Prediction of childhood overweight and obesity at age 10–11: findings from the Studying Lifecourse Obesity Predictors and the Born in Bradford cohorts

Childhood overweight and obesity are linked to adverse adult health outcomes including diabetes, hypertension, and coronary heart disease. Global childhood obesity has increased markedly since 1975. In England, National Child Measurement Programme (NCMP) data show high and recently fluctuating prevalence at ages 4–5 and 10–11 years, with a pronounced deprivation gradient. Weight status tracks from early to later childhood, with many children maintaining their status between Reception (4–5 years) and Year 6 (10–11 years). Although NCMP provides parental feedback, it has limited impact on lifestyle changes. Targeted interventions are most effective when risk can be identified early. The study aimed to develop and externally validate prediction models for overweight/obesity at age 10–11 using routinely collected BMI at age 4–5 and maternal antenatal/birth data.

Prior analyses of UK cohorts and NCMP tracking indicate substantial persistence of overweight/obesity from ages 5 to 11, with higher risks among children already overweight or obese at school entry. Feedback from NCMP improves parental recognition but rarely leads to behavior change; school-based combined diet and physical activity interventions are most effective. A systematic review identified eight child obesity prediction models (four for ages 6–13), commonly including child sex and maternal BMI, but none developed in the UK used routine data with good performance (one applicable model had AUC 0.64). Recent Scottish models for predicting obesity at 12 years included maternal BMI, child sex, child BMI at 5–6 years, and other non-routinely collected predictors; similar models in the Netherlands and Australia also relied on variables not routinely available in UK healthcare systems.

Design and data sources: Prediction models were developed and internally validated using the SLOPE cohort, an anonymised linked dataset of maternal antenatal/birth records and child health records for singleton births at University Hospital Southampton (2003–2018). External validation used the Born in Bradford (BiB) cohort (children followed up at ages 7–11 between 2017 and 2020). Outcome: Overweight/obesity at 10–11 years defined by BMI ≥ 91st percentile (UK1990 reference), with BMI derived from weight/height measured in schools via NCMP at ages 4–5 (Year R) and 10–11 (Year 6). Sample: SLOPE development sample included 6566 children with valid measurements at both ages; BiB external validation sample included 3325 children. Candidate predictors: Staged model building: (1) Year R predictors—child BMI at 4–5 years (with transformation) and sex; (2) addition of maternal variables from first antenatal booking: maternal age, BMI, smoking status, highest educational attainment, ethnicity, employment status, parity, folic acid intake, first language English, and single parent status; (3) birth variables: birthweight, gestational age, and mode of birth. Variables were routinely collected in clinical care. Statistical analysis: Logistic regression with stepwise backward elimination using p-value threshold 0.157 (AIC-equivalent) to reduce overfitting. Multivariable fractional polynomials explored non-linear transformations for continuous predictors (child BMI at 4–5, maternal age, maternal BMI). Events-per-variable assessed with a rule-of-thumb ≥20 EPV. Internal validation via bootstrapping (1000 repetitions) estimated optimism and produced shrinkage factors, applied to coefficients. External validation used BiB measurements collected by trained researchers; pregnancy predictors in BiB with 18–24% missingness were imputed via multiple imputation by chained equations (25 imputations, 10 iterations). Model performance: Discrimination assessed by AUC; calibration by calibration plots, calibration slope, and calibration-in-the-large (CITL). Risk score calculation: log-odds from regression coefficients converted to predicted probability P = 1/(1 + exp(−Y)). Performance metrics (sensitivity, specificity, PPV, NPV) were computed across multiple risk thresholds, with illustrative thresholds at 15%–40%. Ethics: Approvals from HRA (IRAS 242031) for SLOPE and Bradford REC (07/H1302/112; 16/YH/0320) for BiB.

