The prevalence of atopic dermatitis beyond childhood: A systematic review and meta‐analysis of longitudinal studies

Abstract Background There are sparse and conflicting data regarding the long‐term clinical course of atopic dermatitis (AD). Although often described as a childhood disease, newer population‐based estimates suggest the prevalence of pediatric and adult disease may be similar. Methods Our objective was to determine whether there is a decline in the prevalence of AD in population‐based cohorts of patients followed longitudinally beyond childhood. We conducted a systematic review and meta‐analysis including studies assessing AD prevalence across 3 or more points in time. The primary outcome was weighted overall risk difference (percentage decrease in AD prevalence). Results Of 2080 references reviewed, 7 studies with 13 515 participants were included. Participants were assessed at 3‐6 time points, ranging from age 3 months to 26 years. The percentage decrease in prevalence after age 12 was 1%, which was not significantly different from zero (95% confidence interval −2%‐5%). Similar results were found with other age cut‐offs. Conclusion The prevalence of AD in longitudinal birth cohort studies is similar in childhood and adolescence/early adulthood.

predictors of disease persistence or adult-onset disease. [3][4][5] Challenging the traditional notion of AD as a predominantly childhood disease, a growing body of research on the pathophysiology of AD points to genetic causes of altered skin barrier and immune dysfunction that could predispose to episodic disease throughout life. 6,7 Additionally, population-based estimates from cross-sectional surveys suggest that AD may be as common in adults as in children. 8 If AD prevalence does not decline over time, there are important implications for patient counseling, management, and future research.
The clinical course of AD has been particularly challenging to study because the condition is heterogeneous and intermittent. Individuals have different clinical presentations, and many will have periods without symptoms or skin lesions. Clinical trials generally focus on short-term disease control, and cross-sectional studies offer a snapshot of the population and hence cannot be used to generate prognostic information for individuals. Thanks to growing interest in atopic diseases, a number of longitudinal cohort studies included measures of AD disease activity at multiple time points in the same individuals, enabling the estimation of changes in AD prevalence as a given population ages. The primary objective of this study was to systematically review and analyze longitudinal studies of AD that include repeated measurements in the same cohort during and after childhood to test the hypothesis that the prevalence of AD does not decrease with age.

| Data source and searches
This review is reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) and Meta-analysis of Observational studies in Epidemiology (MOOSE) guidance. 9,10 Our research protocol was registered on PROSPERO and was publically available prior to the study start date (registration ID: 42016033553). PubMed and EMBASE were searched from database inception to October 2015 using prespecified search terms, MeSH (medical subject heading) headings and keywords, developed in consultation with a professional medical librarian. Search terms are described in Table S1. Reference lists of included studies were also screened for additional potentially relevant articles.
This study was exempt from IRB review because it included only published data and is not considered Human Subjects Research by the UCSF IRB (http://irb.ucsf.edu/not-human-subjects-research).

| Study selection
We included studies of patients with AD as diagnosed by a physician or using a standardized definition such as the UK working party diagnostic criteria for AD or the International Study of Asthma and Allergies in Children ISAAC criteria. 11,12 Because our objective was to examine changes in AD prevalence within a given cohort as populations aged, we included only longitudinal cohort studies that assessed AD activity in the same patients at 3 or more distinct time points, to account for the episodic nature of AD and allow for observation of any nonlinear patterns of change. Further, to capture AD prevalence beyond childhood, studies were required to include at least 1 assessment with participants over the age of 12 years. We chose this cutoff because prior studies suggested high rates of "remission" in adolescence and a longitudinal prospective study suggested that prevalence declines most rapidly between ages 8 and 11. 13,14 We excluded studies that were clinic-based, or focused on patients with localized forms of AD (eg, hand eczema). We did not select studies on the basis of interventions and our search included studies in all available languages.

| Data extraction and quality assessment
Title and abstract screening followed by full-text screening was performed independently and in duplicate to ascertain whether studies met eligibility criteria. Extracted variables included study information (authors, year of publication); cohort information (country of study, cohort name, method of participant recruitment); study size (number of patients at study start/end, and alternate measures of loss to follow-up); patient characteristics (sex, race/ethnicity, other demographic information); AD diagnostic criteria; and finally, method and age of AD prevalence assessment.
Methodological quality was assessed using the Newcastle-Ottawa Scale for cohort studies. 15 The risk of bias and scoring criteria used in the present review are described in Table S2. Screening, data extraction, and risk of bias assessment were all performed in duplicate by 2 authors (AY and JO), and discrepancies were resolved by group consensus (AY, JO, KA, and SML).

