Search strategy and selection criteria

Our population of interest consists of paediatric patients with acute conditions and/or those treated in ICU or emergency units for acute conditions whose vitamin D status was assessed prior to or during admission. We included published cross-sectional, case–control and cohort studies that measured circulating 25(OH)D levels and either reported prevalence, ORs or data to enable calculation of these measures. Studies were excluded if they were reviews, case reports, surveys, commentaries, replies, not original contributions, experimental in vitro or if they recruited patients who were not treated in emergency, neonatal intensive care units (NICU), paediatric intensive care units (PICU) or for acute conditions. Studies were also excluded if they only enrolled vitamin D deficient patients, investigated healthy populations only or did not measure circulating 25(OH)D levels as an indicator of vitamin D status. When we identified more than one publication using the same cohort, we included the publication which shared our review’s objective to investigate vitamin D levels and prevalence of deficiency.

For purposes of our review, we classified vitamin D deficiency as being 25(OH)D less than 50 nmol/L (equivalent to 20 ng/mL), as suggested by the IOM.16 Different age categories were used to designate patients as ‘children’ in the studies reviewed. We therefore included all ‘children’ (neonates up to 21 years) as defined by each treating facility and this included ‘neonates’, ‘infants’, ‘toddlers’, ‘children’ and ‘adolescents’.

We searched PubMed, OVID, Google Scholar and the Cochrane Library from inception up until 21 December 2018, with no language restrictions. Search terms used across these databases included: ‘critical care’, ‘vitamin D’, ‘pediatric’, ‘child’, ‘neonate’, ‘toddler’, ‘intensive care unit’, ‘sepsis’ and ‘septic shock’. Search terms used in OVID and PubMed are listed in the online supplementary table 2A,B. Literature searches were performed by two investigators independently (MC and AJBT) and included initial screening of titles and abstracts, followed by full-text screening. Any disagreements for study eligibility were resolved by discussion between the two investigators. Reference lists of the selected papers, including reviews, were also checked for relevant titles. Abstracts of relevant titles were then assessed for eligibility. Corresponding authors were contacted to obtain additional information if necessary. A data extraction form was designed a priori in Excel. Variables extracted from each study included year of publication, country of study, clinical setting, cut-off given to define vitamin D deficiency, total number of children, total number of cases, study design and age range.

Study quality assessment

The quality of each included study was assessed using the Newcastle-Ottawa Scale (NOS) adapted a priori for this review, for cohort, case–control and cross-sectional study designs (online supplementary table 3A–C).29 We classified studies as low (1–3), medium (4–6) or high quality (7–9) for purposes of sensitivity analysis.

Prevalence and mortality outcomes

In the majority of studies (n=40), prevalence of vitamin D deficiency was extracted as reported with a threshold of ≤50 nmol/L. If prevalence was not reported directly, it was calculated using data provided in each study (cases ≤50 nmol/L per total number of study participants) (online supplementary table 4A,B). Extracted or calculated prevalence values were then combined in a meta-analysis. For mortality, we calculated unadjusted ORs as:

OR = (vitamin D deficient patients who died * vitamin D non-deficient patients who did not die)/ (vitamin D deficient patients who did not die * vitamin D non-deficient patients who died)

We had sufficient information to calculate ORs <50 nmol/L for 40 studies (77.0%). For the 12 studies with insufficient information, we used the lower cut-off values reported as a conservative approximation (online supplementary table 5). We converted 25(OH)D values using: nmol/L=ng/mL * 2.496.

Data analysis

We obtained proportions of vitamin D deficiency with 95% CIs using the Clopper-Pearson method30 in R. We used a random-effects model31 to account for the variation observed within and between studies due to the different ages and acute conditions in the populations considered. For each meta-analysis we also obtained the 95% prediction interval (PI) to further account for between-study heterogeneity. This helps to evaluate how consistent an observed effect would be in a future study that will investigate the same association.32 We obtained pooled proportions and pooled ORs with fixed-effects model for sensitivity analysis or in cases where heterogeneity was low.33–35 For prevalence we also calculated median and IQR for comparisons with pooled prevalence estimates.

