Disproportionate climate burden of rising temperatures on low birth weight in Pakistan

Abstract

Background

Warming temperatures add to the global health burden, with disproportionate effects on pregnant women and newborns. Low birth weight is a major neonatal health issue in Pakistan, leading to neonatal mortality and impaired long-term health. We assessed the impact of extreme temperatures on low birth weight, identified high-risk subgroups, estimated the heat-attributable burden, projected future risks, and constructed a district-level heat vulnerability index.

Methods

We conducted a space–time series study using nationally representative surveys from 2008 to 2017 across Pakistan’s provinces. We modelled temperature–low-birth-weight associations with distributed-lag non-linear models in a generalised mixed-effects framework, with model averaging to address specification uncertainty. Subgroup analyses considered maternal education, household wealth, urban/rural residence, and air quality. We estimated heat-related population attributable fraction using observed temperature and projections under SSP2-4.5 and SSP5-8.5. Province-level risk estimates combined with district-level indicators, such as mean temperature, multidimensional poverty, and under-5 mortality, were used to develop the heat vulnerability index.

Results

The study included 85,017 participants, with 15,920 (18.72%) infants identified as having low birth weight. Heat-related risks for low birth weight varied across provinces, with relative risks ranging from 1.47 (1.07–2.03, 95% confidence interval) to 1.91 (1.24–2.93) at the 99th percentile of temperature. The heat-related population attributable fraction ranged from 9.39% to 13.15%, translating to 1.24 million heat-related low-birth-weight cases over the study period. Projections indicate that heat-related population attributable fractions will increase by 8.43–10.20% by the 2060s. Subgroup analysis showed higher risk among women exposed to hazardous air pollution, those with less education, and urban residents. Women in southern Punjab, northern Baluchistan, and Sindh faced the highest risks.

Conclusions

Our findings identify Pakistan’s districts most vulnerable to heat-related low birth weight and highlight contributing factors. These insights can inform targeted interventions to mitigate risks. The study advances the understanding of the impacts of rising temperatures, particularly in resource-limited and high-risk settings.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12916-026-04664-8.

Background

Climate change is a major threat to global health, with disproportionate impacts on vulnerable populations such as pregnant women and their newborns [1, 2]. Rising temperatures and more frequent heatwaves driven by climate change have led to higher rates of illness and death, particularly in low- and middle-income countries [3]. In South Asia, the likelihood of heatwaves has increased by approximately 45 times since the pre-industrial era, with temperatures now 0.85 °C higher [4]. This region is among the most vulnerable globally to climate-related impacts [5].

Pakistan is already grappling with socio-economic challenges and is among the countries most severely affected by climate change [6]. Over recent decades, the country has experienced more extreme weather events — heatwaves, devastating floods, and prolonged droughts — largely attributed to shifting climate patterns [7, 8]. These environmental factors disproportionately affect marginalised populations, including pregnant women, who face heightened risks due to physiological vulnerability, widespread poverty, malnutrition, inadequate access to healthcare, and resource scarcity [2, 9, 10]. Pakistan has one of the highest neonatal mortality rates globally, with 42 deaths per 1000 live births, according to the Pakistan Demographic and Health Surveys 2017–2018 [11, 12]. Compounding this issue is the country’s substantial burden of low birth weight, with 19–32% of infants born weighing < 2.5 kg. Low birth weight is a leading contributor to neonatal mortality and is associated with long-term health impairments such as stunted growth and cognitive deficits [13, 14].

Adverse birth outcomes — including low birth weight — are observed to increase during hot seasons, consistent with heat-related maternal dehydration and behavioural adaptations [15, 16]. Although considerable research exists on the effects of temperature extremes on birth outcomes, most of these studies originate from high-income countries with comprehensive birth registry data [17]. In contrast, research remains limited in low- and middle-income countries like Pakistan, where populations are more vulnerable to climate-related impacts.

In addition to temperature extremes, worsening air pollution in Pakistan’s urban centres has emerged as a major public health concern, especially for pregnant women and their newborns [18–20]. Particulate matter (e.g. measured as PM_2.5) can cross the alveolar–blood barrier and trigger oxidative stress and systemic inflammation, disrupting placental function and foetal growth [21]. Accordingly, epidemiologic studies link prenatal PM_2.5 exposure to increased risks of low birth weight and preterm birth and to reduced mean birth weight [22, 23]. Although air pollution is not the primary focus of our study, we included it as a confounder to account for its independent effects on maternal and neonatal health.

Pakistan presents a critical case for investigating the impacts of heat on marginalised and resource-limited communities. The country is highly susceptible to climate change and faces compounded challenges, including rapid population growth, limited healthcare infrastructure, widespread poverty, and socio-political instability. Evidence on the heat–low birth weight association in Pakistan is emerging; for example, two studies have assessed this relationship: one based on district-level analysis and another providing national figures derived from a global multi-country model [10, 24]. Neither provides Pakistan-specific, subnational estimates calibrated on nationally representative survey data. This highlights a gap in locally relevant evidence needed to inform adaptation strategies and policy interventions in the most vulnerable countries.

Our study addresses this gap by providing Pakistan-specific evidence at both the national and provincial scales using the Demographic and Health Surveys and Multiple Indicator Cluster Surveys, coupled with province-level distributed-lag non-linear models to obtain locally calibrated relative risks. We then translate these into temperature-attributable fractions for the baseline and future scenarios and downscale them to a district-level heat vulnerability index to identify high-risk areas for targeted action.