| Data synthesis and analysis
The primary outcome was percentage change in AD prevalence by age, and all measured time points were included. Secondary outcomes included the prevalence of other atopic conditions (asthma, rhinitis/hayfever), treatment information, and severity data. Corresponding authors of included studies were contacted via e-mail to verify extracted study prevalence data used for the analyses.
We tabulated and plotted the AD prevalence by age for each study. We then calculated the standardized difference in mean prevalence before and after the age of 12 years (using all time points from each study), and report individual study differences and the mean weighted risk difference. We tested for heterogeneity with the chi-square test and measured inconsistency using the I 2 statistic, which represents the percentage of total variation across studies. 16 Because of the methodologic and clinical heterogeneity inherent in the data, we used random-effects models for all meta-analyses.
Preplanned sensitivity analyses were conducted to assess the impact of varying the age cutoffs. The primary analysis was repeated using different age cutoffs: <10 years vs ≥10 years; <8 years vs ≥8 years. These analyses were performed to test whether our choice of an age 12 cutoff affected the results. We also repeated the analysis comparing those aged <2 years to 3-11 years and those aged 2-12 years vs >12 years to test whether including very young children ABUABARA ET AL.
| 697 (when the diagnosis of AD can be more challenging) affected our results. Additional stratified analyses were performed to explore sources of heterogeneity, including analyses by age, by country, by loss to follow-up, by the number of measurements, and a "jackknife" analysis that eliminated 1 study at a time. Egger's test was used to evaluate for publication bias. 17 All analyses were performed with Stata (version 14, Stata Corporation, College Station, TX, USA).

| RESULTS
The search yielded 2080 records. After full-text review of 28 manuscripts, 7 studies were selected for inclusion. 14,[18][19][20][21][22][23] The study flow diagram ( Figure 1) lists the reasons why 21 studies were excluded. Table 1 outlines the characteristics of each study in detail. The  Table S3. Data were largely unavailable for the secondary outcomes: Only 4 studies reported data on the prevalence of other atopic outcomes, and the methods for assessment and reporting differed. None of the studies reported data on treatment or severity of AD (Table 1).

| Prevalence estimates
The annual period prevalence of AD ranged from 6% at the age of 26 years 22 to 34% at the age of 12 years. 21 Study-specific figures are presented graphically in Figure 2, and the exact numbers used in the meta-analysis are included in Table S4.
Data on gender were inconsistent. Gough

| Risk difference
The primary outcome was the percentage change in prevalence after the age of 12 years. This ranged from À0.05 (meaning there was a 5% increase in the prevalence after the age of 12 years, 95% CI: The publication only included those with complete data on all health outcomes at every follow-up, which represented 71% of the original cohort. b The publication only included those with complete data on all health outcomes at every follow-up, which represented 47% of the original cohort. c At the age of 7, participants were asked about any AD up to the age of 7, at subsequent ages they were asked about AD over the past year.  (Figure 3). 20 The overall mean weighted percentage difference in prevalence was not significantly different from 0 (0.01, a 1% decrease, 95% CI: 0.02, À0.050). We found similar results when we explored multiple age cutoffs in sensitivity analyses (Table S5), and Eggers test showed no significant publication bias (P = .314). Because of the small number of included studies, funnel plots were not created.

| Heterogeneity
There was significant heterogeneity in the primary analysis (I 2 96.3%).
Multiple sensitivity analyses were performed to explore the impact of study design features on the degree of heterogeneity and the outcome. First, we considered differences in the age structure of the  Table S5 and Figures S1 and S2).

| Risk of bias
The included studies had variable risk of bias with Newcastle-Ottawa scores ranging from 3 to 7 of a total maximum score of 8.  (Table S6). In Burr et al's study, the definition of AD used varied over time and could have introduced bias: They measured the point prevalence of AD through age of 7 years, then annual period prevalence from ages 15 and 23 years, which could make AD rates appear higher during adulthood. A sensitivity analysis excluding this study did not change the results (Table S5).
All studies had good comparability due to equivalent measurements between AD and non-AD controls, and all studies had an adequate length of follow-up. Loss to follow-up was >30% for 3 of 7 studies; excluding these studies did not affect our results (Table S5 and Figure S3).

| Analyses of repeated measures among individuals
Four studies examined whether individuals reporting AD symptoms at each age had a history of prior disease. All found that there was substantial turnover with many new and remitting cases throughout the study period. However, studies were inconsistent in providing study definitions of disease "remission" and "persistence" (Table S3). In 1 study, persistence was defined as AD symptoms at 2 or more observation points at any time within the study 18 ; another required subjects to have AD symptoms over the course of all study assessments, and across consecutive observation points. 23 This inconsistency in the definition we are unable to conclude whether AD tends to get better or worse over time. Nonetheless, the available data suggest that the reason for steady prevalence estimates across ages is due to a combination of active disease in both childhood and early adulthood among some individuals, remission or clearance of disease among others, and lateronset disease among others.

