Associations between multiple perinatal exposures and risk of childhood hospitalisation with infection: a registry-based study in two countries

Two separate registry-based cohorts were assembled from Norwegian and Danish administrative health data. The cohorts were identified from the Medical Birth Register of Norway and Denmark and included all liveborn singleton children born from 1 st January 2008 through 31 st December 2017 in Norway and from 1 st November 1997 through 31 st December 2018 in Denmark. We excluded multiple births, children with gestational age < 24 or > 43 weeks, birth weight < 500 or > 5500 g, children with congenital malformations recorded in the birth registries, and pregnancies where information on maternal smoking was not available (Denmark n = 26738, Norway n = 67816). In the Norwegian data, emigration dates were not available, and therefore, we additionally excluded children who emigrated from the study population (n = 15898). In the Danish data, maternal emigration dates were used as a proxy for censoring due to emigration. Birth records were linked to national hospital and prescription registries for both countries in the relevant periods.

We included seven perinatal exposures for which there was prior evidence of association with infection in offspring or relatedness with other exposures [4‐15]. Start of pregnancy was calculated as the birth date minus the estimated gestational age, which is usually based on ultrasound estimates. For each exposure, a binary variable was created as follows: maternal dispensed prescriptions for antibiotics during pregnancy (any/none), maternal smoking during pregnancy (any/none), maternal diabetes (pre-gestational type 1 or 2 or gestational diabetes versus no diabetes), hypertensive disorders during pregnancy (any of pre-pregnancy hypertension, gestational hypertension, pre-eclampsia, eclampsia, HELLP versus none), mode of birth (caesarean versus vaginal), preterm birth (< 37 versus ≥37 weeks gestational age), and small for gestational age (SGA) (≤10th percentile versus > 10th percentile of birthweight for specific gestational age and sex). Further details on the diagnosis/procedure/prescription codes and registries used to create each exposure variable is presented in Supplementary Table 1.

Associations between perinatal exposures and infection-related hospitalisation were assessed using hazard ratios estimated from Cox proportional hazards regression models with child age as the time scale. This model was a Prentice–Williams–Peterson recurrent events model considering up to three separate infection-related hospitalisations [16]. The Prentice–Williams–Peterson model stratifies by event number (e.g., first, second, third hospitalised infection), estimating a different baseline hazard for each event number. This allows for the risk to vary according to event number which is plausible for our outcome of hospitalised infections. We used the total-time method of the Prentice–Williams–Peterson model. For our main analysis, we included infection-related hospitalisations occurring before 5 years of age to capture the peak period of infection-related hospitalisations in children. Censoring events in the analysis were death, emigration (available in Denmark only) or the end of follow-up (5 years of age or end of the study period, whichever occurred first). Two sensitivity analyses were conducted in which we evaluated all infections the first 2 years and 10 years of life, respectively. Exclusions due to missing exposure data (apart from pregnancy smoking) were minimal.

To evaluate the risks associated with the seven exposures individually and in combinations, we used four separate analytic approaches, basing our methodological approach on that of Nielsen et al. which also considered multiple co-occurring exposures [17]. All four approaches used the regression model as described above and included adjustment for background maternal and perinatal factors that were considered potential confounders, as identified from our causal directed acyclic graph (Supplementary Fig. 1): (1) maternal age, (2) parity, (3) year of birth, (4) season of birth, and as a proxy for socio-economic status (5a) maternal employment status (Denmark only), and (5b) maternal country of birth (Norway only). First, we considered the individual associations between each perinatal exposure and infection-related hospitalisation. Second, we examined the association between the cumulative number of adverse perinatal exposures (1, 2, 3, 4, 5+), where the group with no exposures was the reference category. Third, we estimated the association for each unique combination of the seven perinatal exposures, again where the group with no exposures were the reference category. This analysis was limited to exposure combinations with at least 100 children. In each analysis, separate estimates were generated for the Norwegian and Danish cohorts and these estimates were subsequently combined using fixed-effects meta-analysis. We chose a fixed-effects model as the two populations were relative similar in terms of data sources and the individual estimates. Finally, to examine possible interaction between exposures, we estimated interaction terms on the multiplicative and additive scales for each pairwise combination of perinatal exposures. For the multiplicative scale we report the interaction term as the ratio of hazard ratios (HR₁₁/HR₀₁HR₁₀) and for the additive scale we report the relative excess risk due to interaction (RERI) measure calculated using the InteractionR package [18]. We tested interactions on both scales as each may provide different insight into the joint effects of exposures [19], however we focused our main results to interaction terms on the multiplicative scale in line with our other analyses. The multiplicative interaction term indicates whether the effect of two exposures present together is greater (interaction term > 1) or lesser (interaction term < 1) than the expected joint effect calculated from the product of their individual effects.

