Protocol and registration

The protocol for this systematic review has been published prior to the start of this study.³ In addition, the protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD42018087500).⁴ The changes to the protocol are summarized in Appendix S1 (Supporting Information). The study is reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.⁵

Eligibility criteria

We have excluded studies of localised eczema such as hand eczema, if not mentioned specifically in patients with AD, and other types of eczema such as contact dermatitis and seborrheic dermatitis; literature reviews, case reports and case series; conference abstracts, books and book chapters; and studies on other phenotype categories than defined above (including subgroups based only on age, gender, ethnic populations, presence of triggers, comorbidities, immunology and genetics). Ichthyosis vulgaris, prurigo nodularis and keratosis pilaris in AD patients were considered morphological features.

Search strategy and information sources

A comprehensive literature search strategy was developed in consultation with a clinical librarian. We have searched Ovid EMBASE, Ovid MEDLINE and Web of Science from inception till 12 February 2021. No language restrictions or filters were applied. The Ovid MEDLINE search strategy can be found in Appendix S2 (Supporting Information). In addition, the reference lists from three major review articles were hand-searched for relevant studies.^{1, 2, 6}

Study selection process

The results of the literature search were uploaded into Covidence online software. All titles and abstracts were screened independently by two reviewers, using a screening tool based on our eligibility criteria. Publications that both reviewers recorded as meeting the inclusion criteria were retrieved for full-text review and excluded when not meeting the criteria. Disagreements were discussed with a second reviewer if necessary. Persistent conflicts were resolved with a senior author. Thereafter, full-text publications were reviewed in duplicate by two separate reviewers. Disagreements were resolved after discussion between the reviewers and with a senior author if necessary.

Data extraction process

Data from each full-text publication were independently extracted by two reviewers (A.B., A.A., R.I., K.F. and J.M.), using a data extraction form designed for this purpose. Discrepancies in data extraction were resolved by discussion if necessary.

Data items

We extracted the following data domains from the included publications using our predesigned data extraction form: study data, disease data and outcome data. The study data comprised the following items: year(s) conducted, study design, setting conducted in, country/countries conducted in, World Health Organization (WHO) region, and the number, age and gender of the participants with (atopic) eczema. The following disease data items were extracted: disease description, diagnostic criteria/codes and disease severity definition. The following outcome data items were extracted: qualitative description of the phenotype(s), proportion of individuals in each phenotype (if relevant), qualitative description of the characteristic(s) (of a priori interest) potentially associated with the phenotype(s), result of the statistical analyses on the association, methodological approach for deriving phenotype(s) and/or investigating the association (including a data-driven approach using statistical techniques, rather than the predefinition of phenotypes, if applicable), and whether controls were included (including the number).

Synthesis of results

The results are reported descriptively. We anticipated that both the phenotype definitions and potentially associated characteristics that are investigated would vary between studies. Therefore, we expected heterogeneity in all outcomes. We have grouped studies into categories where possible and composed evidence tables per phenotype category. If more than one phenotype category was applicable to one study, the publication was grouped into all relevant categories.

Risk of bias assessment

Risk of bias was assessed per study using the critical appraisal checklists for analytical cross-sectional studies, cohort studies and case-control studies from the Joanna Briggs Institute (JBI), as appropriate.⁷ In the forms, we have treated the described phenotype as the outcome and the description of the potentially associated characteristics under investigation as the exposure. Traffic light tables were composed according to study design and phenotype category to visualize the qualitative results descriptively.

Quality of the evidence

We aimed to use the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach for assessing the quality of evidence per phenotype category. As we anticipated that the phenotype definitions and potentially associated characteristics would vary between studies, an assessment was made whether the quality of evidence per phenotype category could be investigated.

Search results

We have screened 8511 records and have assessed 675 full-text publications. In total, 186 studies, published between 1966 and 2021, fulfilled the inclusion criteria. Reference searching has yielded 6 additional publications. Figure 1 gives an overview of the study selection process, including reasons for exclusion.