We investigated possible sources of heterogeneity using sensitivity and subgroup analyses. Cochran’s Q was used to assess the heterogeneity and the I² statistic was used to estimate the percentage of total variation across studies which can be attributed to heterogeneity. CIs of I² were calculated to aid interpretation.36 A Q value of <0.05 was considered significant and an I² statistic greater or equal to 75% indicated a high level of variation due to heterogeneity.37 38 We used Egger’s regression test to present results of small-study effects and funnel plot asymmetry39 and generated funnel plots for visual assessment and screening. A p value <0.05 indicated evidence of small-study effects. With few studies, Egger’s test has low power to detect such bias, therefore we only estimated small-study effects for analyses with more than 10 studies.40 When small-study effects were detected based on this threshold, we used trim-and-fill methods to add potentially missed studies and recalculate an adjusted pooled estimate.41

To further assess heterogeneity, we used meta-regression to identify predictor variables that could explain variation in study prevalence estimates. We used restricted maximum likelihood estimations in the model to account for residual heterogeneity42 and the Knapp-Hartung method to adjust CIs and test statistics. This method estimates between study variance using a t-distribution, rather than a z-distribution, yielding a more conservative inference.43 We tested the following continuous predictors: year of study publication, total sample size and quality score. Categorical variables included study setting (PICU, NICU), study design (case–control, cross-sectional and cohort) and country group by geographic region and economic development (group 1, group 2 and group 3) and were dummy coded.

We used R V.3.5.0 and Microsoft Excel 2010 for analyses and data collection. The R packages ‘meta’44 and ‘metafor’45 were used for analyses. Only results of the random-effects model are reported for prevalence due to the expected heterogeneity between populations being considered. Our protocol is registered in PROSPERO (CRD42016050638).

Role of the funding source

The study received funding from the UK Medical Research Council. The funders had no role in data collection, analysis, interpretation or writing of the report. All authors had access to the data in the study.

Patient involvement

No patients were involved in this study. We only used data from previously published studies.

Screening and study characteristics

After title and abstract screening, we identified 2890 potentially relevant studies (figure 1) and 85 full-text articles were assessed for eligibility. Rationale for study exclusion included: studies including adults, study populations other than critically ill children or with acute conditions and studies of circulating vitamin D levels and deficiency in healthy children or in children with chronic conditions. Four studies46–49 were excluded due to insufficient data reporting (online supplementary table 6). We also excluded three studies50–52 that used the same cohort of children and included a single study to represent the cohort.53 Ultimately, 52 studies met criteria for inclusion (online supplementary table 7).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Flow chart of study selection process. ICU, intensive care unit.

The primary objective of most included studies was to determine circulating vitamin D concentration (‘status’) in children and/or prevalence of vitamin D deficiency. Secondary objectives included investigation of associations between deficiency of circulating vitamin D and various outcomes, such as hospital mortality length of stay, requirement of ventilation and/or illness severity (online supplementary table 8).

All included studies reported vitamin D measurement assay methods used (online supplementary table 9) and stated that samples were collected and analysed within the first 24 hours of hospital admission. Studies reported ethical approval and consent for participation from parents or guardians (online supplementary table 10). Included studies were published between 2004 and 2018, with the majority (n=40, 77.0%) published between 2014 and 2017 (online supplementary table 7). In total, 7434 children were hospitalised in PICU or NICU or emergency units or for acute conditions. Sample sizes of critically ill children ranged from 2554 to 1016.55 In 18 studies the total number of cases was greater than 100.

Studies originated from 15 countries, with the majority from India8 56–65 (n=11) or Turkey54 66–71 (n=7) (online supplementary table 7). All were of medium or high quality (NOS score median 7, range 4–8). The score range for cohort studies was 6–8 (n=30), for case–control studies 5–8 (n=18) and for cross sectional 4–6 (n=4). Studies used a broad range of ages to classify patients as ‘children’. Seven studies (13.5%)54 65 67 69–72 included only neonates. In two67 72 of these studies, neonates were preterm. The largest age range was seen in the study of Ayulo et al 73, which included individuals between 1 and 21 years of age (online supplementary table 11). Forty-two of the included studies (80.8%) included patients admitted for medical conditions and the other 1053 61 66 74–79 included both surgical and medical patients. Of the 52 included studies 26 used a control group and had a total number of 2479 controls of which 773 (31.2%) were vitamin D deficient.

All studies included both female and male participants. For mortality, 4 of the 18 studies (22.0%) carried out multivariate regression analysis with adjustment for confounders. The remaining studies presented results using a variety of methods, including Spearman’s correlation analysis, Χ² or Fisher’s exact test or descriptive statistics.