Our analyses included 1,113,708 singleton children born in Denmark between 1997 and 2019 and 470,270 born in Norway between 2008 and 2018, giving a total analytical sample of 1,583,978. A total of 23.6% and 21.5% of the study population in Denmark and Norway respectively had one or more infection-related hospitalisation before the age of five. The distribution of maternal and perinatal characteristics for the two countries are shown in Table 1. The proportion of children who had each perinatal exposure of interest were similar between the two countries, albeit with somewhat higher maternal antibiotic prescriptions during pregnancy in the Norwegian study population (27.3% vs. 24.4%, respectively), and a higher proportion of smoking during pregnancy in the Danish study population (15.6% vs. 8.9%, respectively).

Adjusted estimates of the individual risk for each exposure are shown in Fig. 1. All seven perinatal and pregnancy exposures were each associated with increased risk of infection-related hospitalisation. Similar estimates were observed across the two countries. In meta-analyses of hazard ratios across the two populations, the magnitude of risk was smallest for SGA (aHR 1.08, 95% CI 1.08–1.09); relatively similar across smoking during pregnancy (aHR 1.16, 95% CI 1.15–1.17), maternal diabetes (aHR 1.16, 95% CI 1.15–1.18), hypertensive disorders of pregnancy (aHR 1.17, 95% CI 1.16–1.18), prenatal antibiotics (aHR 1.19, 95% CI 1.18–1.20), and caesarean section (aHR 1.20, 95% CI 1.19–1.21); and highest in magnitude for preterm birth (aHR 1.45, 95% CI 1.43–1.46).

Bild vergrößern

Figure 2 shows the adjusted estimates from the analysis of infection-related hospitalisation according to the number of cumulative exposures. In the pooled analysis, a dose-response pattern was observed with the number exposures; the adjusted hazard ratio for hospitalised infections was 1.17 (95% CI 1.17 to 1.18) for one exposure compared to no exposures, and 1.88 (95% CI 1.78 to 1.99) for five or more exposures.

Bild vergrößern

The adjusted estimates for the association between each unique combination of the seven exposures and hospitalised infections are shown in Table 2. A total of 86 and 68 unique exposure combinations with 100 or more exposed children were analysed for Denmark and Norway respectively. Nearly all combinations of exposures were associated with increased risk of hospitalised infections. The results showed similar patterns to the analyses of individual exposures, where combinations including preterm birth constituted the greatest risk for hospitalised infection, with estimates ranging from 1.43 (95% CI 1.31–1.56) for preterm children also exposed to maternal diabetes, to 2.14 (95% CI 1.93–2.38) for preterm children also exposed to prenatal antibiotics, caesarean section, and born small for gestational age. The four highest risk combinations in meta-analyses were children with prenatal antibiotic exposure, born preterm and by caesarean section in combination with another exposure. In combinations of exposures not involving preterm birth, the largest hazard ratios were approximately 1.6 and included four-way combinations of the most common exposures (prenatal antibiotic exposure and caesarean section) and two other exposures.

Table 3 shows the multiplicative interaction terms estimated for each pairwise combination of exposures and Supplementary Table 5 shows the full interaction analyses. The most notable interaction terms– where we observed similar results across both countries and the 95% confidence intervals for the interaction term were not inclusive of the null– were: (1) maternal diabetes together with pregnancy hypertension, and (2) maternal diabetes together with preterm birth, where the risk associated with the exposures together was less than what would be expected from their individual risks on the multiplicative scale; (3) caesarean section together with preterm birth on the additive and multiplicative scale; and (4) SGA together with preterm birth on the additive scale, where the combined risks were greater than that expected from their individual risks.

Supplementary Tables 2–4 shows the results from sensitivity analyses where follow-up was until 2 years or 10 years of age rather than 5 years, which was the pre-specified period in the main analysis. The magnitude of risk tended to be slightly higher in analyses where follow-up was until 2 years and slightly lower in those until 10 years, but similar overall patterns to those in the main analyses were observed.