Our objective is to assess the impact of extreme heat on low birth weight in Pakistan, identify high-risk subgroups and geographic areas, and project future risks under two climate change scenarios. By integrating maternal health and environmental datasets within a robust statistical framework, we provide decision-ready, subnational evidence to inform climate-resilient health planning in low- and middle-income settings.

Methods

Study area

Pakistan’s provinces span diverse zones — Punjabin the east along the Indus plains, Sindh in the south–southeast including the Thar Desert, Khyber Pakhtunkhwa in the northwest, Baluchistan in the arid southwest, Gilgit-Baltistan in the high northern Karakoram–Himalaya, Azad Jammu and Kashmir in the north-eastern foothills, and the Islamabad Capital Territory in the north-central Pothohar Plateau. Summers are hot in the south/centre, the July–September monsoon concentrates rainfall in the east and north, and winters are cold in the north but milder in the south. These contrasts generate wide spatial variation in temperature exposure relevant to maternal risk. Additional File 1: Fig. S1 shows the provincial boundaries and geographic location of Pakistan in the South Asia region.

Data sources

Health and socio-economic data

We used publicly available, internationally recognised datasets in this study. In total, we compiled 11 datasets: 1 Pakistan Demographic and Health Survey (PDHS 2012–2013) [25] and 10 UNICEF Multiple Indicator Cluster Surveys (MICS), spanning 2010–2020 [26]. The Multiple Indicator Cluster Surveys included two from MICS-4 (2010–2011), four from MICS-5 (2014–2017), and four from MICS-6 (2017–2020). These surveys provided a nationally representative sample of maternal and child health data for women aged 15–45 years. To avoid temporal overlap with MICS-6 (2017–2020), we excluded the Pakistan Demographic and Health Survey (PDHS 2017–2018) (Additional File 1: Table S1).

The datasets provided individual-level information on maternal and child health. The surveys record (at minimum) the month and year of birth, newborn sex, birth order, whether a birth weight was recorded, and — if weight was unavailable — maternal recall of size at birth (‘very small’, ‘smaller than average’, etc.). We defined low birth weight as ≤ 2.5 kg when a weight was recorded or as ‘very small/smaller than average’ when only recall was available. While birth size recalled by mothers is a subjective measure and inherently prone to recall bias — relying on the mother’s perception rather than objective measurement — this limitation is well-documented in the literature [27].

In the context of Pakistan’s Demographic and Health Surveys and Multiple Indicator Cluster Surveys, birth size is often the only available proxy for birth weight, especially in rural or resource-limited settings where accurate birth weight measurements are rarely recorded. For instance, birth weight data were reported for only 16.7% of children born in the last 5 years, and 90% of these were based on maternal recall rather than on documented weight. Although birth size is a weaker predictor on its own, previous studies have assessed its usefulness and suggested 86% agreement between birth weight and mothers’ perceived size at birth [27, 28]. We acknowledged this limitation in the following sections and conducted a sensitivity analysis to assess the associations of heat with recorded birth weights only.

Socio-economic variables included maternal education, wealth index, and place of delivery. We obtained district-level data on the multidimensional poverty index and under-5 mortality rate (2000–2017) from the United Nations Development Programme’s 2014–2015 data (undp.org) and Global Health Data Exchange (ghdx.healthdata.org), respectively [29, 30] (Additional File 1: Fig. S2). We used publicly available datasets for all analyses, which are accessible upon request from their data custodians. Because these datasets are de-identified and publicly accessible, ethical approval was not required.

Environmental data

We sourced monthly gridded meteorological datasets from Copernicus ERA5-Land (cds.climate.copernicus.eu) at a 9 km × 9 km resolution, including mean temperature (°C), precipitation (m), and dew point (°C) [31]. We obtained mean monthly air pollution data for particulate matter (PM_2.5 in μg m⁻³) from the global data repository by the Atmospheric Composition Analysis Group at a spatial resolution of 11 km (sites.wustl.edu) [32]. We extracted all datasets for January 2008–December 2017.

We extracted ERA5-Land/PM_2.5 within each district and then aggregated to the province level by taking the arithmetic mean across constituent districts. In a sensitivity analysis, we constructed population-weighted provincial time series using WorldPop 2020 district populations as weights. Population-weighted and arithmetic-mean series were almost identical (precipitation R = 0.996, mean temperature R = 0.997, relative humidity R = 0.996; all n = 864; Additional File 1: Fig. S3, Additional File 1: Table S2), so we retained the arithmetic-mean aggregation in the main analyses. This two-step approach preserved sub-provincial variability and produced province-level estimates that reflected differences across districts rather than a single, province-wide raster value.

Study design

Space–time series design

We employed a space–time series design to investigate long-term, delayed, and non-linear associations between monthly mean temperatures and low birth weight at the provincial level in Pakistan. This design used generalised linear mixed-effects models with province-level random intercepts and a distributed-lag non-linear framework to capture delayed effects [33]. To address model uncertainty, we compared prespecified distributed lag non-linear model specifications and calculated Akaike information criterion (AIC)-weighted model averages of the coefficient vectors and variance–covariance matrices [34]. We handled uncertainty arising from missing data by pooling estimates across multiply imputed datasets using Rubin’s rules for multiple imputation [35] (see the ‘Analysis’ section). By leveraging this design, we effectively addressed challenges associated with model uncertainties, spatiotemporally sparse data, sampling gaps, and potential missingness inherent in the survey-based nature of Demographic and Health Surveys and Multiple Indicator Cluster Surveys.

