Ambient temperature during pregnancy and fetal growth in Eastern Massachusetts, USA

Abstract

Background

Left unabated, rising temperatures pose an escalating threat to human health. The potential effects of hot temperatures on fetal health have been under-explored. Here, we examined the association between prenatal ambient temperature exposure and fetal growth measures in a Massachusetts-based pregnancy cohort.

Methods

We used ultrasound measurements of biparietal diameter (BPD), head circumference (HC), femur length and abdominal circumference (AC), in addition to birthweight (BW), from 9446 births at Beth Israel Deaconess Medical Center from 2011 to 2016. Ultrasound scans were classified into three distinct gestational periods: 16–23 weeks, 24–31 weeks, 32+ weeks; and z-scores were created for each fetal growth measure using the INTERGROWTH-21st standards. We fitted distributed lag models to estimate the time-varying association between weekly temperature and fetal growth, adjusting for sociodemographic characteristics, seasonal and long-term trends, humidity and particulate matter (PM_2.5).

Results

Higher ambient temperature was associated with smaller fetal growth measures. The critical window of exposure appeared to be Weeks 1–20 for ultrasound parameters, and high temperatures throughout pregnancy were important for BW. Associations were strongest for head parameters (BPD and HC) in early to mid-pregnancy, AC late in pregnancy and BW. For example, a 5ºC higher cumulative temperature exposure was associated with a lower mean AC z-score of -0.26 (95% CI: -0.48, -0.04) among 24–31-Week scans, and a lower mean BW z-score of -0.32 (95% CI: -0.51, -0.12).

Conclusion

Higher temperatures were associated with impaired fetal growth. This has major health implications given that extreme temperatures are more common and escalating.

Key Messages.

• Recent work has shown that increased heat exposure during pregnancy was associated with lower birthweight. Assessing fetal growth in utero using ultrasound measures could provide further insights on the timing of when the growth-restricting effects of heat manifest and the differential effect of temperature on fetal structures.

• We found that increased heat exposure during pregnancy was associated with smaller fetal growth measures (ultrasound parameters and birthweight), with associations strongest for head parameters (biparietal diameter and head circumference) in early to mid-pregnancy, abdominal circumference later in pregnancy and birthweight.

• The first 20 weeks of pregnancy appeared to be a critical window of heat exposure for ultrasound biometric parameters (head parameters and abdominal circumference), and higher temperatures throughout pregnancy were important for birthweight

• Our findings contribute to the growing body of evidence documenting the overall health impact of rising temperatures, highlighting the need for urgent and transformational action to combat the climate crisis.

Introduction

The climate crisis has led to warmer global temperatures, with eight of the 10 hottest years on record having occurred in the past decade.^1–3 Apart from having been recognized as an environmental emergency, several organizations, including the World Health Organization and the Lancet Countdown, have identified human health as one of its major consequences.^4,5 Pregnant individuals and their fetuses have been recognized as one group that is particularly vulnerable to heat stress.^6,7 The physiological and anatomical changes that occur during pregnancy (e.g. increased internal heat production with fetal and placental metabolism) present challenges to thermoregulation. The impaired ability to respond to high temperatures can result in cell death, disturbance of cell migration, disruption of gene expression, and damage to blood vessels and the placenta.^8–12

Two recent systematic reviews have shown that heat exposure during pregnancy is associated with lower birthweight,^13,14 which has implications for subsequent health and development.^15–17 However, there is still inconsistency with regards to the critical window of exposure for birthweight, which is likely due to the coarse temporal resolution used to explore biologically relevant windows (e.g. many prior studies used trimester-specific exposures),^13,14 and the lack of adjustment for other exposure periods during pregnancy. A distributed lag model (DLM) framework using temporally-resolved exposure data can avoid the above pitfalls, as it does not pre-specify a critical window and also produces week-specific associations that are mutually adjusted for exposures in other weeks.¹⁸ Furthermore, the exclusive use of birthweight to assess fetal growth does not provide insights on: (i) the timing of when heat-related influences on growth may manifest during pregnancy or (ii) the differential impact on various fetal structures. Routine ultrasound measurements make the ontogenetic processes of fetal development observable, and thus are an invaluable tool for identifying developmental windows during which a fetus is susceptible to heat stress. Thus, we used DLMs to examine the time-varying association between weekly-resolved temperature during pregnancy and fetal growth in a pregnancy cohort with routine ultrasounds from Eastern Massachusetts, USA.