• Sample and prevalence: SLOPE n=6566; overweight/obesity prevalence 14.6% at 4–5 years and 26.1% at 10–11 years; 10.8% were overweight/obese at both timepoints. Most healthy-weight children at 4–5 remained healthy-weight at 10–11 (80.7%), and 74.3% of those overweight/obese at 4–5 remained so at 10–11. • Predictor selection: Final models retained child BMI at 4–5 years (with transformation) and sex; with the pregnancy-augmented model additionally retaining maternal age, BMI, smoking status, highest educational attainment, employment status, ethnicity, and parity. Birth variables (birthweight, gestational age, mode of birth) were not retained. • Model coefficients (shrunken): Year R-only model: intercept −0.88; BMI at 4–5 (transformed: BMI − 16.2) coefficient 1.00; female sex −0.29. Year R + pregnancy model: intercept −1.69; BMI at 4–5 (BMI − 16.20) 0.97; female −0.33; maternal age (age − 27.9) 0.02; maternal BMI (BMI − 25.2) 0.07; smoking ex 0.19, current 0.50 (ref never); education college 0.39, secondary/below 0.48 (ref university+); employment student 0.54, unemployed 0.06 (ref employed); ethnicity Asian 0.96, Black 0.73, Other 0.77, Mixed −0.05 (ref White); parity 1: 0.05, 2: −0.25, ≥3: −0.02 (ref 0). • Discrimination: AUC development/validation: Year R-only 0.82/0.83; Year R + pregnancy 0.84/0.85. • Calibration: Calibration slopes ~1.00; CITL ~0.00; shrinkage factor 0.98, indicating minimal overfitting. • Risk thresholds (illustrative): At 30% cut-off, Year R-only identifies 31.3% at risk with sensitivity 65.8%, specificity 80.9%, PPV 55.0%, NPV 87.0. With pregnancy factors, 37.7% at risk with sensitivity 71.3%, specificity 74.2%, PPV 49.5%, NPV 87.9. Higher thresholds improve specificity and PPV at the expense of sensitivity. • External validation: Performance in BiB closely matched development performance with good discrimination and calibration.

The study shows that routinely collected data at school entry (age 4–5) can predict overweight/obesity risk at age 10–11 with good accuracy, and that incorporating maternal antenatal data yields small but meaningful improvements in discrimination while maintaining excellent calibration. These models operationalize risk stratification at Year R within existing NCMP workflows, enabling identification of children at elevated risk who may benefit from targeted preventive interventions, especially outside extreme BMI centiles where proactive follow-up is inconsistent. Given socioeconomic disparities in obesity, such tools could support prioritization of resources to families in more deprived contexts. Compared with existing international models, the present models perform better in a UK setting and rely solely on routinely collected variables, increasing implementability. While pregnancy variables slightly improve sensitivity, the Year R-only model remains practical where linked maternal data are unavailable. The models’ high NPV suggests reliable identification of low-risk children; moderate PPV indicates that some identified at-risk children will not progress, underscoring the need to evaluate the population-level impact and cost-effectiveness of risk-guided interventions.

Prediction models using routine measurements at age 4–5 can quantify individual risk of overweight/obesity at age 10–11. Adding maternal pregnancy data modestly improves prediction. These models offer a feasible basis for integrating risk identification into early years care to target preventive support and potentially reduce the prevalence of childhood obesity.

• Missing data constrained inclusion of early-life variables (e.g., breastfeeding 99% missing; weights at 1–2 years 65% missing). • Outcome data were unavailable for many eligible children due to movement, recording differences, or schooling outside state sector, potentially introducing selection bias, although observed prevalence aligned with national estimates. • No definitive literature-based risk threshold; chosen thresholds balance sensitivity/specificity and proportion flagged, which may vary by context. • Moderate PPV (e.g., 55% at 30% cut-off in the Year R-only model) implies some false positives among those flagged as at risk. • External validation was in a single UK cohort (BiB) with specific demographic characteristics; broader validation could further assess generalizability.