Study overview

Of the included articles, 59% (109/186) was hospital-based (medical specialist setting). Regarding study design, 76% (141/186) was cross-sectional studies. In 7% of studies (13/186), a data-driven approach was used to derive phenotypes, including two studies using existing data-driven phenotypes. The number of included AD patients ranged from seven to 526 808. Ninety-one (49%) publications investigated phenotypes based on disease severity (phenotype group 1). Phenotypes based on disease trajectory (phenotype group 2) were investigated in 56 (30%) studies. Thirty-six (19%) studies investigated morphological characteristics in other phenotypes (phenotype group 5). A morphology-based phenotype (phenotype group 3) and a phenotype of AD patients having eczema herpeticum (phenotype group 4) were investigated in 22 (12%) and 11 (6%) studies respectively. There was an overlap between phenotype categories in 26 studies, with two (n = 22) to three (n = 4) phenotype categories being investigated in one study. An overview of the study characteristics per study grouped per phenotype category can be found in Table S1a–e (Supporting Information).

Risk of bias

The risk of bias of studies is reported in Table S2a–e (Supporting Information), demonstrating the qualitative results of the JBI critical appraisal checklists according to study design and phenotype category. We decided not to give an overall estimation of the risk of bias per paper but to descriptively report the checklist results per paper.

In various papers (30%, 27/91) within the disease severity category (phenotype group 1), no predefined scoring system or severity cut-offs were reported, resulting in the score unclear for outcome in the risk of bias assessment.^8-10 In many papers (27%, 15/56) on disease trajectories (phenotype group 2), age cut-offs were unclear or it was unclear who assessed the age of onset (i.e. whether it concerned reports by patient, parent or physician).^11-13 A lack of detail was identified regarding phenotypes based on morphological features and the investigation of morphological characteristics in other phenotypes (phenotype group 3 and 5). Often (in 69%, 41/59), it was unclear who performed the assessment or no criteria or further specifications for the assessment of morphological characteristics were reported (i.e. when characteristics were considered present or not).^{14, 15} Overall, in many cross-sectional studies (60%, 84/140), the subjects and setting were not described in sufficient detail.^16-18 In addition, the absence of inclusion of potentially confounding factors in the analyses of many studies (55%, 102/186) was noteworthy. A major source of bias across studies related to the two latter factors in the checklists.

Quality of the evidence

We found heterogeneity in the phenotypes and investigated characteristics that were reported in studies and the results of this review are descriptive. Therefore, following discussions with author M.L., an international leading GRADE researcher, assessing the quality of the evidence with GRADE, was considered not relevant.

Study results

An overview of all studies in alphabetical order per phenotype category and details of the results are found in Table S1a–e (Supporting Information). The results of the statistical analyses are summarized per phenotype category and per category of associated characteristics in Appendix S3 (Supporting Information). An overview of all phenotypic groupings and their investigated characteristics can be found in Fig. 2 (graphical abstract).

Summary of evidence

We have undertaken a comprehensive analysis of the published literature on phenotype definitions used in literature and have described the characteristics associated with phenotypes. Phenotypes of patients with AD have been identified based on various features, including disease severity, disease trajectories, morphology and predisposition to eczema herpeticum. With this systematic review, we have gained insight on how these phenotype categories are reported in the literature, thereby contributing to developing a better understanding of AD. This systematic review highlights the heterogeneity that currently exists in the phenotyping of the AD population. In the literature, many phenotypes based on many features are described. At the present time, no consensus exists on how these phenotypes of AD should be defined, and the potential role of phenotypes in guiding both diagnostic and therapeutic management of patients is unknown.