Prevalence of vitamin D deficiency

We included 52 studies representing a total of 7434 children hospitalised with critical or acute conditions. Of these, 3473 (47.0%) were classified as vitamin D deficient (<50 nmol/L). Prevalence of deficiency ranged from 5.0%80 to 95.0%,60 median (IQR) 56.3% (31.9%–75.2%) (online supplementary table 12). Sample sizes ranged from 25 to 1016, with a median of 82 individuals (online supplementary table 13). Using a random-effects model, the pooled prevalence estimate of vitamin D deficiency was 54.6% (95% CI 48.5% to 60.6%) with a high proportion of variation attributed to heterogeneity (I²=95.3%, 95% CI 94.5% to 96.0%, p<0.0001) (figure 2) and evidence of funnel plot asymmetry (p=0.01, Egger’s test) (table 1 and online supplementary figure 1). Trim and fill analysis estimated 11 unpublished studies in the lower left-hand side of the funnel plot (online supplementary figure 1). The recalculated adjusted pooled estimate was lower 43.6% (95% CI 37.5% to 50.0%) with significant heterogeneity (p<0.0001).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Pooled prevalence estimate for vitamin D deficiency in acute and critically ill children by study design. Forest plot shows results from the random-effects model. Each diamond represents the pooled proportion of 25(OH)D deficiency for each of the subgroups (case–control, cohort, cross-sectional study designs). The diamond at the bottom represents the overall pooled proportion of all the 52 studies together. Each square shows the prevalence estimate of each study and the horizontal line across each square represents the 95% CI of the prevalence estimate. VD, vitamin D.

Sensitivity analysis for prevalence

We did not detect material differences in prevalence after exclusion of the 12 studies which did not directly report prevalence <50 nmol/L (53.0%, 95% CI 46.4% to 59.5%; I²=95.5%, 95% CI 94.5% to 96.2%, p<0.0001) (online supplementary table 14). 

When examining results by median sample size (defining ‘large’ as ≥82 and ‘small’ as <82), we found that the 268 53 55–58 66–68 72–75 78 79 81–91 studies with larger sample size included 6094 total individuals and gave a prevalence estimate of 51.5% (95% CI 43.6% to 59.4%; I²=96.8%, 95% CI 96.0% to 97.4%, p<0.0001). The remaining 26 studies with ‘smaller’ sample sizes included 1340 total children and estimated pooled prevalence as 58.2% (95% CI 47.5% to 68.2%; I²=90.9%, 95% CI 87.9% to 93.2%, p<0.0001) (online supplementary table 14).

We also conducted analysis by study design. Cohort studies (n=30) yielded a prevalence estimate of 49.6% (95% CI 42.7% to 56.4%; I²=94.9%, 95% CI 93.6% to 95.9%, p<0.0001). In case–control studies (n=18) the estimate was 68.1% (95% CI 56.5% to 77.8%; I²=93.0%, 95% CI 90.4% to 94.9%, p<0.0001) and in cross-sectional (n=4) 34.8% (95% CI 12.8% to 66.0%; I²=96.7%, 95% CI 94.0% to 98.2%, p<0.0001) (online supplementary table 14, figure 2).

We assessed whether studies’ country of origin influenced results. Studies in India gave an estimate of 68.9% (95% CI 54.9% to 80.1%; I²=96.7% (95% CI 94.0% to 98.2%, p<0.0001). Similarly, we found higher pooled prevalence estimates for studies from Turkey (76.3%, 95% CI 60.9% to 87.0%; I²=91.1%, 95% CI 84.2% to 95.0%, p<0.0001). We also grouped studies by geography and economic development. Group 1: USA, Chile, Australia, Canada, Ireland, Japan, Spain; group 2: South Africa, China, Egypt, Iran, Turkey, Saudi Arabia; and group 3: Bangladesh, Thailand and India. Prevalence was 37.2% (95% CI 29.7% to 45.5%) for group 1 (n=20), 61.8% (95% CI 53.2% to 69.7%) for group 2 (n=19) and 70.8% (95% CI 58.3% to 80.7%) for group 3 (n=13) (online supplementary figure 2). Variation attributable to heterogeneity was still high in the three subgroups (I²>90.0%).