Data aggregation

We aggregated low birth weight counts by birth month–year because exact birth dates were unavailable in the survey. We combined mean monthly environmental data, including mean temperature, humidity, and air pollution, with monthly health data, using the province, month, and year information as linking variables. We did all analyses at the provincial level to mitigate issues of zero inflation inherent in district-level analyses and to capture regional trends and vulnerabilities. We aggregated monthly low-birth-weight cases and environmental variables at the provincial level.

Analysis

Modelling framework

We modelled the association between monthly mean temperature exposure and low birth weight using generalised linear mixed-effects models with a negative binomial error distribution to account for overdispersion. We included province-level random intercepts to account for spatial nonindependence and an offset for the log of the population of women of reproductive age. We specified temperature effects via a distributed-lag non-linear model cross-basis on scaled temperature, with a quadratic exposure function (degree = 2) and a first-degree lag function over 0–7 months, to capture long-term lagged effects across the entire pregnancy period, thereby reflecting cumulative maternal exposure to temperature during gestation [36]. The selection of the lag period was informed by existing studies [36] and further validated by assessing model performance based on AIC. We centred all distributed-lag non-linear models at the province-specific baseline median temperature. We chose the median a priori for interpretability and stability across provinces.

We specified a priori a causal directed acyclic graph for the association between mean temperature (exposure) and low birth weight (outcome), informed by the epidemiological literature and domain expertise (Additional File 1: Fig. S4, Additional File 1: Table S3 [24, 37–51]). The directed acyclic graph encodes relations among temperature, PM_2.5, province (structural/contextual factors), calendar trend, urbanicity, relative humidity, precipitation, maternal socio-economic status (including education and wealth index), and place of delivery. We used the DAGitty package in R [52] to identify a minimal sufficient adjustment set for estimating the total effect of temperature on low birth weight and to fit our primary models accordingly. In this specification, PM2.5 is treated as a confounder (solid paths in Additional File 1: Fig. S4). Humidity and precipitation are identified as secondary pathways (dotted), so we excluded them from the minimal set, but we added robustness models to assess sensitivity to alternate specifications (see Model comparison and averaging). Because PM_2.5 might plausibly lie on a pathway linking heat exposure to adverse birth outcomes (Additional File 1: Fig. S4, Additional File 1: Table S3), we re-estimated models excluding PM_2.5 in the sensitivity analysis.

We used monthly mean temperature (rather than a composite heat index) as the exposure in the distributed-lag non-linear models to isolate the temperature–low birth weight association, avoid embedding humidity (a potential effect modifier/mediator) into the exposure itself, and because temperature-only specifications consistently showed better model fit (lower AIC) across provinces, especially in cooler regions (e.g. Gilgit Baltistan) where standard heat index formulations are less informative. We adjusted for education and household wealth index in the main models and excluded place of delivery from the main models to avoid potential overadjustment and because of survey heterogeneity, but we assessed its effects regardless in sensitivity analyses.

Model comparison and averaging

We compared six candidate model specifications as follows: (0) null: adjusted for PM_2.5, maternal education, and wealth index; (1) linear: included a linear temperature-response relationship; (2) quadratic: a quadratic temperature-response relationship adjusted for PM_2.5, maternal education, and wealth index; (3) quadratic: a quadratic temperature-response relationship adjusted for PM_2.5, humidity, maternal education, and wealth index; (4) quadratic: adjusted for PM_2.5, precipitation, maternal education, and wealth index; and (5) quadratic: adjusted for PM_2.5, humidity, precipitation, maternal education, and wealth index.

We did multi-model comparisons based on AIC. For inference, we computed AIC weights across the quadratic distributed-lag non-linear model specifications and combined their coefficients and variance–covariance matrices using these weights (Additional File 1: Table S4). We used model-averaged coefficients to obtain relative risks at specific percentiles (1st, 10th, 90th, and 99th) and across the temperature range. We present results as relative risks with 95% confidence intervals.

Handling missing data

We imputed missing data for socio-economic categorical variables — maternal education, place of delivery, wealth index, mother’s age group, and urban residence — using multiple imputations (via the mice package in R) [53]. We employed polytomous logistic regression (polyreg) as the imputation method, leveraging relationships between these variables and other predictors in the dataset. We generated five imputed datasets (m = 5) to account for uncertainty. For each imputed dataset, we fit all candidate distributed-lag non-linear models; we then pooled the cross-basis coefficient vectors and the variance–covariance matrices across imputations using Rubin’s rules [35] and subsequently performed AIC-weighted model averaging to obtain the final coefficient vector and the covariance. Uncertainty from imputation and model selection was propagated via the pooled/averaged covariance when deriving relative risks and the 95% confidence intervals (and in downstream attributable fraction calculations). All models adjusted for PM_2.5, pre-specified socio-economic covariates, a province random intercept, and the population offset.

Subgroup analyses

We assessed effect modification by conducting subgroup analyses for several domains including the following: maternal education: low (no education or primary) versus high (secondary or higher), wealth index: low (poorest, poorer, and middle) versus high (richer and richest), air quality: PM2.5 < 25 μg m⁻³ (‘fair’) versus PM2.5 ≥ 25 μg m⁻³ (‘poor/hazardous’), and region type: urban versus rural. To assess between-group heterogeneity within each domain, we fit a single interaction model using the same distributed lag non-linear model specification as the main analysis. The model included cb (temperature) × subgroup plus covariates — PM_2.5 (quadratic), a cubic polynomial in year; we included education and wealth index except when they defined the subgroup (to avoid collinearity). We used treatment contrasts and obtained subgroup-specific log risks at the group-specific 90th and 99th percentiles of temperature (centred at the overall median) using the linear combinations of the main and interaction distributed-lag non-linear model coefficients, propagating the full model variance–covariance to form 95% confidence intervals. We tested between-group heterogeneity at each percentile with multivariate Wald tests (general linear hypotheses) built from covariance-aware contrasts (reported as p-het).

