SARS-CoV-2 infection by trimester of pregnancy and adverse perinatal outcomes: a Mexican retrospective cohort study

Study population

The study employed a retrospective birth cohort from Mexico between January 2020 and November 2021. The Mexican Institute of Social Security (IMSS—Instituto Mexicano del Seguro Social) created an IMSS COVID-19 dataset that was used for this analysis. IMSS, a comprehensive and vertically integrated insurance and healthcare system inspired by the Bismarck Model, provides services to approximately 70 million individuals in Mexico.38 The system’s primary focus is on serving formal private sector employees and their dependents, which constitutes a significant proportion of the Mexican population. Students 18 years or older and self-employed individuals are also eligible for the IMSS services. During the COVID-19 pandemic, IMSS extended its services to non-beneficiaries as part of its response to the public health crisis.

For the study period, the IMSS social security dataset containing COVID-19 information was cross-referenced with the IMSS routine hospitalisation dataset, using national identifications numbers, to identify individuals who underwent SARS-CoV-2 testing while pregnant and subsequently delivered. The datasets comprised maternal age, indigenous status, marital status, smoking status, socioeconomic status, mode of delivery, COVID-19 symptoms, vaccinations and pre-existing conditions (eg, obesity, diabetes, hypertension, cardiovascular diseases, asthma, pneumonia and immunosuppressive disorders). Neonatal information comprised sex, gestational age at birth, birth weight and intrapartum complications. Additionally, the hospitalisation dataset provided information on hospitalisation, intubation, intensive care unit (ICU) admission and death.

The IMSS COVID-19 and routine hospitalisation datasets contained 2 996 438 and 3 473 439 records, respectively. Of these, 825 822 were suspected of COVID-19 and had a hospitalisation record during the study period (online supplemental figure 1). Among those, 399 951 hospitalisations were for deliveries, of which 48 079 were suspected of COVID-19 and 33 719 were tested. Of those tested, 33 325 were singletons, 393 were multiple gestations and 1 was missing. The overall sample for this analysis comprised 17 340 singleton pregnancies with complete covariate information and COVID-19 tests conducted during pregnancy (the remainder were tested either before or after the pregnancy (online supplemental figure 1). 33 pregnancies (0.002%) were tested more than once. Only singleton births were included to eliminate confounding by multiple gestations. We filtered COVID-19 cases during pregnancy using the test date, date of delivery and gestational age at birth. A woman was considered to be positive if she tested positive on either an RT-PCR or a rapid antigen test.39 Mild COVID-19 disease was defined as a positive test with one or more of the mild symptoms (headache, pain swallowing, myalgia, arthralgia, rhinitis, fever or chill, nasal congestion, difficulty speaking, abdominal pain, conjunctivitis, dyspnoea, diarrhoea, chest pain, fast breathing, coryza, loss of smell or taste40). Severe COVID-19 disease was defined as a positive test with the presence of one or more of the following conditions: hospitalisation, intubation, ICU admission or death. Asymptomatic COVID-19 disease was defined as a positive test without mild symptoms or severe disease.

Outcomes

WHO definitions were used for PTB (birth at gestational age less than 37 completed weeks41 and LBW (weight at birth below 2500 g42). Birth weight-for-gestational age z-scores were calculated using the Intergrowth-21 standards.43 SGA and LGA were defined as birth weight-for-gestational age below 10th and above 90th percentiles, respectively.44

We used International Classification of Diseases, Tenth Revision (ICD-10) codes to identify stillbirth (P95), fetal death (O36.4), missed abortion (O02.1) and spontaneous abortion (O03) because IMSS originally used ICD-10. The vast majority (98.1%) of these were coded as fetal death, and we analysed it as one composite outcome.