Given the broad age range in included studies, we combined studies with only neonates54 65 67 69–72 and observed a prevalence estimate of 83.0% (95% CI 73.1% to 89.8%) with less variation attributable to heterogeneity (I²=76.6%, 95% CI 51.0% to 88.9%, p=0.0003). In all other studies (n=45) that included children of other age ranges, estimated prevalence was lower at 49.7% (95% CI 43.5% to 55.8%; I²=95.2%, 95% CI 94.3% to 96.0%, p<0.0001) (online supplementary table 14 and figure 3).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

Bubble plots of univariate meta-regressions. Each study is represented by a circle. Predictor variables: (A) clinical setting and (B) country groups are shown on the x-axis and the effect measure logit transformed proportion shown on the vertical (y-axis). Country group 1: USA, Chile, Australia, Canada, Ireland, Japan, Spain. Country group 2: South Africa, China, Egypt, Iran, Turkey, Saudi Arabia. Country group 3: Bangladesh, Thailand, India. grp, country group; NICU, neonatal intensive care unit; PICU, paediatric intensive care unit.

Post hoc investigation to determine sources of heterogeneity

To investigate the substantial heterogeneity observed in prevalence estimates, we incorporated study-specific characteristics (year of publication, total study sample size, quality score, study design, country group and clinical setting) as covariates in a random-effects meta-regression model. We identified clinical setting and country groups as significant predictors, p<0.01 (figure 3). We found that the model fitted with all available covariates can explain 32.9% of I² with F=4.57, p=0.001 (online supplementary table 15). We also conducted univariate meta-regressions for each of the six predictors (online supplementary figure 4).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Pooled prevalence estimate for vitamin D deficiency in children with sepsis. Forest plot shows result from the random-effects model. The diamond represents the overall pooled proportion of 25(OH)D deficiency from the meta-analysis of the 18 studies. Each square shows the prevalence estimate of each study and the horizontal line across each square represents the 95% CI of the prevalence estimate. VD, vitamin D.

Prevalence of vitamin D deficiency in children with sepsis and in those with respiratory tract infections

A total of 889 (median 42, range 9–160) patients had a diagnosis of sepsis, of which 565 (63.5%) were vitamin D deficient. Sixteen of the 18 studies including patients with sepsis were cohort (88.9%) and two (11.1%) were case–control (online supplementary table 16). Most studies originated from India (n=7), Turkey (n=3) or Ireland (n=2) and 16 were published between 2014 and 2017. Thirteen studies took place in a PICU and the remaining65 67 70 72 in NICUs. We found that all studies were of medium to high quality (median NOS score 7, range 6–8). Pooled prevalence of vitamin D deficiency was 64.0% (95% CI 52.0% to 74.4%) (figure 4) and median (IQR) 68.5% (50.4%–71.6%). Variation attributable to heterogeneity was high (I²=89.3%, 95% CI 84.6% to 92.5%, p<0.0001). Funnel plot was symmetric (p>0.05) suggesting no small-study effects (p=0.81, Egger’s test) (online supplementary figure 5).

• Download figure
• Open in new tab
• Download powerpoint

Figure 5

Pooled OR of risk of mortality in vitamin D deficient versus vitamin D non-deficient acute and critically ill children. Forest plot shows result from the random-effects model. Diamond represents the overall OR (with corresponding 95% CI). Each square shows the OR of each study and the horizontal line across each square represents the 95% CI of the estimate.

We also separately analysed studies of patients admitted for respiratory tract infections (n=25) such as acute lower respiratory tract infection, pneumonia and bronchiolitis. Of these 2699 total individuals (median 50), 1076 (39.9%) were vitamin D deficient. These studies were of high to medium quality (median NOS score 7, range 6–8). Most originated from India (n=6) and Spain (n=4). We found a prevalence estimate of 48.7% (95% CI 38.2% to 59.3%; I²=94.3%, 95% CI 92.7% to 95.6%, p<0.0001) and median (IQR) at 36.7% (24.3%–83.6%) with marginally non-significant evidence of bias (p=0.05, Egger’s test) (table 1). We therefore applied the trim and fill method and obtained an adjusted pooled estimate of 37.4% (95% CI 27.6% to 48.4%) after four studies were added.

Sensitivity analysis for prevalence in children with sepsis

Exclusion of the studies64 67 72 92 using thresholds other than <50 nmol/L for deficiency yielded a similar estimate of prevalence at 62.0% (95% CI 47.3% to 74.7%; I²=89.7%, 95% CI 84.5% to 93.2%, p<0.0001) (online supplementary table 17).