Temperature-related population attributable fractions

We estimated the population attributable fraction of low-birth-weight cases attributable to temperature exposure under current and future climate scenarios. For the baseline (2008–2017), we used observed mean temperature data from ERA5-Land. For future projections (2048–2057 and 2068–2077), we applied province-level projected mean temperatures obtained from the World Bank’s Climate Change Knowledge Portal (climateknowledgeportal.worldbank.org). The scenarios included Shared Socio-economic Pathways (SSP) SSP2-4.5 (‘intermediate’ greenhouse-gas emissions) and SSP5-8.5 (‘very high’ greenhouse-gas emissions) [54]. We used SSP2-4.5 and SSP5-8.5 to bracket an intermediate and a high-emission future, reflecting widely used coupled model intercomparison project (CMIP) 6 pathways that span policy-relevant uncertainty.

For each province and scenario, we derived temperature-response functions from our distributed-lag non-linear model estimates from the main analysis. We defined a heat–cold split using the province-specific baseline median temperature (2008–2017) and classified ‘heat exposure’ as months above this temperature threshold. As a sensitivity analysis, we repeated the risk estimation and population attributable fraction calculations using the province-specific minimum mortality temperature as the threshold.

We calculated population attributable fractions as the mean excess fraction among exposed months:

$$\mathit{100}\times {\mathrm{mean}}_{d\in \epsilon }[(r_{\mathrm{dln}}-1)/r_{\mathrm{dln}}]\times P$$ 1

where r_dln = the relative risk derived from distributed-lag non-linear models for month d, ε = the set of exposed months, and P =|ε|/(all months) = the proportion of months classified as heat exposure. We applied Monte Carlo simulations (1000 iterations) to incorporate uncertainty in risk estimates. We calculated the confidence limits for the 2.5th and 97.5th percentiles of the simulated attributable fractions.

Heat vulnerability index

We developed a district-level heat vulnerability index with three standard pillars, including exposure, sensitivity, and adaptive capacity, and combined them into a single composite score [55]. We included the district-level average mean temperature for heat exposure. We used mean temperature to ensure internal consistency with the exposure metric used in the models and to avoid redundancy or double counting with humidity in the broader modelling framework.

We included the average PM_2.5 concentrations over the entire study period as a co-exposure — a known environmental stressor that exacerbates vulnerability during heat events [56]. We accounted for sensitivity in terms of socio-economic deprivation. We used the multidimensional poverty index, which reflects socio-economic vulnerabilities such as limited access to healthcare, education, and inadequate living standards [57]. To capture pre-existing child health status and the effectiveness of local health systems, which are the main components of both population sensitivity and adaptive capacity, we included the under-5 mortality rate. Under-5 mortality reflects the joint influence of environmental exposures, socio-economic disadvantage, and health-system performance. We therefore interpret this index as a summary marker of baseline health and health-system capacity rather than an exposure metric per se, with higher mortality indicating greater underlying vulnerability and more limited adaptive capacity [58].

To capture empirical relationships between heat exposure and health outcomes, we incorporated province-specific relative risk estimates at the 99th percentile of temperature from the distributed-lag non-linear models as susceptibility multipliers. To account for model uncertainty, we used the confidence intervals of these estimates, ensuring a robust representation of risk across regions.

To ensure comparability across variables, we scaled and centred the variables (mean = 0, standard deviation = 1). We then formed a baseline composite score as an equal-weighted sum of z-scores for four components: mean temperature, PM_2.5, under-5 mortality, and multidimensional poverty (weights = 0.25 each). This approach ensures comparability and equal contribution.

We employed Monte Carlo simulations (1000 iterations) to account for uncertainty in the province-level risk estimates and vulnerability indices. Uncertainty in risk was propagated on the log scale, with the standard deviation (σ_log) derived from the width of the confidence intervals using the formula:

$${\sigma }_{\text{log}}=\frac{{\text{log}}_e(u)-{\text{log}}_e(l)}{2z}$$

where z = critical value for a 95% confidence interval (z = qnorm (0.975)≈1.96), with u = the upper bound, and l = the lower bound. For each iteration k, we resampled the relative risk (r_resamp) as follows:

$$r_{\text{resamp}}=e^{N\left({\text{log}}_e,\left(r_{\text{obs}}\right),,,{\sigma }_{\text{log}}^2\right)}$$

where r_obs = observed relative risk. For each iteration, we multiplied the baseline composite by a province-matched resampled relative risk to obtain the simulated heat vulnerability index and then summarised the mean and the 95% uncertainty intervals (2.5th–97.5th percentiles).

We calculated the baseline heat vulnerability index for district j as follows:

$$H_j={\sum }_{i=1}^k\left(w_i,x_{\mathit{ij}}\right)$$ 2

where w_i = 0.25 = weight assigned to variable i, x_ij = scaled and centred value of variable i for district j (mean = 0, standard deviation = 1), and k = number of variables included in the calculation of H. The susceptibility-adjusted index is as follows:

$$H_j^{\ast }=H_j{r}_j$$

where r_j is the province-matched relative risk (or its Monte-Carlo resample in uncertainty analyses).

We computed a heat vulnerability index for future scenarios (SSP2-4.5 and SSP5-8.5: 2048–2057 and 2068–2077). We used projected mean temperature and z-scored it using the baseline mean temperature and standard deviation to retain comparability over time. We held non-climate components (PM_2.5, under-5 mortality, multidimensional poverty) at their baseline values (i.e. no future change assumed). We then applied the same equal weights (0.25 each) and province-matched projected relative risk integration (multiplicative), yielding a projected district-level heat vulnerability index.

