12-month prevalence of atopic dermatitis in resource-rich countries: a systematic review and meta-analysis

There is a lack of robust prevalence estimates of atopic dermatitis (AD) globally and trends over time due to wide variation of populations and age groups studied, different study methodologies and case definitions used. We sought to characterize 12-month AD prevalence across the life span and change over time in resource-rich countries focusing on population-based studies and using a standardized AD case definition. This systematic review was conducted according to PRISMA guidelines. Medline (Ovid), Embase, WOS core collection, Cinahl, and Popline were searched for studies published since inception through August 15, 2016. Studies were synthesized using random effects meta-analysis. Sources of heterogeneity were investigated using subgroup analyses and meta-regression. From 12,530 records identified, 45 studies met the inclusion criteria. Meta-analysis with random effects revealed the 12-month period prevalence of 9.2% (95% confidence interval 8.4–10.1%). The prevalence was significantly higher among 0–5-year-old children (16.2%; 95% confidence interval 14.2–18.7%) than in older age groups. Studies using a random sampling strategy yielded lower prevalence estimates than studies relying on other sampling methods. There was no clear time trend in AD prevalence over the period of 1992–2013.

Information sources and search strategy. We searched Medline (Ovid), Embase, WOS core collection, Cinahl, and Popline from their inception until 15th August 2016. Search strategies, adapted for each search engine, included terms for "atopic dermatitis", "atopic eczema" and "prevalence" and individual names of EU/ EEA countries (and Switzerland), or "Europe", or the non-European high-income countries (Australia, Canada, Chile, Israel, Japan, Korea, New Zealand, USA) (see Supplementary material 1). In addition, we searched reference lists for eligible studies. The full electronic search strategy for Medline (Ovid) is presented in Supplementary table 1.
Selection process. Pairs of qualified reviewers independently screened the titles and abstracts, then by full text to determine eligibility for final inclusion (following the predefined inclusion criteria) and recorded the results onto standardized forms of a preformatted data collection template. At both stages of screening, any differences between reviewers were discussed and a consensus decision for eligibility and inclusion was made for all articles; a third reviewer resolved differences between reviewers if necessary.

Data collection process. A data extraction sheet was developed based on the Cochrane Consumers and
Communication Review Group's data extraction template 26 , pilot-tested on eight randomly selected but included studies and refined accordingly. If multiple publications reported one study, we extracted data from the primary publication (assigned as the publication with the most detailed description of the methods and the most data on specified prevalence measures). Data reported in the primary publication were used in case of inconsistencies between the publications based on a same source study. The two reviewers compared the extracted data and resolved differences by discussion. If there was still a discrepancy, a third reviewer adjudicated. We did not contact authors for additional information.
Data items. The following information was extracted: study design; country; setting (national or subnational); demographic characteristics (age, gender); sampling method (random-simple, stratified, multistaged, cluster; or convenience sampling); numbers of eligible, invited and participating subjects; number of subjects excluded and those with detected AD; definition of AD used; estimated prevalence and 95% confidence intervals (CI) reported in the study.
Study risk of bias assessment. We used published guidelines for cross-sectional prevalence studies by Boyle 27 to assess the risk of bias related to methodological aspects of included studies and disagreement at any stage was solved by consensus or arbitration. The items assessed included: representativeness of the target and source populations; similarity of responders and non-responders; attained sample size; use of standardized/valid AD measurement/definition; appropriateness of statistical methods; and response rate. We pre-specified criteria to determine whether each of the features in a specific study could be rated as attributing a low or high risk of bias, or if there was insufficient information to decide (unclear). The overall risk of bias generally corresponded to the highest risk of bias in any of the items. However, if a study was judged to have 'unclear' risk of bias for multiple (two or more) items, it was regarded as at high risk of bias overall. Publication bias was assessed qualitatively using funnel plot symmetry as a surrogate for low risk of publication bias. Data analysis and synthesis. We estimated the 12-month AD prevalence using the number of individuals with AD and the number of people tested (confidence intervals (CI) are based on the Clopper-Pearson method).
Outcomes of the study. The primary outcome was 12-month prevalence of AD. Secondary outcomes included the prevalence of AD across age, sex, study decade, AD case definition and country/region. For the current review, AD was pre-defined as (1) an itchy skin condition with a chronic and/or relapsing course and affecting the folds of the elbows, behind the knees, in front of the ankles, under the buttocks, or around the neck, ears or eyes 32,33 , or (2) diagnosed by a physician: (i) based on the self/parent report on atopic dermatitis/eczema diagnosis; (ii) observed in the study; or, (iii) extracted from a healthcare maintenance/administrative database according to the International Statistical Classification of Diseases and Related Health Problems (ICD-9/10) AD-specific diagnosis codes. Institutional review board approval was not sought as only data from already published studies was used.

