Data source and study population:
This study included singleton births from a prospective observational study in 16 hospitals in Benin, Malawi, Tanzania and Uganda collected as part of the Action Leveraging Evidence to Reduce perinatal morTality and morbidity (ALERT) study.45 Benin, Malawi, Tanzania and Uganda are low- and lower middle-income countries, facing a large perinatal mortality burden.1 In each of the countries, four medium-size hospitals were included with more than 2500 births per annum. Hospitals were typically district or regional public and private-for-nonprofit (faith-based) facilities, although the national referral hospital in Cotonou, Benin also took part in the study. In this analysis, we included all mother-baby pairs admitted for childbirth in any of the hospitals between 1st July 2021 and 30th September 2023. We excluded mother-baby pairs who were referred to the hospitals after giving birth. Data was entered daily into a pre-programed harmonized perinatal e-registry in the maternity ward of each hospital using the Research Electronic Data Capture (REDCap) platform available on tablets; details are published elsewhere.46 Observations with missing outcome (N=81) and exposure (N=211) data were removed from the dataset (Supplementary Fig. S1). The ALERT study was approved by the Swedish Ethical Authority and in each participating country with details provided in our protocol paper.45
Outcome assessment: Stillbirths and perinatal deaths
Our main outcomes were stillbirths (including both ante- and intrapartum deaths) and very early neonatal deaths documented in REDCap by nurse-midwives in each hospital.46 Stillbirths were defined as the death of a foetus before birth, weighing at least 1000 grams, and which could not be resuscitated after birth. Perinatal deaths included both stillbirths and deaths within the first 24 hours (very early neonatal deaths). We chose 24 hours as a cut-off as mothers were usually discharged the day after birth. Data collectors and nursing staff in the study hospitals were trained to distinguish antepartum and intrapartum stillbirth based on skin appearance, rather than in relation to onset of labor. Since skin disintegration is considered to begin 12 hours after death, macerated stillbirth is assumed to have occurred before the onset of labor and fresh stillbirth after and these were used as a proxy for antepartum and intrapartum stillbirth, respectively.
Other obstetric variables for stratified analysis
For the stratified analysis we considered obstetric variables which were collected as part of our perinatal e-registry. Information was abstracted by nurses or midwives from antenatal care cards (e.g. maternal age and HIV status), retrieved from admission files (e.g. complications such as hypertensive disorders and antepartum bleeding) or clinical files (e.g, babies’ sex, birth weight) and entered into the electronic REDCap perinatal e-registry form.46
Exposure assessment: Temperature
We obtained daily mean temperatures from the European Centre for Medium-Range Weather Forecasts (ECMWF) at a 28 km × 28 km resolution (0.25 ° × 0.25 °). Data was freely downloaded from the Copernicus Climate Data store (https://cds.climate.copernicus.eu/). We linked the daily mean temperatures at the specific grid of ECMWF within which each hospital was located. We further extracted 10 hourly data from ECMWG (12 pm, 1 am, 2 am, 3 am, 4 am, 12 am, 1 pm, 2 pm, 3 pm, 4 pm) to estimate the mean, maximum and minimum values. In addition, we included daily dew point concentrations obtained from ECMWF as a proxy of relative humidity.
Statistical analysis
We estimated the association between multiple short-term air temperature metrics and perinatal mortality using a case-crossover design. This design has been widely utilized in short-term studies, including for the association between temperature and perinatal mortality.47 The major advantage of the case-crossover design is that each case serves as its own control, allowing for precise adjustment of individual-level confounders that remain constant or change slowly over time, such as age, and obstetric risk factors.47 In our study, we adopted a time-stratified strategy to select control days for each case. This strategy involves defining control days as the same day-of-the-week within the same month as childbirth. By employing this approach, we can effectively control for day-of-week and other short-term temporal trends, as the cases and controls are closely matched in time. Additionally, the separation of cases and controls by seven days helps reduce serial autocorrelation in exposures and outcomes.47 To evaluate the association between temperature metrics and perinatal mortality, we first applied country-specific conditional logistic regression to estimate odds ratios (ORs) and 95% confidence intervals (CI) of perinatal mortality, comparing the 99th percentile vs the 75th percentile of the annual temperature distribution. We applied this percentile to accurately reflect the warm season as the estimated minimum mortality temperature (MMT) in some countries was set at very low levels of the temperature distribution. To account for the nonlinear shape of the relationship, we applied a natural cubic spline with three internal knots placed at the 10th, 75th, and 90th percentiles of country-specific temperature distributions and the lag-response curve with a natural cubic spline with three internal knots placed at equally spaced values in the log scale. We considered the 7-day lag period (0-6) as our reference exposures as previous studies have indicated to be the period of highest susceptibility.11,16,35,36 In the second stage, we applied a random-effects meta-analytical model to pool the country-specific estimates of associations of temperature and perinatal deaths. We calculated I2 statistics and Cochran's Q-test to evaluate the between-country heterogeneity. We further applied a multivariate meta-regression of the country specific predictions of the exposure-response curve to obtain a cumulative exposure-response curve. Statistical analyses were performed in R software (version 3.6.2).
To evaluate the robustness of our results, we applied several sensitivity analyses. First, we evaluated the effect of temperature on stillbirths during the six consecutive hottest months for each country in order to capture extreme heat events. Second, we explored different lag patterns including single lag of same day (lag0), 2-day lag (lag 0-1), and a 3-day lag (lag 0-2), to identify different windows of vulnerability. Third, we evaluated the associations of daily maximum and minimum temperature on perinatal mortality. Fourth, we used the 50th percentile as our reference value (in comparison to the 75th value) in order to evaluate a consistent increase in temperature across countries. Finally, we adjusted for dew point residuals as a proxy of humidity.
We further stratified the analyses by some key effect modifiers that could potential modify the associations between short-term exposure of ambient air pollution and still births. The key modifiers included maternal age, HIV status, hypertensive disorders, number of pregnancies, sex, preterm births, low birthweight, labour problems and Antepartum Haemorrhage (APH).