We examined pooled prevalence estimates according to median sample size (<42 vs ≥42). Studies with a smaller sample size (n=9; 204 total individuals) showed a pooled prevalence estimate of 64.7% (95% CI 52.5% to 75.3%) with moderate variation attributable to heterogeneity (I²=57.9%, 95% CI 11.8% to 79.9%, p=0.015). For the remaining nine studies (sample sizes ≥42, 685 total individuals) the estimate was 63.2% (95% CI 44.6% to 78.5%) with high variation attributable to heterogeneity (I²=94.3%, 95% CI 91.1% to 96.3%, p<0.0001).

There was no material change in prevalence estimates when analysed according to study design. The 16 cohort studies (749 total individuals) gave an estimate of 61.4% (95% CI 48.6% to 72.8%) with high variation attributable to heterogeneity (I²=88.8%, 95% CI 83.5% to 92.4%, p<0.0001). Case–control studies (n=2; 140 total individuals) showed a pooled prevalence of 80.0% (95% CI 58.8% to 91.8%; I²=81.3%, 95% CI 20.5% to 95.6%, p<0.0001) (online supplementary table 17 and figure 6).

Studies from India (n=7) gave a prevalence estimate of 66.0% (95% CI 51.4% to 78.1%; I²=81.1%, 95% CI 61.8% to 90.6%, p<0.0001). The three studies from Turkey assessing patients with sepsis gave a pooled estimate of 59.2% (95% CI 13.6% to 93.1%; I²=97.8%, 95% CI 95.8% to 98.8%, p<0.0001) (online supplementary table 17).

The pooled prevalence estimate in the four studies65 67 70 72 including neonates with sepsis was 73.7% (95% CI 60.3% to 83.8%; I²=76.0%, 34.1–91.3, p=0.006). The 14 studies with children of different ages, excluding neonates, gave a pooled estimate of 60.7% (95% CI 45.5% to 74.0%); I²=90.1%, 95% CI 85.2% to 93.4%, p<0.0001) (online supplementary table 17). Four of the studies56 61 87 89 included children admitted with either sepsis or respiratory tract infections.

Mortality in acute and critically ill children

We identified 18 cohort studies8 53 56–59 61 64 66 72 73 75–79 89 92 assessing vitamin D status and mortality. These studies included a total of 2463 individuals, from which 220 deaths (17.2%) were observed in 1278 (51.9%) individuals with vitamin D deficiency and 99 deaths (8.4%) were observed in 1185 individuals without deficiency (48.1%).

All 18 studies took place in a PICU apart from one,72 which considered only NICU patients. Sixteen of these studies (89.0%) were published between 2014 and 2017. Almost half (n=7) of the studies originated from India. Quality scores ranged from 5 to 8 with a median of 6.5.

Using a random-effects model, we found that vitamin D deficiency in critically ill children significantly increased the risk of death (OR 1.81, 95% CI 1.24 to 2.64, p=0.002) with low, non-significant heterogeneity (I²=25.7%, 95% CI 0.0% to 58.0%, p=0.153) (figure 5). However, small-study effects cannot be easily excluded (p=0.084, Egger’s test) (online supplementary figure 7) and the 95% PI (0.71–4.62) included the null value.

Sensitivity analysis for mortality in acute and critically ill children

We obtained similar results through the fixed-effects model (OR 1.72, 95% CI 1.27 to 2.33, p=0.0005) (online supplementary figure 8). When excluding studies with thresholds other than <50 nmol/L indicating deficiency, we found the association between vitamin D deficiency and increased risk of mortality still significant but lower, both with the random (OR 1.59, 95% CI 1.05 to 2.41, p=0.028; I²=24.3%, 95% CI 0.00% to 59.9%, p=0.191) and fixed-effects models (OR 1.52, 95% CI 1.08 to 2.13, p=0.016) without clear indication of small-study effects (p=0.120, Egger’s test) (online supplementary table 18).

The association was positive but not significant when pooling the seven studies from India with the random-effects model (OR 1.08, 95% CI 0.70 to 1.69, p=0.710; I²=0.0% 0.0–62.4, p=0.589) and similar with fixed effects (OR 1.08, 95% CI 0.70 to 1.69, p=0.710) (online supplementary table 18).

Mortality in patients with sepsis and respiratory tract infections

We were unable to identify a sufficient number of studies assessing vitamin D and mortality for meta-analysis in individuals with sepsis. Three studies8 64 67 measured vitamin D levels in paediatric patients with sepsis. One study8 assessed mortality and did not find a significant association in children from 1 to 12 years with sepsis (n=124). None of the studies with children admitted for respiratory tract infections looked at the association between vitamin D deficiency and childhood mortality.