Results
After removing duplicates, the search retrieved 8,856 records. After title and abstract screening, 7485 records were excluded. The remaining 1371 articles were read in full and screened for eligibility. After excluding 1326 articles for ineligibility, 45 studies remained for analysis. Expressly, out of the 532 studies excluded for covering other diseases, in 432 cases the disease investigated was defined as "eczema" (e.g., eczema or patient or doctor reported eczema or childhood eczema or food-sensitized and non-sensitized eczema or hand eczema) with no details allowing further clarification. In the rest of excluded studies hay fever (n = 39), food allergy (n = 28), asthma (n = 14), allergic rhinitis (n = 12) or atopy (n = 7) were studied. The flow chart of the selection process and reasons for excluding studies is detailed in Fig. 1.
Included studies. The data of the 45 included studies had been collected from 27 countries between 1992 and 2013 and published during the period from 1998 to 2016. Altogether, AD prevalence was assessed in 75,203,859 individuals (3,494,054 when excluding individuals whose data was obtained from healthcare databases) ( Supplementary Fig. 1 [74][75][76][77][78] , providing information also from France, Greece, Estonia, Latvia, Iceland, Belgium, Portugal and Hungary in addition to the aforementioned states. Table 1 summarizes the characteristics of each study. Risk of bias assessment and response rate. The assessments of risk of bias for each included study are presented in Supplementary table 2. Given the very limited information on data collection methods and response rates provided by the five studies of multiple countries [74][75][76][77][78] , formal risk of bias could not be assessed in these studies. Of note is that all of these studies compiled data from an international multisite collaboration research project based on a standardized study design 32,33 . From the remaining 40 studies, twelve (27%) were considered to have high risk of bias, and in 15 (33%) the risk was unclear, at least for one item assessed (Fig. 2). Thirty six studies out of the 45 (80%) utilized random sampling, six studies convenience sampling 34,35,42,46,54,73 and in three 45,47,61 the sampling strategy was unclear. In 24 of the studies utilizing random sampling, the sampling unit was a school 36,37,40,42,43,47,48,52,[54][55][56][61][62][63][64][65][66]68,70,[74][75][76][77][78] , in five the sampling unit was at an individual level 39,53,58,70,72 , four studies used multistage sampling (regional and individual) 36,59,60,68 , and three studies were based on health insurance data 40,50,51 . The studies using convenience sampling, or for which the sampling strategy was unclear, were regarded to be at high risk of bias in relation to the sampling from the source population.
The target population was assumed to represent the general population in 30/45 (67%) of studies. However, none of the studies provided a comparison between participants and non-participants and none of the studies  32 , and in 7 studies 37,38,52,59-61,69 modifications of the ISAAC instrument, were used. In the remaining two studies 53,57 , the self-reported score of symptoms were guided by criteria suggested by Hanifin and Rajka 79 , and in one study 58 by modified UK diagnostic criteria 80 . In three studies 36,39,48 , a self-report of physician diagnosis of AD was used and in one study 54 AD was diagnosed by a physician based on a UK refinement of diagnostic criteria 81 . In three of the included studies 40,50,51 assessment was based on ICD-10 diagnostic codes for AD in administrative health data.
12-month prevalence of AD. AD prevalence estimates ranged from 0 to 24% (Fig. 3). Meta-analysis identified overall 12-month period pooled prevalence of AD across all included studies of 9.2% (95% CI 8.4-10.1%) with a high level of heterogeneity. Meta-regression was used to explore potential variables that may  , the data could be drawn from just two countries, South Korea and the UK. Sensitivity analysis of age grouping did not reveal significant bias (Supplementary table 2). Further, the effect of study quality on the primary outcome was tested. Limiting the analysis to studies with low risk of bias (10 studies with 34 prevalence estimates [48][49][50][51]56,57,60,65,67,70 ) gave an AD prevalence of 8.9% (95% CI 7.2-11.0). This did not differ significantly from the prevalence estimate based on all included studies. Of note is that no studies with low risk of bias were detected within populations designated to age limits of either 0-5 or over 18 years.
Besides age and gender, the methodological characteristics of studies and the period of data collection were significantly associated with AD prevalence in meta-regression. The prevalence of AD was lower in studies reporting data from 2001-2010 compared to 1991-2000 (7.4 vs. 10.4%; p = 0.0004). Next, we stratified the time trend analysis by age group. There was a slight decrease in 12-month prevalence among 0-5-year-old and 6-12-year-old children from mid-1990s to the late 2000s (not statistically significant, p = 0.9119, and p = 0.8259, respectively; Fig. 4). Among individuals aged 13-18 years, a significant downtrend in predicted AD prevalence was observed over the same time (p = 0.0122), from 12.3% (95% CI 8.5-17.4%) in 1993 to 4.3% (95% CI 2.5-7.6%) in 2012. In the adult group, we saw an increase in AD prevalence over the past decades (p = 0.0036), from 5.7% (95% CI 4.0-8.7%) in 2005 to 13.7% (95% CI 9.6-19.2%) in 2012 (Fig. 4).  www.nature.com/scientificreports/  www.nature.com/scientificreports/ Studies using random sampling yielded lower AD prevalence than studies of non-random sampling (8.9 vs 13.9%; p = 0.005). There were no differences between Asia and other regions either in overall AD prevalence ( Table 2) or in trends across age groups (Fig. 4). We further explored the effects of the case definition of AD through sensitivity analysis. AD prevalence was 9.2% (95% CI 8.3-10.1%) in studies using self-report on symptoms, 7.0% (95% CI 5.5-8.8%) in studies using self-report on physician diagnosis (p = 0.2066) and 10.8% (95% CI 8.7-14.9%) in studies using other measurement methods (p = 0.2247).