Statistical analysis

We used the targeted maximum likelihood estimators (TMLE)45 to quantify the associations, which has been shown to perform better in estimating associations and CIs while achieving optimal bias-variance trade-off in the event of model misspecification.46 TMLE obtains the initial estimate and performs the updating step through the clever covariate using the estimated probabilities.47 TMLE uses a semiparametric substitution estimator that is efficient, doubly robust and has advantages over other well-known semiparametric approaches such as the augmented inverse propensity weighted estimator.48 TMLE approach does not assign a parametric form to the distribution of the data, rather it remains an element of the semiparametric model. The model is represented as ψ(p₀)=E_W,0 [E₀(Y|X=1, W)/E₀(Y|X=0, W)], where ψ(p₀) is the parameter of interest (eg, risk ratio, RR) and is a function of the unknown probability distribution. Y represents an adverse perinatal outcome, X represents maternal COVID-19 (diseased=1, disease free=0), W represents a set of covariates, E₀ is the expected probability and E_W,0 is the average of all the observed covariates. A TMLE model first generates an initial estimate of the parameter using super learner algorithms and identifies the best model that minimises the squared error loss function. The second step refines the initial model by optimising the bias-variance trade-off. This procedure has been demonstrated to be robust because it reduces bias and improves efficiency.49 TMLE has been shown to generate unbiased estimate when either the outcome or the exposure is inconsistently estimated.50 This is a significant advantage of the TMLE model, as it is expected to provide an unbiased estimate, in the presence of exposure misclassification. The estimate from the TMLE was interpreted as an RR of an outcome if all participants in the study population were exposed versus all were unexposed. The models were adjusted for available and necessary potential confounders, as described in the directed acyclic graph (online supplemental figure 2).

We implemented an ensemble machine learning method, the super learner, to avoid imposing assumptions on the data. The super learner considers multiple machine learning models by evaluating the specified loss function in cross validation, then it leverages the most predictive model(s) by creating the best-weighted combination.51 We considered both parametric and non-parametric candidate models to cover a wide model space. The candidates were ridge regression, least absolute shrinkage and selection operator (LASSO) regression, multivariate adaptive regression splines, random forest, extreme gradient boosting, generalised additive model and Bayesian generalised linear model.48 We fit the TMLE models in the: (1) overall sample, (2) those with mild symptoms only and (3) those with severe conditions only. We also fit the TMLE models by trimester when COVID-19 was diagnosed, using the overall sample.

In a sensitivity analysis, we adjusted the TMLE models using area-level characteristics at the level of AGEB, equivalent of a census tract in Mexico (https://www.inegi.org.mx/programas/ccpv/2020/%23Microdatos). We used data pertaining to the census tract in which a primary care facility was located. There were many characteristics, including proportions of the AGEB population from different religions, ethnicities, literacy levels, employment, residential ownership, health service affiliation, etc (online supplemental table 1). We used principal component analysis and selected the top three components that together described 87% of the total variation in the area level variables (online supplemental table 2). Early in the pandemic, the testing regimen in Mexico for SARS-CoV-2 was not universal, rather based on symptoms or potential exposure.52 Thus, some asymptomatic pregnancies would not have been tested. In a second sensitivity analysis, we examined if lack of information on non-tested pregnancies could induce bias. To examine the potential bias that would have been induced by the unavailability of test status of all pregnancies, we applied predictive value weighting (PVW) recommended by Lyles and Lin.53 As the PVW method has not been rigorously examined in the TMLE environment, we fit multilevel logistic regression following the original method. We assumed several scenarios of non-tested pregnancies (between 0% and 30%) using non-differential and differential distributional assumptions by COVID-19 status. The upper level of 30% was based on a recent meta-analysis.54 In the non-differential distribution, we assumed an equal proportion of non-tested pregnancies were missing in the negative and positive groups. In the differential assumption, we assumed unequal proportions are missing in the negative and positive groups. We re-estimated the expected true proportion of COVID-19 and the corresponding sensitivity and specificity for each scenario, ensuring that the observed proportion (π) is smaller than the sensitivity and greater than 1–specificity (a conditional requirement). Next, we replicated the data for several combinational levels of outcome, exposure and covariates with one set hypothetically assigned as exposed and another as unexposed. Finally, we fit the models with this new dataset to obtain new estimates that accounted for missing non-tested pregnancies. All hypothesis tests were two sided at the 5% significance level. The analysis was conducted in STATA V.17.0 and R.

Patient and public involvement

Procedures and methods followed the IMSS research guidelines. Anonymised data were used for analysis. Research questions and outcome measures were chosen in consultation, to answer questions of key concern to study partners to understand the potential effect of COVID-19 on pregnancy and the newborn. Patients were not involved in the design or conduct of the study, other than being a participant. Study results were shared with the in-country teams. However, results were not disseminated directly to women.