Methods

Study population

This study population has been described previously.¹⁹ Briefly, we used prenatal and obstetric data from Beth Israel Deaconess Medical Center (BIDMC, Boston, MA, USA). Pregnant individuals, who delivered at ≥20 weeks of gestation from 2011 to 2016 and received prenatal care from the practices for which all obstetric ultrasounds are performed through BIDMC, were eligible (n = 12 967). Of these, we excluded pregnant individuals with multifetal gestations (n = 844) because the growth trajectory is different compared with singleton pregnancies, and those with a residential address outside Massachusetts (n = 408). We also excluded participants without any ultrasound measurements (n = 1377), and those with any missing covariate data (n = 160). Finally, we excluded individuals who had their first and only scan after the 23rd week of gestation, as these pregnancies likely did not have their standard prenatal care at BIDMC (n = 732); they likely moved homes (and transferred to BIDMC) later in pregnancy, and so it would have been inappropriate to assign exposures based on their residential address at birth (Supplementary Figure S1, available as Supplementary data at IJE online). Full addresses at time of birth were available for each delivery and were geocoded to latitude and longitude using the Google Maps Application Programming Interface.

Ambient temperature

We obtained daily average ambient temperature data at a 30-arcsec resolution, equivalent to an 800-m grid, from the Parameter-Elevation Regressions on Independent Slopes Model (PRISM) developed by the Spatial Climate Analysis Service at Oregon State University.²⁰ Briefly, grid-level daily temperatures were interpolated using temperature data from all weather stations across the conterminous USA as inputs in a weighted regression that accounts for elevation and other climatological factors.²⁰ The PRISM interpolation procedure is regarded as one of the most reliable for climatic data on a small scale—it not only is used by the National Aeronautics and Space Administration (NASA) and almost all professional weather services,²¹ but also, has been shown to be useful in environmental epidemiology studies.²² For each delivery, we assigned the 800-m grid in which the pregnant individual reported to have resided at the time of birth, and created weekly averages from conception to the 37th gestational week.

Fetal ultrasound

Ultrasound scans were performed and interpreted by either maternal-fetal medicine specialists or radiologists. To assess the association with fetal growth, we used four ultrasound biometric parameters—biparietal diameter (BPD), head circumference (HC), femur length (FL) and abdominal circumference (AC), all of which were recorded in millimetres (mm). The gestational age at the ultrasound examination was based on the best obstetric estimate, combining information from the last menstrual period and the earliest ultrasound performed in pregnancy.²³

We made the a priori decision to classify the ultrasound measurements into three groups: <24 weeks, 24–31 weeks and 32+ weeks. This was informed by both the standard clinical practice at BIDMC and the sequence of events in the fetal developmental process. That is, the standard of care is one ultrasound scan at about 18–23 weeks of gestation to screen for any abnormalities in the fetal anatomy, whereas those conducted later in pregnancy are typically done for diagnostic purposes (e.g. fetal growth restriction, obstetric indications etc.). These later scans (i.e. after 24 weeks) were further divided into two distinct periods (24–31 weeks and 32+ weeks) because these periods likely involve different developmental processes; for example, the brain’s functional networks (i.e. short- and long-range connections between different brain regions) begin to form in the second half of pregnancy with the development of network connections peaking between 27–30 weeks, whereas the generation of oligodendrocytes (cells that are involved in axonal myelination) peaks after 30 weeks.^24,25

For a given gestational week, we considered ultrasound measurements four standard deviations (SDs) away from the mean of the cohort implausible, and thus excluded these measurements (736 for BPD, 839 for HC, 827 for FL and 922 for AC—about 3% of all measurements for each parameter). Furthermore, to enable comparisons across gestational weeks, we generated age-specific z-scores for each of the fetal ultrasound parameters by applying the INTERGROWTH-21st standards for fetal growth.²⁶ Since these standards are only available up to 40 weeks of gestation, ultrasound scans conducted after 40 weeks were excluded. Finally, we also abstracted birthweight (BW), which was reported in grams, from the medical records and generated age- and sex-specific z-scores using the INTERGROWTH-21st standards for newborn size.²⁶ Our analysis of BW was restricted to term deliveries (≥37 weeks) so that they all have complete exposure histories up until the 37th week of gestation.

Covariates

From the medical records, we abstracted data on the following maternal and fetal characteristics: maternal age (continuous), race (White, Black, Asian, Hispanic or Other), maternal educational attainment (college and higher, lower than college, or not specified), insurance type (private or public/uninsured), parity (nulliparous or parous), self-reported smoking during pregnancy (smoker or non-smoker) and fetal sex (male or female). Furthermore, for area-level socioeconomic status, we used the national percentile rankings of the Area Deprivation Index (ADI), which ranges from 1 to 100, with 1 being the least disadvantaged and 100 being the most.²⁷ The ADI was calculated at the census block group level and represents a composite measure of neighbourhood socioeconomic disadvantage derived from 17 census variables on income, employment, and housing from the American Community Survey.²⁷ We also adjusted for humidity, as the amount of moisture in the air can interfere with thermoregulation and exacerbate the effects of high temperatures.²⁸ To do so, we downloaded data on specific humidity, defined as the mass of water vapour in a unit mass of moist air, with 12-km spatial resolution from Phase 2 of the North American Land Data Assimilation Systems (NLDAS-2) at the NASA Earth Sciences Data and Information Services Center.²⁹ These covariates were chosen a priori as they were potential confounders for the exposure-outcome association.