Discussion
Our study describes the prevalence and trends of AD over the past three decades in resource-rich countries. Drawn from pooled data from countries of Europe, North America, East Asia and Oceania, we have come to two main findings. Firstly, it ascertains that nearly one-tenth (9.2%) of all people have experienced AD during last 12 months. Secondly, the prevalence of the disease has remained stable during the last decades.
We saw the highest 12-month AD prevalence of 16.3% in the youngest age group (0-5 years old), being almost twice as high as in older age groups. Our finding of higher AD prevalence among the youngest children corroborates previously published evidence that AD occurs more frequently in early life 3,4,7,82 . Also, longitudinal cohort studies revealing distinct disease trajectories in childhood and adolescence have depicted that the most prevalent subphenotypes are the 'early-onset-early-resolving' ones 9,10 .
In our analysis, the prevalence of AD in children aged 6-18 years and adults did not differ. This is concordant with the finding from a systematic review of longitudinal studies from Northern European countries of similar AD prevalence in the age groups of up to 12 years and older 83 . Likewise, in two British birth cohorts with a longer follow-up period, the annual period prevalence of AD at the age of 5 years and onwards ranged from 5 to 14% with no clear trend across ages 16 . The steady prevalence across ages older than 5 years probably reflects a balance of different disease trajectories, i.e., persistent disease as well as the phenotypes of AD that have resolved or relapsed for that period and later-onset disease. The observed 12-month prevalence of 9.3% among adults is echoed in the latest study from Finland where the prevalence an AD of 10.1% in the adult population was detected 84 .
Previous studies on gender differences in AD prevalence have come to conflicting results 85 . We found the overall female preponderance of 1.4:1.0. This finding is in agreement with the recent systematic review documenting a higher burden of AD among women throughout all age groups and geographic regions 3 . It has been speculated that skin care practices, occupational exposures, higher awareness or disease misclassification can play a role in this phenomenon, but to our knowledge these factors have not been formally studied. In a recent analysis, AD was not associated with endogenous sex hormones, neither in adolescents nor adults 86 .
Although there was a transient decrease of reported AD prevalence in the period 2001-2010, no convincing time trend was disclosed across the three decades. This observation is in line with the findings of the collaborative research looking at the secular trends in childhood AD and documenting the levelling off or decreasing prevalence of AD in some formerly high prevalence sites from high-income countries 21 . The nature of current study precludes assessment of causes of the drop in AD prevalence witnessed among 13-18-year-olds over time. Neither does it discriminate whether children born at late 1980s and early 1990s were less likely to have persistent AD or to develop AD later at school age, or both. So far, no age-group specific individual or environmental risk factors have been identified in the pathogenesis of AD 87 . Across studies, the most important predictors of AD persistence beyond childhood have been earlier age of onset, disease severity, allergic multimorbidity, family history of atopy, filaggrin gene mutations, and urban environment 88 . None of these appear to be modified easily. We propose that there could be a cohort effect attributable to not yet known changes in exposome since 1990s (e.g., more stringent use of antibiotics or promotion of breastfeeding). Another way to explain the decline would be more efficient treatment of AD (e.g., liberal use of emollients or proactive therapy) leading to changes in disease course. Interventional or prospective cohort studies of longer duration than usual are needed to clarify www.nature.com/scientificreports/ this. Concomitantly, we saw some increase of AD prevalence among people aged 19 years and older over time.

