Development of second and third-trimester population-specific machine learning pregnancy dating model (Garbhini-GA2) derived from the GARBH-Ini cohort in north India

Abstract

AbstractBackgroundThe prevalence of preterm birth (PTB) is high in lower and middle-income countries (LMIC) such as India. In LMIC, since a large proportion seeks antenatal care for the first time beyond 14-weeks of pregnancy, accurate estimation of gestational age (GA) using measures derived from ultrasonography scans in the second and third trimesters is of paramount importance. Different models have been developed globally to estimate GA, and currently, LMIC uses Hadlock’s formula derived from data based on a North American cohort. This study aimed to develop a population-specific model using data from GARBH-Ini, a multidimensional and ongoing pregnancy cohort established in a district hospital in North India for studying PTB.MethodsData obtained by longitudinal ultrasonography across all trimesters of pregnancy was used to develop and validate GA models for second and third trimesters. The first trimester GA estimated by ultrasonography was considered the Gold Standard. The second and third trimester GA model named, Garbhini-GA2 is a multivariate random forest model using five ultrasonographic parameters routinely measured during this period. Garbhini-GA2 model was compared to Hadlock and INTERGROWTH-21st models in the TEST set by estimating root-mean-squared error, bias and PTB rate.FindingsGarbhini-GA2 reduced the GA estimation error by 23-45% compared to the published models. Furthermore, the PTB rate estimated using Garbhini-GA2 was more accurate when compared to published formulae that overestimated the rate by 1·5-2·0 times.InterpretationThe Garbhini-GA2 model developed is the first of its kind developed solely using Indian population data. The higher accuracy of GA estimation by Garbhini-GA2 emphasises the need to apply population-specific GA formulae to improve antenatal care and better PTB rate estimates.FundingCentre for Integrative Biology and Systems Medicine, IIT Madras; Department of Biotechnology, Government of India; Grand Challenges India, BIRAC.Panel: Research in ContextEvidence before this studyThe appropriate delivery of antenatal care and accurate delivery date estimation is heavily dependent on accurate pregnancy dating. Unlike GA estimation using crown-rump length in the first trimester, dating using foetal biometry during the second and third trimesters is prone to inaccuracies. This is a public health concern, particularly in low and middle-income countries like India, where nearly 40% of pregnant women seek their first antenatal care beyond 14 weeks of gestation. The dating formulae used in LMIC were developed using foetal biometry data from the Caucasian population, and these formulae are prone to be erroneous when used in ethnically different populations.Added value of this studyThis study developed a dating model, the Garbhini-GA2 model for second and third trimesters of pregnancy using multiple candidate biometric predictors measured in a North Indian population. When evaluated internally, this model outperformed the currently used dating models by reducing the errors in the estimation of gestational age by 25-40%. Further, Garbhini-GA2 estimated a PTB rate similar to that estimated by the Gold Standard in our population, while the published formulae overestimated the PTB rates.Implications of all the available evidenceOur Garbhini-GA2 model, after due validations in independent cohorts across the Southeast Asian regions, has the potential to be quickly translated for clinical use across the region. A precise dating will benefit obstetricians and neonatologists to plan antenatal and neonatal care more exactly. From an epidemiologist standpoint, using the Garbhini-GA2 dating formulae will improve the precision of the estimates of pregnancy outcomes that heavily depend on gestational age, such as preterm birth, small for gestational age and stillbirth in our population. Additionally, our dating models will improve phenotyping by reducing the risk of misclassification between outcomes for mechanistic and biomarker research.