Software and tools

We did all data preprocessing and analyses in R (version 4.1.0, R Core Team 2024) with the dlm [59], glmmTB [60], and MuMIn [61] packages and visualisation with ggplot2 [62]. The analysis code is available at 10.5281/zenodo.14373384.

Results

After cleaning and preprocessing the data, we retained 85,017 records from mothers who provided birth weight or size information for their most recent child born within the last 5 years of the study period. Among these, 15,920 children (18.7%) had low birth weight (≤ 2.5 kg) or were reported as smaller than average in size at birth. By extrapolating to the national level and using an average birth rate of 29.69 per 1000 population (from 2008 to 2017) and the annual total population estimates [23], we calculated that approximately 60.15 million babies were born in Pakistan between 2008 and 2017 (mean = 6.01 million per year), of which 18.72% (~ 11.26 million) had low birth weight.

Nearly three-quarters (74.65%) of participants had primary or no education, 13.47% had completed secondary education, and 11.87% had attained higher education (Table 1). About one-third (29.01%) of participants belonged to the ‘poorer/poorest’ wealth category, 12.57% were in the middle-income group, and 23.79% were classified as ‘richer/richest’ (Table 1). The median temperature during the study period was 19.5 °C, and the median air pollution concentration was 44.8 μg m⁻³ (Table 1).

Table 1.

Prevalence of low birth weight and distribution of sociodemographic and environmental variables in Pakistan from 2008 to 2017 (N = 85,017)

Demographic and socioeconomic variables
Variable	%	Variable	%
Low birth weight (n = 15,920)	18·72
Education		Mother age
Primary or no education	74·65	15–29	59·37
Secondary	13·47	30–39	35·53
Higher	11·87	40–49	5·11
Wealth index		Province
Poor	29·01	Baluchistan	3·95
Middle	12·57	Gilgit Baltistan	4·98
Rich	23·79	Khyber Pakhtunkhwa (including FATA)	13·02
Region		Punjab	62·31
Rural	66·69	Sindh	15·74
Urban	33·31
Environmental variables
Variables	Median (range)	Variable	Median (range)
Mean temperature (°C)	19·5 (− 20·0–35·1)	Humidity (%)	54·3 (18·9–80·2)
Air pollution PM2·5 (μg m−3)	44·8 (4·5–123·9)	Precipitation (mm)	46·60 (0·03–427)

Note: Percentages may not sum to 100 due to rounding and/or missing data

The graphical analysis of monthly trends in low birth weight, mean temperature, precipitation, humidity, and air pollution (PM_2.5) revealed that cases of low birth weight are higher during hotter months when temperatures, humidity, and precipitation are highest (Fig. 1).

Fig. 1.

Monthly province-level trends of a cases of low birth weight along with average mean temperature, b average precipitation, and c average particulate matter (PM_2.5)

Province-level exposure–response association between temperature and low birth weight

We observed a strong positive association between low birth weight and heat exposure across all provinces (Fig. 2). The relative risk estimates ranged from 1.37 (1.04–1.81, 95% confidence intervals) to 1.59 (1.16–2.17) at moderate heat (90th percentile) and from 1.47 (1.07–2.03) to 1.91 (1.24–2.93) at extreme heat (99th percentile), compared to the median reference temperature (Fig. 2). In contrast, we found no evidence of a relationship between the risk of low birth weight and cold exposure for all provinces except Gilgit Baltistan. In our findings, heat exposure had a stronger impact; therefore, the results and discussions primarily emphasise the effects of heat. We present the cumulative exposure–response associations for all candidate models in Additional File 1: Fig. S5 and province-specific risk estimates for all provinces in Additional File 1: Table S5.

Fig. 2.

Exposure–response association between counts of low birth weight and mean temperature (with 95% confidence intervals). Arrows indicate the minimum risk point, while province-level letter symbols (B, Baluchistan; K, Khyber Pakhtunkhwa; P, Punjab; S, Sindh) and vertical lines represent the 90th and 99th percentiles for relative risk estimation

Our stratified subgroup analysis indicates that poor air quality (≥ 25 μg m⁻³) exacerbates the risk of low birth weight during hot conditions: 2.22 (1.28–3.85) at the 90th percentile and 2.77 (1.37–5.60) at the 99th percentile, with between-category heterogeneity at the 99th (p-het = 0.0083). For education, women with primary or no education had higher heat-related risk: 1.78 (1.16–2.74) at the 90th percentile and 2.12 (1.22–3.66) at the 99th percentile. There was no statistical evidence for heterogeneity (p-het = 0.459 and p-het = 0.560, respectively) (Table 2).

Table 2.

Subgroup-stratified relative risk estimates (with 95% confidence intervals) of low birth weight at the 90th and 99th percentiles of mean temperature

Subgroup	Risk estimates (90th percentile)	Risk estimates (99th percentile)	p-het* (90%)	p-het* (99%)
Air quality status
Fair (< 25 μg m−3)	6.46 (0.35–118.25)	0.72 (0.42–1.22)	0.457	0.0083
Poor–hazardous (≥ 25 μg m−3)	2.22 (1.28–3.85)	2.77 (1.37–5.60)
Education status
Primary or no education	1.78 (1.16–2.74)	2.12 (1.22–3.66)	0.459	0.560
Secondary or above	1.24 (0.46–3.31)	1.40 (0.34–5.73)
Wealth status
Low	2.44 (1.61–3.71)	3.38 (1.93–5.95)	0.001	0.001
High	1.02 (0.58–1.79)	1.04 (0.51–2.15)
Type of region
Urban	2.41 (1.11–5.19)	2.90 (1.16–7.28)	0.257	0.278
Rural	1.58 (1.04–2.41)	1.79 (1.07–3.02)