AGE GROUPS
One can speculate that there is no real downtrend of prevalence in adolescents, but the manifestation of AD is postponed into adulthood due to environmental or behavioural changes. Nevertheless, considering that lateonset AD was largely ignored until the 1990s 14,89 , it cannot be excluded that the increase of AD prevalence found in this age group is due simply to a rise in awareness of AD in adults. It has been debated whether anyone could develop clinical syndrome of AD if exposed to enough key risk factors or a finite number of only genetically predisposed individuals are susceptible to it 90,91 . On a large scale, stable AD prevalence throughout time and ages gives support to the latter hypothesis and reassures that the epidemic is not increasing infinitely. In our review, the sampling method was the only study design item associated with AD prevalence. Most of the included studies used a random sampling strategy and the remainder were mainly based on convenience samples. The latter yielded significantly higher prevalence estimates. This could have been anticipated since non-probability sampling based studies are known to have disadvantages and often oversample individuals with the condition studied due to overrepresentation either by respondents' self-selection, membership bias or non-responding [92][93][94] .
To-date, there is no consensus on how to capture AD at population level. Without a pathognomonic biomarker available, history taking and clinical signs are needed for diagnosis of AD and a clinician's assessment is considered the gold standard 7,95-97 . The fluctuating course of the disease further complicates collecting reliable data in population settings 91 . Previously, higher prevalence of AD has been reported in studies using participants self-report of AD compared to health care practitioner's assessment 98 . Interestingly, in the current analysis, we saw no difference in AD prevalence whether the outcome was measured by the diagnosis made by a health care practitioner or by a participant's self-report. Likewise, we did not reveal any differences across Asia vs other regions that have been described in some previous studies 3,12 . Despite the regional variations in single features of AD 99 our results indicate that the entity of AD might be the same in the setting of resource-rich countries. Thus, a standardized set of self-reported items delineating pruritic, inflammatory skin condition, characterized by a chronic and relapsing dermatitis in typical anatomical sites 100 could reliably detect AD in population-based research.
There are several strengths to this review. To our knowledge, this is the largest systematic review of AD prevalence with a comprehensive literature search, succinct focus on source studies' design validity (population-based studies, delineating sampling strategies) and using a pre-defined standardized AD definition as an inclusion criterion. Importantly, research findings may differ substantially if different definitions of AD are used 19,101 . In our study, despite within-sample heterogeneity, the AD prevalence estimates were quite similar irrespective of the ascertainment method.
There are also limitations that should be considered in the appraisal of the evidence presented by this review. High heterogeneity was observed between studies both overall and across subgroups. We explored whether study design aspects, geographic variation, or study period could explain part of the heterogeneity. However, the remitting and relapsing nature of the disease and residual study-level differences could have contributed to unexplained heterogeneity in outcome estimates. Namely, source population and sampling characteristics were often not described in enough detail and most studies using multistage sampling disregarded the complex design when estimating AD prevalence. High heterogeneity may affect interpretation and generalization of the results. Secondly, high-income countries included in the current study were defined by their EU/EEA and/or OECD status. This has closed out data from some other high-income countries such as Singapore and Taiwan. Thirdly, as we aimed to involve only data from affluent countries, the results cannot be attributed directly to low-income regions of the world. And last, but not least, in the analysis we were able to reckon only with the factors available in the original studies. Therefore, it was not possible to assess the effect of risk factors, such as living environment, migration or filaggrin gene mutations. However, these potential limitations seem unlikely to have accounted for the clear patterns observed in this study and we believe that the results allow inferences to be made in terms of the prevalence and trends of AD in the resource-rich populations.

Conclusions
Determining prevalence is a crucial step in understanding the impact of AD on affected people and health systems. Our results confirm that in affluent countries one-tenth of the general population suffers from AD annually and suggest that AD prevalence has not increased over time.

Data availability
The data extracted from included studies and used for analyses are available upon reasonable request from the authors. www.nature.com/scientificreports/