| Comparison with other studies
A recent review focused on "AD persistence" concluded that "80% did not persist by 8 years." 24 It measured the change in the average proportion of a population reporting AD symptoms upon second measurement at any age. Upon first glance, these findings may seem contrary to ours. However, the authors caveat that "it is possible that some of the patients reported to have AD remittance had unobserved recurrences later in life." Thus, although their results are not directly comparable to ours because they included studies with variable populations, definitions of AD, and timing of follow-up, their findings also highlight the episodic nature of the condition.
Our results concur with previous work showing high rates of AD persistence in a longitudinal cohort of children and young adults treated with pimecrolimus, which may be biased toward more persistent disease. 25 This study did not meet inclusion criteria for this review because enrollment was clinic-based and it included only patients with AD which did not allow for estimates of disease prevalence. The population-based ISAAC studies calculated AD prevalence at 2 age points in sites around the world, but were cross-sectional in nature and did not follow the same individuals over time. There was variation in prevalence by site and by age, but as with our findings, the average prevalence at both time points was similar (7.9% at the age of 6-7 and 7.3% at the age of [13][14]. 26 It is important to highlight that diagnostic definitions may impact prevalence estimates, possibly biasing toward smaller numbers at older ages. Notably, the 1 study that required symptom onset occur in "early childhood" for the definition of AD found the sharpest decline in prevalence, 20  Party Criteria, assign extra points if the AD diagnosis is made in early childhood. 11,27 Subsequently, studies using these criteria may report smaller numbers of individuals with later-onset disease, which should be taken into account when considering estimates of adult AD.

| Strengths and limitations
Major strengths of our review and meta-analysis include the comprehensive search, careful selection and critical appraisal of studies, and inclusion of large representative cohorts followed from birth over 2-3 decades. Of note, we did not specify the initial age of follow-up in our review; it was coincidental that all of the studies that met inclusion criteria were designed as birth cohort studies.
Limitations also warrant discussion. First, there is variability in the terminology used to describe AD: "Atopic dermatitis," "atopic eczema," and "eczema" are all used to refer to the condition. 28 In some settings, "eczema" is considered a less specific term that may include other types of dermatitis. Therefore, we carefully reviewed the diagnostic criteria used in each study (Table S3) and assessed the impact on our results as described above. Second, we found a high level of statistical heterogeneity between studies. This might be due in part to the use of a risk difference as our outcome measure, as risk differences are often less homogeneous. 16,29 We used a model with random effects to help account for the statistical heterogeneity, and found no appreciable differences in the results after multiple analyses to address differences in the methods for and timing of AD ascertainment between studies (Table S5). Nonetheless, our results should be interpreted with caution and replicated as additional follow-up data from longitudinal cohorts become available. Third, most studies lacked important information on secondary outcomes including treatment use, disease severity, and other atopic conditions. These data may impact the clinical course of AD and are important to report.
Fourth, as is common for studies that follow large cohorts of patients over decades, attrition was substantial, ranging from 10 to 53%. Most of the studies (7 of 9) were not designed to specifically study AD, and none were organized through clinic visits so dropout was likely random with respect to AD and would therefore be unlikely to bias prevalence estimates. However, other factors differentially associated with attrition such as race/ethnicity and socioeconomic status may also be associated with AD resulting in selection bias. 30 These data should be reported in future longitudinal studies. Finally, all of the studies identified were from Northern European countries, and patients were only followed into the third decade of life. Additional work is needed to understand the clinical course throughout adulthood and in more diverse settings.

| CONCLUSIONS AND IMPLICATIONS FOR PRACTICE AND RESE ARCH
Our finding of similar AD prevalence before and after childhood supports the emerging paradigm of AD as a lifelong genetic predisposition to episodic skin lesions. 3 Studies examining predictors of individual disease course are a high priority for additional research and can help to elucidate the relationship between genetic susceptibility and environmental influences on the epidemiological trends described herein.

ACKNOWLEDGMENT
We are indebted to Ms. Carol Mita for her assistance with developing and validating our study search strategy.

CONF LICT OF I NTEREST
None reported.

AUTHOR CONTRIBU TI ONS
All authors were involved in the study design. Dr. Abuabara had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Drs. Abuabara, Langan, and Allen involved in study concept and design. Yu and Okhovat involved in acquisition, analysis, and interpretation of data. Yu and Abuabara drafted the manuscript. All