^*p-het is the type 1 error for heterogeneity

For wealth, women from both low-wealth groups had a high risk of low birth weight linked to extreme heat, 2.44 (1.61–3.71) and 3.38 (1.93–5.95), respectively. There was statistical evidence for heterogeneity (p-het = 0.001). By region, risks were raised in both urban areas: 2.41 (1.11–5.19) and 2.90 (1.16–7.28) and in rural areas: 1.58 (1.04–2.41) and 1.79 (1.07–3.02), with no evidence of heterogeneity (p-het = 0.257 and 0.278, respectively). Estimates with wide confidence intervals (e.g. fair air quality at the 90th percentile: 6.46 (0.35–118.25)) should be interpreted cautiously (Table 2). Results from additional sensitivity analysis based on different modelling choices are presented in Additional File 1: Table S6. Consistent estimates and AIC indicate model robustness and support the modelling choices.

Current and projected population attributable fraction of heat- and cold-related low-birth-weight cases

The analysis of the population attributable fraction revealed regional variation in temperature-related impacts on low birth weight under current and future climate scenarios (Fig. 3, Additional File 1: Table S7). In Baluchistan, the heat-related attributable fraction increased from 11.83% (4.41–18.17) during the baseline period to 20.26% (8.44–30.29) under SSP5-8.5 for 2068–2077, representing an increase of 8.43% by the 2060 s under a high-emissions scenario. Similarly, Khyber Pakhtunkhwa experienced an increase in the heat-related population attributable fraction from 9.39% (1.63–16.33) at baseline to 16.25% (3.50–26.57) under SSP5-8.5 for 2068–2077, showing an increment of 6.86% by the 2060s.

Fig. 3.

Heat-related population attributable fraction (%) with 95% confidence intervals by province under SSP2-4.5 (upper panel) and SSP5-8.5 (lower panel) for the baseline (2008-2017) and future periods (2048-2057 and 2068-2077)

In Punjab, the heat-related population attributable fraction rose from 13.15% (4.90–19.59) at baseline to 21.87 (9.49–31.73) under SSP5-8.5 for 2068–2077, an increase of 8.72% by the 2060s. In Sindh, the heat-related population attributable fraction increased from 8.02% (3.17–12.56) at baseline to 18.22% (6.92–26.75) under SSP5-8.5 for 2068–2077 (increasing by 10.20% by the 2060 s). These findings underscore the intensifying impacts of heat under high-emissions scenarios, highlighting regional disparities in vulnerability to temperature-related risks on low birth weight. As a robustness check, we re-estimated attributable fractions using province-specific minimum-risk temperature as the reference; results were directionally consistent and are reported in Additional File 1: Table S8.

District-level heat vulnerability index

The heat vulnerability index map (Fig. 4) illustrates the spatial distribution of heat vulnerability for low-birth-weight outcomes across Pakistan’s provinces and districts. The map highlights areas with varying vulnerability to heat-related impacts, ranging from low to high. Districts in southern Punjab, northern Baluchistan, and Sindh exhibit the highest vulnerability, indicating a greater risk of heat-related low birth weight in these regions. These areas are likely influenced by higher heat exposure and compounding socio-economic and environmental conditions, such as poverty and air pollution. In contrast, northern regions such as Gilgit Baltistan and parts of Khyber Pakhtunkhwa show lower vulnerability due to their cooler climates. Projected heat vulnerability indices are presented in Additional File 1: Fig. S6.

Fig. 4.

District-level spatial patterns of the heat vulnerability index for heat-related low birth weight

Discussion

Our study highlights the impact of heat exposure on low birth weight in Pakistan, emphasising the health implications of rising temperatures and compounded vulnerabilities in resource-limited settings. Pakistan stands out for having some of the worst maternal and child-health outcomes among low- and middle-income countries with similar socio-economic status [63, 64]. Climate change and extreme temperature conditions further exacerbate this vulnerability. Dimitrova et al. [65] highlighted that Pakistan (along with Mali, Sierra Leone, and Nigeria) experiences a disproportionately high burden of neonatal mortality linked to temperature fluctuations, with > 160 temperature-related neonatal deaths per 100,000 live births. Similarly, Zhu et al. [24] estimated the temperature-related burden of low birth weight across several countries, reporting the greatest reduction in birth weight — − 257 g (− 498.06 to − 17.15 g) — among children in Pakistan. These statistics underscore the intersection of climate vulnerability and systemic healthcare challenges in Pakistan, highlighting that heat exposure is not just an environmental issue but also a social determinant of health that disproportionately impacts marginalised populations.

In one of our recently published studies [66], we assessed the impacts of temperature on low birth weight across pregnancy months leveraging a subset of the same data sources and identified lag 1 as a salient exposure window, with considerable uncertainty in other lags. In contrast, in this study, we focused on subnational, province–month estimates, harmonises measured birthweight and maternal-reported size at birth, and employed an updated analytic pipeline with improved preprocessing and advanced methods. Consistent with this shift in scope and methods, effect estimates are more precise and robust across sensitivity analyses (including a weight-only analysis, see supplementary material).

Our study builds on this evidence, providing detailed insights into the risk of low birth weight associated with heat (~ 30–70%), identifying contributing factors, pinpointing high-risk areas, and estimating future population attributable fractions. This localised analysis informs targeted interventions and adaptation strategies for regions facing severe climate-related health challenges.