Finally, we also included particulate matter less than 2.5 μm (PM_2.5) in our models so that our estimates represent the direct effect of temperature (i.e. the effect of temperature not mediated by air pollution). This is because the direct effect of temperature is the scientific question that is most relevant to climate-related policies. That is, if there is a direct effect, then policies should focus on curbing exposure to rising temperatures; however, if we do not identify a direct effect, then policies should instead focus on lowering emissions of air pollutants, which we have shown to be associated with fetal growth in this setting.¹⁹ The PM_2.5 data were obtained from a well-validated ensemble model that predicts daily PM_2.5 concentration for each 1-km grid across the continental USA (cross-validated R² = 0.86).³⁰ Both humidity and PM_2.5 were assigned to each pregnancy, based on which grid they resided at birth, and weekly averages were calculated for both variables.

Statistical analysis

We estimated the time-varying association between mean weekly ambient temperature and fetal growth by fitting DLMs, which regresses the fetal growth measure on weekly temperature: where $Y_i$ is the fetal growth measure for fetus $i$; ${\mathrm{\beta }}_0$ is the intercept; $s_t\left(\mathit{Tem}{p}_{it};{\mathbf{\alpha }}_t\right)$ for $t=1,\dots ,k$ is a distributed lag nonlinear function for weekly temperature up until the week of measurement $k$ that is characterized as a linear function of regression coefficients ${\mathit{\alpha }}_t$.³¹${\mathrm{\delta }}_t$ and ${\mathrm{\psi }}_t$ are regression coefficients during gestational Week $t$ for humidity and PM_2.5, respectively; ${\mathit{z}}_i$ is a vector of baseline covariates with corresponding coefficients $\mathit{\gamma }$; and ${\mathrm{ϵ}}_i$ are independent and identically distributed errors. For each fetal growth measure, we first used three degrees of freedom (df) for the temperature-response function to allow for a potentially nonlinear relationship, as temperature health-effects have been previously shown to be either U- or J-shaped.^14,32 Due to the high correlation among measures of weekly temperature, DLMs typically constrain the lagged associations as a function of time.³³ Here we used a natural spline constraint, varying the df from 2 to 6, and selected the one that minimized the Akaike information criterion (AIC). To adjust for confounding by seasonal and long-term trends, we included a year indicator and a sine-cosine pair for each year. To adjust for specific humidity and PM_2.5, we used distributed linear lags with df for the lag constraint that matched the one for temperature. All other covariates previously described were included in the model, with linear and quadratic terms used for continuous variables (maternal age, and ADI). Furthermore, we included a random intercept for each pregnancy because ultrasounds within each pregnancy were likely to be correlated. For each model, we estimated: (i) the week-specific association between ambient temperature and the fetal growth measure, and (ii) the cumulative association, which is the expected mean change in the fetal growth measure associated with a 5ºC higher ambient temperature sustained throughout the lag period (i.e. from conception up until the fetal growth measure). Although we found that the temperature-response function did deviate from linearity in a few weeks in some models (e.g. exposure in Week 21 for BPD measured in Weeks 24–31), the overall temperature-response function (across models and weeks) did not display U- or J-shaped associations (Supplementary Figures S2–S6, available as Supplementary data at IJE online), and so we present estimates from models using distributed linear lags for the temperature-response function.

$$Y_i={\mathrm{\beta }}_0+{\sum }_{t=1}^k{s}_t\left({\mathit{Temp}}_{it};{\alpha }_t\right)+\mathrm{ }{\sum }_{t=1}^k{\mathit{Humid}}_{it}{\delta }_t+{\sum }_{t=1}^k{PM}_{it}{\psi }_t+{\mathit{z}}_i^T\mathbf{\gamma }+{\mathrm{ϵ}}_i,$$

We conducted several additional analyses. First, we conducted a sensitivity analysis adjusting for short-term temperature variability to disentangle associations from those related to short-term acclimatization, which also have been shown to be associated with pregnancy outcomes.³⁴ To do so, we calculated the SD of daily temperatures in each gestational week and included a distributed linear lag for weekly temperature SD in our DLMs. Second, we examined whether there were potential interactions between ambient temperature and PM_2.5 in order to determine whether the direct effect of temperature (i.e. the effect not through PM_2.5) depends on local PM_2.5 levels; the presence of exposure-mediator interactions, if ignored, could lead to biased effect decompositions.³⁵ Finally, we conducted subgroup analyses to assess for potential effect modification, focusing on effect modifiers reported in prior studies. Specifically, prior work has shown that the growth-restricting effect of heat exposure is larger in pregnant individuals at the age extremes (<22 or >40 years),^36–38 those who were Black or Hispanic^37,38 and those of low socioeconomic position.^37,38 We also considered potential effect modification by fetal sex, as other research has shown that prenatal exposures may be more harmful to male fetuses.^39–42 Furthermore, we also assessed whether season of conception (spring, summer, fall, winter) modified the association, as temperatures extremes during specific periods of pregnancy could be more harmful to the fetus.³⁶ For maternal race and education, individuals in the ‘Other’ and ‘Not Specified’ categories, respectively, were excluded from these analyses. We present the individual lag-responses predicted by the DLMs for each level of the modifier. To test whether the week-specific associations differed across levels of the modifier, we examined the p-value for the product term between the temperature cross-basis and the modifier. All analyses were performed in R (version 4.1.0; R Foundation for Statistical Computing, Vienna, Austria).⁴³

