so, the researchers say the models performed well enough to support further refinement, and the external validation suggests they may be transportable beyond Beijing. For parents and caregivers, the findings underscore a simple point: early habits, family health and routine behaviour can all shape a child’s long-term risk, and the earlier those factors are identified, the better the chance of preventing obesity before it becomes entrenched.
Disclaimer: This content is for informational purposes only and is not intended to be a substitute for professional medical judgment, advice, diagnosis, or treatment.