We found that women with lower education and from lower-wealth households experienced greater heat-related risks, consistent with prior work [67]. This heightened vulnerability reflects multiple, overlapping disadvantages: poorer housing quality and limited access to cooling, higher exposure in outdoor or physically demanding occupations, constraints on healthcare access and prenatal care, and greater sensitivity to power outages that reduce the effectiveness of heat-mitigation strategies [68]. These pathways align with evidence that socio-economic deprivation amplifies heat impacts on maternal and newborn health [67, 69].

We identified that Punjab appears to be the most vulnerable to heat-related low birth weight. The heat vulnerability index revealed that districts in lower Punjab, southern Sindh, and northern Baluchistan are the most at risk. These regions are among the hottest in Pakistan and face substantial risks from climate change. The spatial analysis underscores the importance of prioritising these areas for public health interventions, providing local health authorities with scientific evidence to inform resource allocation and health policies. We estimated that 9.39–13.15% of low-birth-weight cases in Pakistan were attributable to hot conditions across various provinces. Based on this estimate, if approximately 11.26 million children were born with low birth weight in Pakistan from 2008 to 2017, an average of ~ 1.24 million cases can be attributed to hot weather conditions. These estimates are projected to increase by ~ 8–10% in high-emission scenarios by the 2060s. The estimates suggest that high temperatures are a major threat to progress in maternal health and could widen social and health inequalities in low- and middle-income countries.

We did not observe an association between cold exposure and low birth weight, except for Gilgit-Baltistan, a region with a harsh tundra climate and severe winters. The area’s high altitude, extreme cold, and limited healthcare access during the winter likely contribute to this finding. Further research is needed to understand the combined impacts of heat, glacier melting, flooding, and landslides in these terrains and their interaction with maternal and child health to guide adaptation strategies.

Pakistan’s vulnerability is intensified by its rapidly growing population, a young average age, and cultural barriers limiting women’s autonomy and access to prenatal care. Political instability, civil unrest, and periodic terrorism further strain the country’s capacity to prioritise public health. Limited healthcare infrastructure and research resources hinder the development of adaptive strategies to address climate-related health impacts. Together, these factors create a backdrop of socio-economic fragility, heightening the vulnerability of pregnant women and newborns to adverse health outcomes, including low birth weight. Addressing the intersection of climate change and maternal health requires comprehensive strategies that integrate climate adaptation and mitigation with strengthening health systems, promoting gender equality, and building community resilience. Collaborative efforts at the global, regional, and national scales are essential to mitigate the impact of climate change on maternal health and to achieve the sustainable development goals.

Educational programmes should inform pregnant women about heat risks and mitigation strategies, while the government must improve access to prenatal check-ups and maternal healthcare services. Previous studies have explicitly identified education as an important protective factor in protecting young children from the impacts of climate change [70]. Investments in prenatal and postnatal care in high-risk areas are necessary for enhancing maternal and child health. Tailored early warning systems can alert pregnant women and healthcare providers to heatwaves through mobile networks, local media, and antenatal care centres. Lady health workers and non-governmental organisations should promote heat adaptation, hydration, and support access to cooling facilities. Collaborations with the Pakistan Meteorological Department and telecommunication companies can extend outreach via mobile apps, voice messages, and community networks.

As climate change progresses, it will be necessary to develop and test adaptation strategies to protect pregnant women in urban and rural areas from heat-related health risks. Continuous monitoring and evaluation of these strategies will be essential to ensure their effectiveness over time. With the heat vulnerability maps developed, high-risk areas can be specifically targeted for interventions, optimising the allocation of resources and improving outcomes for the most vulnerable populations.

Limitations

We used data from Pakistan’s Demographic and Health Surveys and Multiple Indicator Cluster Surveys, representative of national maternal and child health indicators. However, these datasets have limitations:

• (i)The lack of exact childbirth dates required monthly data aggregation, which might de-emphasise short-term temperature peaks, even though the approach effectively captures broader temporal patterns. Aggregated data also enable reliable estimates of long-term burden, aligning with the principles of Basagaña and Ballester [71].
• (ii)Missing spatial coordinates required aggregation at monthly and provincial scales, accounting for regional variation while addressing data sparsity. To complement this, we developed a district-level heat vulnerability index to identify high-risk areas for targeted interventions.
• (iii)We addressed inherent variability and gaps in survey data using weighted coefficients from multiple models and random intercepts for provinces.
• (iv)Birth size, a crude measure used alongside birth weight, can introduce biases.
• (v)We excluded data from the 2017–2018 Pakistan Demographic and Health Surveys to avoid redundancy and overlap with the 2018 Multiple Indicator Cluster Surveys.
• (vi)Gridded meteorological data served as a proxy for exposure, capturing local trends but lacking individual-level spatial and temporal precision. Monthly averages might obscure variability and household factors influencing low birth weight.
• (vii)Projected heat vulnerability indices were derived by chaining projected heat risk (relative risk) and projected mean temperature onto a baseline heat vulnerability index in which non-climatic components (e.g. multidimensional poverty index, under-5 child mortality rate, particulate matter) were held constant; therefore, projections only reflect climate-driven changes in risk.

The study’s aggregated nature limits control over individual-level confounders, so we encourage caution when interpreting the relationships observed as causal. Including spatial coordinates and precise childbirth dates in Pakistan’s Multiple Indicator Cluster Surveys, along with robust maternal and child health data systems, is essential. Improved data collection will enhance understanding of environmental impacts on health and support more targeted interventions.