Results

Characteristics of the 9446 pregnancies that constituted our study population are shown in Table 1. On average, participants were 31 years of age at conception, with few individuals <22 years or >40 years (approximately 5% each). The majority were White (52%) and had private insurance (79%). About half had completed college or higher (49%), and about half were nulliparous (49%). Furthermore, the mean ADI percentile was 22 (median nationwide is 50), which indicates that our cohort comprised individuals who lived in neighbourhoods with less disadvantage relative to the rest of the USA. Most pregnancies had two or more ultrasound measurements (Table 1; Supplementary Figure S7, available as Supplementary data at IJE online), and summary statistics for the fetal growth measures are shown in Table 2. Throughout pregnancy, we observed that BPD was on average smaller than the international norm,²⁶ whereas, FL, AC and BW were larger (Table 2). Week-specific distributions and correlations of mean temperature, PM_2.5, specific humidity and temperature variability are presented in Supplementary Figures S8–S10 (available as Supplementary data at IJE online). Weekly mean temperature was about 10ºC throughout pregnancy (Supplementary Figure S8), and this did not differ between those with one ultrasound (the standard of care) and those with two or more (pregnancies with possibly higher risk of complications) (Supplementary Figure S9). Weekly SD temperature (i.e. a measure of temperature variability) was about 3ºC, average PM_2.5 was about 7 µg/m³, which is below current annual ambient standards,⁴⁴ and average specific humidity was about 6 mg/kg (Supplementary Figure S8). Mean ambient temperature was positively correlated with specific humidity but showed weak correlations with PM_2.5 and temperature variability throughout pregnancy (Supplementary Figure S10).

Table 1.

Maternal and fetal characteristics of deliveries at Beth Israel Deaconess Medical Center, Boston, Massachusetts from 2011 to 2016 (N = 9446)

Characteristics	N (%) or mean ± SD
Age at conception (years)
Mean ± SD	31.4 ± 5.4
<22 years	507 (5)
22–40 years	8497 (90)
>40 years	442 (5)
Education
College or higher	4608 (49)
Lower than college	3193 (34)
Not specified	1645 (17)
Race
White	4916 (52)
Black	1626 (17)
Asian	899 (10)
Hispanic	882 (9)
Other	1123 (12)
Parity
Nulliparous	4655 (49)
Parous	4793 (51)
Smoking
Smoker	226 (2)
Non-smoker	9220 (98)
Fetal sex
Female	4655 (49)
Male	4791 (51)
Insurance
Private	7436 (79)
Public or uninsured	2010 (21)
Area Deprivation Index (percentile)
Mean (SD)	22 (20)
ADI quartile 1 [1, 9)	2611 (28)
ADI quartile 2 [9, 17)	2209 (23)
ADI quartile 3 [17, 28)	2343 (25)
ADI quartile 4 [28, 100)	2283 (24)
Conception season
Spring	1994 (21)
Summer	2401 (25)
Autumn	2604 (28)
Winter	2447 (26)
Number of ultrasounds
1 ultrasound	2106 (22)
2+ ultrasounds	7340 (78)

SD, standard deviation; ADI, Area Deprivation Index.

Table 2.

Summary statistics for fetal ultrasound measures and birthweight from deliveries at Beth Israel Deaconess Medical Center, Boston, Massachusetts in 2011–16 (N = 9446)

Growth measure	N	Mean ± SD	Median
			[25th percentile, 75th percentile]
Biparietal diameter
16–23 weeks	10153	–0.68 ± 1.10	–0.67 [–1.38, 0.06]
24–31 weeks	5632	–0.87 ± 1.15	–0.87 [–1.63, –0.11]
32–40 weeks	9324	–0.83 ± 1.09	–0.81 [–1.52, –0.10]
Head circumference
16–23 weeks	10117	0.37 ± 1.09	0.38 [–0.32, 1.06]
24–31 weeks	5608	0.38 ± 1.20	0.35 [–0.43, 1.18]
32–40 weeks	9287	0.32 ± 1.11	0.30 [–0.37, 1.03]
Femur length
16–23 weeks	10097	0.82 ± 1.13	0.81 [0.12, 1.50]
24–31 weeks	5610	1.05 ± 1.19	1.06 [0.31, 1.82]
32–40 weeks	9305	1.15 ± 1.02	1.19 [0.54, 1.79]
Abdominal circumference
16–23 weeks	10074	0.52 ± 1.16	0.49 [–0.24, 1.25]
24–31 weeks	5603	0.66 ± 1.13	0.63 [–0.07, 1.39]
32–40 weeks	9293	0.51 ± 1.06	0.47 [–0.18, 1.15]
Birthweight
Term (37–43 weeks)	8636	0.35 ± 0.99	0.34 [–0.31, 1.01]

