Data
The dataset was derived from pregnancies managed in the UK national GAP programme which was running during the study period in a total of 132 NHS Trusts, Health Boards or Health and Social Care Trusts in England, Wales, Scotland and Northern Ireland. Information was entered into the GROW App software (version 1.2.6.1, www.gestation.net) which produces customised growth charts and birthweight centiles adjusted for each pregnancy according to early pregnancy maternal weight, height, parity and ethnic origin 12. Also recorded was whether the pregnancy was considered at increased risk of growth restriction on the basis of obstetric or medical history and risk factors such as smoking or high body mass index, and whether serial third trimester ultrasound biometry was instituted according to protocol. The number of growth scans was not collected in the database but, according to routine audits within the programme, usually consisted of 3 to 4 third trimester scans from 28 weeks to delivery. Pregnancies were dated according to the routine 1st trimester ultrasound scan. Outcomes recorded after delivery included birthweight, sex and gestational age at delivery, and whether live birth or stillbirth. Stillbirth was defined as a baby born with no signs of life from 24.0 weeks gestation. Gestational age for stillbirths was based on the age at delivery minus 2 days to adjust for the average delay in third trimester between intrauterine demise and delivery 19.
There were 1,641,897 singleton pregnancies delivered> 24.0 weeks gestation between January 2015 and January 2020. After excluding late fetal losses (725) and cases with missing or incomplete data (68,355), the final study cohort consisted of 1,572,817 pregnancies. All data were fully anonymised including institution of origin, date of birth and all other maternal, new-born and pregnancy-related identifiers, and hence ethics approval was not required.
Statistical Methods To assess the association between customised GROW centiles and stillbirth, we developed receiver operator curves (ROCs) according to standard methodology20 on the overall cohort as well as the low and high risk subgroups. Risk status was recorded in 1,308,967 (83%) of the cases, with the remainder missing mostly because the detailed data items were not contained in the version of the local electronic system in use during the study period.
Analyses were carried out using SAS software (version 9.4, SAS Institute, NC, USA). Optimal centile cut-offs were determined by balancing sensitivity and specificity using Youden’s index21, and model performance was assessed according to area under the curve (AUC) analysis. Differences between pregnancy characteristics were determined using the Kruskal Wallis test for continuous variables and a Z Test between proportions. Outcome between groups was compared using relative risk (RR) with 95% confidence interval (CI).