Finally, we used district-level under-5 mortality as a proxy for baseline health and health-system capacity in our heat vulnerability index. We acknowledge that under-5 mortality is shaped by the combined effects of environmental exposures, socio-economic conditions, and service performance, and therefore does not represent a ‘pure’ measure of adaptive capacity. Our framing treats it as a pragmatic summary indicator of underlying health vulnerability at the district level, but this conceptual overlap should be considered when interpreting the index.

Conclusions

Pregnant women in Pakistan face a heightened risk of delivering infants of low birth weight following exposure to extreme temperatures, with percent excess risk varying among provinces from ~ 30% to ~ 70% at high percentiles. An estimated 9.39–13.51% of low-birth-weight cases are attributable to heat, and this burden is projected to increase by ~ 8–10% by the 2060 s under high emission scenarios. Regions in southern Punjab, northern Sindh, and Baluchistan have the highest risks.

Addressing this issue requires the development, implementation, and rigorous evaluation of comprehensive intervention strategies. These approaches must be multifaceted, leveraging collaboration between researchers, government agencies, and local communities. Enhancing and adapting existing strategies will be essential for improving health outcomes and protecting maternal and child health in the future. The intersection of climate change, socio-economic disparities, and environmental degradation poses a serious threat to maternal and neonatal health in Pakistan and other low- and middle-income countries. Therefore, interventions designed to mitigate the effects of climate change, enhance access to essential healthcare services, and promote sustainable development are urgently required.

Interventions should include (i) research and monitoring that continuously track temperature trends and their impact on maternal and infant health to inform decision-making; (ii) development of policies that prioritise maternal and child health in the context of climate change, ensuring they are inclusive and adaptable to local needs; (iii) public health interventions that include community-based programmes to educate and support pregnant women in coping with extreme heat events; (iv) infrastructure improvements to strengthen healthcare facilities with climate-resilient infrastructure to maintain functionality during extreme weather events; (v) community engagement to involve local communities in developing and implementing interventions, ensuring their insights and needs guide the process; and (vi) resource allocation to secure sufficient funding and resources to support these initiatives with a focus on long-term sustainability.

Abbreviations

AIC

Akaike information criterion

ΔAIC

Difference in AIC between the top-ranked and the current model

wAIC

Akaike weight (model weight based on AIC)

CB

Cross-basis term (as in ‘cb (temperature) × subgroup’)

CMIP 6/CMIP 6

Coupled Model Intercomparison Project Phase 6

DAG

Directed acyclic graph

DAGitty

Tool/R package for DAG-based adjustment set identification

DLNM

Distributed lag non-linear model

ERA5-Land

Copernicus/ECMWF ERA5-Land reanalysis dataset

GHDx

Global Health Data Exchange

HVI

Heat Vulnerability Index

ΔHVI

Change in Heat Vulnerability Index (relative to baseline)

KPK

Khyber Pakhtunkhwa

MICS

Multiple Indicator Cluster Surveys

MICS-4/MICS-5/MICS-6

MICS survey rounds 4, 5, and 6

MPI

Multidimensional Poverty Index

MuMIn

R package (model selection/model averaging utilities)

NA

Not available/not applicable

n

Number of observations

PDHS

Pakistan Demographic and Health Survey

PM2.5

Fine particulate matter (≤ 2.5 μm)

R

Pearson correlation coefficient

SES

Socioeconomic status

SSP

Shared socioeconomic pathways

SSP2-4.5

Intermediate emissions scenario

SSP5-8.5

Very high emissions scenario

Tmean

Mean temperature

UN

United Nations

UNDP

United Nations Development Programme

UNICEF

United Nations Children’s Fund

Authors’ contributions

SHF: Conceptualisation, Data Curation, Formal Analysis, Methodology, Visualisation, Project Administration, Writing-original draft preparation. CJAB: Formal Analysis, Methodology, Visualisation, Writing-reviewing and editing. ZAB: Validation, Writing-reviewing and editing. PB: Validation, Writing-reviewing and editing. JKD: Validation, Writing-reviewing and editing. SM: Conceptualisation, Validation, Writing-reviewing and editing. ZSL: Conceptualisation, Validation, Funding Acquisition, Writing-reviewing and editing. All authors read and approved the final manuscript.

Authors’ social media handles

BlueSky: @hira2019.bsky.social (Syeda Hira Fatima).

BlueSky: @conservbytes.bsky.social (Corey J. A. Bradshaw).

X: @Zohralassi (Zohra Lassi).

Funding

This work was supported by the Australian National Health and Medical Research Council Investigator Grant (2009730) awarded to Z. Lassi. The funder had no role in the design of the study; in the collection, analysis, or interpretation of data; or in writing the manuscript.

Data availability

The analysis code used to preprocess the data and fit all models is openly available in the Zenodo repository at 10.5281/zenodo.18397130. The individual-level survey microdata underpinning this analysis are third-party, de-identified data obtained from the Demographic and Health Surveys (DHS) Program [72, 73] and UNICEF Multiple Indicator Cluster Surveys (MICS) [74–83] under data use agreements that do not permit the authors to share the raw datasets. The specific PDHS and MICS datasets analysed are cited in the reference list [72–83]. These data are freely available to other researchers upon registration and approval of a data request through the DHS Program (www.dhsprogram.com) and UNICEF MICS (mics.unicef.org), in line with their standard access procedures.

Declarations

Ethics approval and consent to participate

Ethical approval for secondary analysis of already existing, de-identified survey microdata was granted by the Flinders University Human Research Ethics Committee (Project ID 8026, Application Ref.: HEG8026-2). As this study used anonymised secondary data with no participant contact, no additional consent was required for this analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.