SD, standard deviation

Figures 1 and 2 show the week-specific associations between ambient temperature and fetal growth measures. For almost all fetal biometric measurements, we found the associations were strongest with temperature earlier in pregnancy—the critical window of exposure appeared to be from conception to about the 20th gestational Week (Figure 1). However, this was not the case for FL measured late in pregnancy and BW, as we found that temperatures experienced in other parts of pregnancy appeared important; for example, higher temperatures in Weeks 18–24 were negatively associated with FL among 32–40 -Week scans (Figure 1), and average entire pregnancy temperatures appeared important for BW (Figure 2). Table 3 shows the cumulative associations between ambient temperature and the fetal growth measures. We found that the cumulative association was particularly strong for head parameters measured in early to mid-pregnancy; for example, a 5ºC higher cumulative temperature exposure was associated with a lower mean HC z-score of -0.17 (95% CI: -0.27, -0.08) among 16–23-Week scans, and -0.31 (95% CI: -0.50, -0.12) among 24–31-Week scans (Table 3). On the other hand, associations appeared to be stronger later in pregnancy for parameters related to fat accumulation; for example, a 5ºC higher cumulative temperature exposure was associated with a lower mean AC z-score of -0.26 (95% CI: -0.48, -0.04) among 24–31-Week scans, and a 5ºC higher cumulative temperature exposure throughout the first 37 weeks was associated with a lower mean BW z-score of -0.32 (95% CI: -0.51, -0.12). Finally, we also found suggestive evidence that higher cumulative temperature may be negatively associated with FL among 32–40 Week scans; for example, a 5ºC higher cumulative temperature exposure was associated with a lower mean FL z-score of -0.20 (95% CI: -0.42, 0.02) among these later scans. In our sensitivity analysis adjusting for temperature variability, we found that the estimates were similar to our primary analyses (Supplementary Figures S11 and S12 and Supplementary Table S1, available as Supplementary data at IJE online). Furthermore, we did not find compelling evidence of an interaction between ambient temperature and PM_2.5 (Supplementary Figures S13 and S14 and Supplementary Table S2, available as Supplementary data at IJE online).

Figure 1.

The time-varying association between weekly ambient temperature and ultrasound parameters of fetal growth from deliveries at Beth Israel Deaconess Medical Center, Boston, Massachusetts in 2011–16 (N = 9446) estimated using distributed lag models adjusted for maternal age, race, education, insurance type, parity, smoking, fetal sex, conception year, day of the year of conceptions, Area Deprivation Index, humidity and PM_2.5. The lagged associations were constrained using natural splines with degrees of freedom selected based on the smallest Akaike information criterion. Black solid lines show the lag-response estimates, grey shaded areas show the 95% confidence intervals and the black dashed line indicates the null hypothesis of no effect across all weeks. PM, particulate matter

Figure 2.

The time-varying association between weekly ambient temperature and birthweight from term deliveries (≥37 weeks) at Beth Israel Deaconess Medical Center, Boston, Massachusetts in 2011–16 (N  =  8636) estimated using distributed lag models adjusted for maternal age, race, education, insurance type, parity, smoking, fetal sex, conception year, day of the year of conceptions, Area Deprivation Index, humidity and PM_2.5. The lagged association was constrained using a natural spline with the degree of freedom selected based on the smallest Akaike information criterion. Black solid lines show the lag-response estimates, grey shaded areas show the 95% confidence intervals and the black dashed line indicates the null hypothesis of no effect across all weeks. PM, particulate matter

Table 3.

Distributed lag model (DLM) estimates and 95% CIs for the cumulative association between weekly temperature and fetal growth measures (ultrasound parameters and birthweight) from deliveries at Beth Israel Deaconess Medical Center, Boston, Massachusetts in 2011–16 (N = 9446)

Fetal growth measure	Cumulative estimate (95% CI)
Biparietal diameter
16–23 weeks	−0.23 (−0.33, −0.13)
24–31 weeks	−0.21 (−0.40, −0.03)
32–40 weeks	−0.03 (−0.25, 0.20)
Head circumference
16–23 weeks	−0.17 (−0.27, −0.08)
24–31 weeks	−0.31 (−0.50, −0.12)
32–40 weeks	−0.03 (−0.26, 0.20)
Femur length
16–23 weeks	−0.16 (−0.27, −0.06)
24–31 weeks	−0.14 (−0.34, 0.06)
32–40 weeks	−0.20 (−0.42, 0.02)
Abdominal circumference
16–23 weeks	−0.14 (−0.25, −0.03)
24–31 weeks	−0.26 (−0.44, −0.07)
32–40 weeks	−0.26 (−0.48, −0.04)
Birthweight
Term (≥37 weeks)a	−0.32 (−0.51, −0.12)

The cumulative association is the expected mean change in the fetal growth measure associated with a 5ºC higher cumulative ambient temperature exposure; all DLMs were adjusted for maternal age, race, education, insurance type, parity, smoking, fetal sex, conception year, day of the year of conception, Area Deprivation Index, humidity and particulate matter (PM_2.5).

^aThere were 8641 term deliveries.

In our effect modification analyses (Supplementary Figures S15–S26 and Supplementary Tables S3–S8, available as Supplementary data at IJE online), we found that maternal age, fetal sex and season of conception potentially modified the associations (Supplementary Figures S15–S20 and Supplementary Tables S3–S5). For maternal age, associations with head parameters in early pregnancy appeared stronger for pregnant individuals aged <22 years compared with those aged 22+ years (Supplementary Figures S15–16 and Supplementary Table S3, available as Supplementary data at IJE online). For fetal sex, we found that associations with ultrasound parameters for females were consistently stronger than for males (Supplementary Figures S17 and S18 and Supplementary Table S4, available as Supplementary data at IJE online). For season of conception, associations with ultrasound parameters appeared stronger for autumn conceptions compared with those that occurred in other seasons (Supplementary Figures S19 and S20, Supplementary Table S5, available as Supplementary data at IJE online). We did not find any evidence that maternal race, educational attainment or ADI modified the associations (Supplementary Figures S21–S26 and Supplementary Tables S6–S8, available as Supplementary data at IJE online).

Discussion

In this large pregnancy cohort from Eastern Massachusetts, USA, we found that increased ambient temperature was associated with all fetal growth measures, which has implications for subsequent health and development as each of these measures has been associated with later health outcomes.^15,45–50 Although we found strong associations with both head parameters in early to mid-pregnancy and parameters related to fat development later in pregnancy, the associations with other fetal growth measures in other windows are still important for public health, given that exposure to high temperatures is common and escalating.^1–3

In our analyses of critical exposure windows, we found that ultrasound parameters were sensitive to higher temperatures in the first 20 weeks of pregnancy. One possible mechanism is that increased heat exposure leads to the production of heat-shock proteins, which can disrupt normal protein synthesis in early pregnancy, leading to altered fetal organ development.⁵¹ Given the strong associations we see for head size parameters measured in early to mid-pregnancy, the ontogenetic processes of brain development during early to mid-pregnancy (e.g. neurogenesis, neuronal migration etc.) appeared to be particularly vulnerable, where any perturbations to these processes could potentially have long-term effects on brain development.⁵² Furthermore, we also see lagged associations with other parameters, in that higher temperatures in early to mid-pregnancy were associated with FL and AC measured at 32 weeks and beyond. FL late in pregnancy is driven mainly by rapid bone growth and mineralization,⁵³ whereas AC in this period is driven mainly by the enlargement and accumulation of adipocytes.⁵⁴ Heat stress early in pregnancy could affect these processes late in pregnancy through interfering with the formation of the placenta. The placenta is not fully formed until the 16th week of gestation,⁵⁵ and so heat-related disruptions to this organ can disrupt placental blood flow and transplacental nutrient exchange throughout gestation, thus continually depriving the fetus of adequate nutrition.

In our BW analysis, we found that higher temperatures throughout pregnancy appeared to be important. Unlike the analysis of ultrasound parameters, we found that BW was also sensitive to higher temperatures later in pregnancy (after 20 weeks), which suggests that heat stress could also disrupt transplacental nutrient exchange through mechanisms other than interfering with placental formation (which occurs before 20 weeks). One possibility is that overheating is more common as the fetus grows and maternal body mass index increases, and so thermoregulation is more difficult as the pregnancy progresses; this could divert too much blood away from the developing fetus and interfere with the accrual of fat late in pregnancy.⁵⁶ Several prior studies of BW have explored the question of critical exposure windows during pregnancy, but this body of work has produced results that are inconclusive.¹⁴ Some have found associations with high temperatures in early to mid-pregnancy, whereas others found them with high temperatures late in pregnancy.^13,14 This inconsistency is likely due to the coarse windows considered (e.g. typically trimester-specific exposures) and also to the lack of mutual adjustment for exposures in other periods which can lead to biased estimates.¹⁸ Two recent studies have used DLMs (which avoids the issues described above) to examine the relationship between temperature and BW.^32,34 The first study conducted in France did not find an association between mean temperature and BW, but this is likely due to the small sample size as they had about 4500 participants.³⁴ The second study conducted in Israel was much larger as they used the national birth cohort registry of over 600 000 participants—they found that BW was associated with higher temperatures in Weeks 3–9 and 19–34.³² The results of our BW analysis were similar to those of the Israeli study but differed in that we found that higher temperature in Weeks 10–18 also appeared to be relevant. One possible explanation for this discrepancy is that the two settings are not comparable in terms of climate (e.g. temperature distribution and seasonality, dryness, humidity etc.). That is, although summer temperatures may be similar, the New England area experiences overall colder temperatures in other parts of the year. Thus, differences in the estimates may reflect the degree of acclimatization to the local environment, in that hotter temperatures in an already warm environment may not be as detrimental as those in a typically cold environment.

One unexpected finding was that cold temperatures did not seem to have an impact on fetal growth. We found the temperature-response functions in our DLMs to be relatively linear. Though there were few deviations from linearity for some weeks, overall there was no evidence for U- or J-shaped temperature-health effects as has been previously reported.^14,32 This may be because our cohort comprises individuals who live in urban neighbourhoods with little disadvantage, such that they may be better at mitigating the effects of extreme cold (e.g. heating, adequate clothing, staying indoors). Furthermore, colder winters in New England are shorter and are becoming less common,⁵⁷ where the proportion of participants experiencing weekly mean temperatures below freezing ranged from 10–15% (depending on the gestational week). Thus, although there may be an effect of cold temperatures on fetal growth, our limited sample size at temperature extremes may not make it possible to detect such an effect. These considerations suggest that perhaps the linear dose-repose relationship we observed may not be generalizable to other settings with different population characteristics, temperature distributions and seasonality patterns.

Results from our effect modification analyses should be interpreted with caution, as we had few individuals in certain subgroups (e.g. <22 years, >40 years, Asian, Hispanic etc.), such that associations could be attributed to random chance and statistical fluctuation. Yet, we present these estimates to earmark them for future investigation as they may be potentially meaningful. We found suggestive evidence that maternal age could potentially modify the association between heat stress and impaired fetal growth. Negative associations were, in general, observed to be stronger for pregnant individuals <22 years, which coincides with past literature suggesting that those at the extremes of reproductive age are at an increased risk of adverse birth outcomes.⁵⁸ We also found that associations potentially differed by fetal sex, with females consistently experiencing stronger negative lag-responses. This may at first appear to contradict previous literature showing that male fetuses were more vulnerable to stressors during pregnancy,^39–42 but the two findings can be reconciled given that male fetuses are less likely to survive to birth in response to stressful exposures, as shown in several studies by a drop in the secondary sex ratio.^40–42 Thus, if higher temperatures do preferentially deplete the male fetal population, then the estimates we present here are biased upwards as they would be estimated in a healthier subpopulation of male fetuses. Furthermore, we also found that season of conception could potentially modify the associations with autumn conceptions appearing more vulnerable for some parameters. Similarly to why there may be discrepancies between the Israeli study and ours, this may be because autumn conceptions experience colder temperatures earlier during their pregnancy and so are not acclimatized to hotter than expected temperatures during that season, whereas spring and summer conceptions may be better adapted to hotter temperatures through more air conditioning usage and altered behaviours (e.g. choice of clothing). Although we did not find differences by maternal race, there is some indication that there may be racial disparities (e.g. though comparisons were likely underpowered, BW estimates appeared stronger for non-White individuals). This may be because non-White individuals are more likely to experience more intense urban heat islands.⁵⁹ However, we did not have enough non-White individuals in our sample to identify these disparities. We found no evidence of effect modification by either educational attainment or ADI. One possible explanation, akin to the rationale for the lack of effect for cold temperatures, is that unlike previous studies, the socioeconomic distribution of our cohort falls within a narrow range. Past studies that found differences were able to do so because, perhaps, they included individuals from a broader range of socioeconomic contexts. For example, a study using birth certificates in California, in which individuals from lower socioeconomic groups were better represented in the analytical sample, found differences by educational attainment.³⁸ Further research is warranted in other settings with different demographic and socioeconomic compositions to better understand the intersectionality between race, ethnicity and socioeconomic position on the impact of rising temperatures during pregnancy.

A key strength of this work is the use of fetal ultrasound parameters, which allowed us to investigate the timing of when the growth-restricting effects of heat manifest and the differential impact on different fetal structures. Other major strengths include the use of temperature data with 800-m spatial resolution,²⁰ which helped minimize exposure misclassification, as well as our DLM analyses of weekly temperature to explore critical exposure windows during pregnancy. The DLM framework allows for the mutual adjustment of temperature in other weeks, which helps minimize potential confounding by seasonality of temperature during pregnancy. Furthermore, we also adjusted for PM_2.5 concentrations, so that our estimates represent the direct effects of temperature. This is important to identify because it suggests that the heat-related growth-restriction is due to the heat itself (likely through mechanisms related to thermoregulation), rather than the downstream effects of air pollution which primarily act through oxidative stress.⁶⁰ Other studies that did not control for air pollution could be identifying air pollution effects and falsely attributing them to heat, in which case lowering temperature would do little to improve fetal health if air pollutants were still being emitted. Finally, we also conducted several additional analyses to assess the robustness of our findings. We found that our estimates were independent of temperature variability, as controlling for weekly SD temperature did not change our results. Furthermore, we also did not find any compelling evidence of an interaction between temperature and PM_2.5, which suggests that the direct effects we estimated were unbiased.³⁵ That is, in the absence of exposure-mediator interactions, it is appropriate to estimate the controlled direct effect.

We also acknowledge several limitations. Although we used temperature data with high spatial and temporal resolution, there were other sources of exposure measurement error: (i) temperature estimates were based on ambient values and so may not reflect personal exposure; and (ii) assignment was based on residential address at delivery, and so is agnostic towards both residential mobility and time-activity patterns during pregnancy. However, these errors were likely non-differential with respect to fetal growth conditional on the covariates.^61–63 Second, although most of the temperature-response functions appeared linear across models and weeks, it is possible that the exposure-response function is mis-specified for a few weeks. Although this may result in slightly biased estimates for some weeks (as we are smoothing over something curvilinear), it would not lead to associations in the opposite side of the null, given that we did not observe U- or J-shaped exposure-response functions. Furthermore, our estimates are vulnerable to live birth bias since our analyses were restricted to live-born children.⁶⁴ That is, if heat exposure also causes pregnancy loss, then the effect of heat would likely appear less harmful as fetuses who are more susceptible to the effects of heat do not survive to birth. Given that we used medical records to construct our cohort, we could not enumerate the number of pregnancy losses in this cohort because most losses occur before the individual engages with the medical system. For example, in this cohort, we were able to identify about 1600 losses (miscarriages, spontaneous abortions and stillbirths) from the medical system, which amounts to about 11% of pregnancies; however, given that prior work has shown that as much as 30% of all pregnancies end in loss,⁶⁵ this suggests that there were many more losses that were not captured because they occurred before the individual was aware that she was pregnant, and so would not be clinically recognized. Finally, our study was based on retrospective ultrasound data, and so ultrasound scans from the second half of pregnancy were likely enriched with high-risk pregnancies, as the standard of care at BIDMC dictates that a follow-up scan after 24 weeks is not required for pregnancies without complications. However, we show that the exposure distribution did not differ between those who had a single ultrasound scan (the standard of care) and those who had follow-up scans (those with two or more ultrasound scans). Thus, although this may not cause selection bias, it could potentially limit the generalizability of our estimates to other populations with a different distribution of healthy pregnancies.

Conclusion

In conclusion, we showed that heat exposure during pregnancy was associated with smaller fetal growth measures in this Eastern Massachusetts cohort. Our DLM analyses identified the first 20 weeks as a critical window for ultrasound parameters and higher temperatures throughout pregnancy appeared important for birthweight. Future research should explore this topic further in other settings with different temperature distributions/seasonality and demographic/socioeconomic compositions. These novel findings contribute to the growing body of evidence documenting the overall health impact of rising temperatures, which further highlights the importance of investment into preventive measures for pregnant individuals, heat warning systems, and more broadly, advocacy for regulations to mitigate the climate crisis.

Ethics approval

The institutional review boards of Harvard T.H. Chan School of Public Health and Beth Israel Deaconess Medical Center approved this study.

Data availability

The PRISM temperature data are available from the Spatial Climate Analysis Service at Oregon State University. The humidity data are available from Phase 2 of the North American Land Data Assimilation Systems at the NASA Earth Sciences Data and Information Services Center. The Area Deprivation Index data are available from the Neighborhood Atlas. The PM_2.5 data are available from the corresponding author on reasonable request. To protect the privacy of participants, the Beth Israel Deaconess Medical Center dataset with outcomes could not be shared.

Supplementary data

Supplementary data are available at IJE online.

Author contributions

M.L., F.L., B.C., M.G.F., A.M.M., M.R.H., B.J.W. and S.P. designed the research and directed its implementation. M.L., A.M.M., M.R.H., H.S.I., J.E.H., Y.W. and J.S. prepared the datasets. M.L. and S.P. analysed the data. M.L., F.L., B.C., M.G.W. and S.P. wrote the paper and all authors contributed to the revision of the manuscript.

Funding

This work was supported by grant P30ES000002 from the National Institutes of Health (NIH), and United States Environmental Protection Agency (USEPA) grant RD-835872. Its contents are solely the responsibility of the grantee and do not necessarily represent the official views of the USEPA. Further, USEPA does not endorse the purchase of any commercial products or services mentioned in the publication.

Conflict of interest

None declared.