Across the phenotype categories there were both differences and similarities in study characteristics. We identified mostly cross-sectional studies (n = 141, 76%) in predominantly hospital-based settings (n = 109, 59%). Hospital-based studies could be subject to selection bias due to referral criteria, potentially leading to the identification of other phenotypes than when a population-based approach was used. Therefore, hospital-based studies need to be interpreted cautiously.¹⁹⁴ Phenotypes based on disease severity were most frequently studied (n = 91, 49%). The methodological approach for investigating phenotypes differed between studies. Besides the predefinition of a subgroup of patients based on certain features (e.g. cut-offs for age or severity), statistical data-driven approaches were also used to identify phenotypes in the minority of studies (n = 13, 7%), for example, by using latent class analysis or cluster analysis. Though these data-driven approaches are only used in a minority of studies aiming to investigate phenotypes, this can be considered a relatively unbiased way to identify phenotypes, in contrast to an approach using an investigator-imposed predefinition. The most frequently used data-driven approach is latent class analysis. Data-driven approaches have the potential to identify patterns that are not obvious to clinical observation. Unfortunately, this only was performed in a small number of studies.

Strengths and limitations

No previous systematic reviews were undertaken to map the current evidence on AD phenotypes in the literature. Librarians were involved in composing a comprehensive and broad search strategy. The protocol of this systematic review was published and preregistered. Moreover, we adhered to PRISMA guidelines in the reporting of this study.

Limitations include that since both the phenotype definitions and the a priori defined characteristics of interest differed between studies, we were unable to pool results and did not use GRADE to assess the quality of evidence. Accordingly, no meta-analyses could be undertaken due to this heterogeneity in study outcomes, and therefore, we have reported on all studies separately in the evidence tables (Table S1a–e, Supporting Information) and Appendix S3 (Supporting Information). Meta-bias resulting from publication bias or selective outcome reporting bias could not be assessed formally because of the qualitative nature of the study. However, both types of bias are deemed unlikely because of our rigorous search and descriptive nature of the studies. Studies were retrieved by our search when the term phenotype or synonyms of phenotype were specifically mentioned. In other words, studies that have used other terminologies (i.e. studies that describe phenotypes, but do not use the terminology phenotype or synonyms of phenotype) could have been missed. A bias for recent studies may have been introduced by the absence of these terminologies at inception of the used databases. Case reports and case series, for example, describing morphological phenotypes were excluded. Although we report associations between phenotypes and characteristics, these do not prove any causal relationship, and many are based on small sample sizes in hospital-based populations and hence should be interpreted with caution due to the possibility of referral and selection bias. In context of the scoping nature of this systematic review, we did not restrict to a specific study setting, size or confounder adjustment. Lastly, because the term phenotype is used in numerous ways in the literature, we had to define phenotype for consistency, and in the context of this systematic review, we have defined phenotype as a subtype or subgroup of patients with AD. In the context of precision medicine, a semantic distinction with endo(pheno)types would be of interest. It was not feasible to include all potential phenotypic groupings in this study. Therefore, we were forced to make choices on which phenotypes to focus, which resulted in focussing only on the most clinically relevant phenotypes. Excluded phenotype categories include subgroups based only on age, gender, ethnic populations, presence of triggers, (allergic) comorbidities, immunology and genetics.

Implications and recommendations for future research

At the moment, the therapeutic management of AD is generally not based on phenotypes that could reflect potentially relevant differences in characteristics between patients, with the exception of severity. In theory, these differences in phenotypes could be associated with variations in treatment outcome. In the context of personalized medicine, stratification according to phenotype would be of interest to enable investigation of which patients are likely to respond best to certain therapies. In order to facilitate comparative or pooled analyses across studies in the future, phenotypes should be uniformly defined and consistently used. Ideally, researchers should use the same definitions for AD phenotypes in research, similarly to using the same core outcome set for outcome measurements in clinical trials and clinical practice (www.homeforeczema.org/). This core outcome set already includes the recommendation of using the EASI to measure disease severity. A previous study has determined and recommended the following severity strata for EASI: 0: clear, 0.1–5.9: mild, 6.0–22.9: moderate and 23.0–72: severe.¹⁹⁸ We should preferably use the same outcome measurements and cut-offs to describe disease severity phenotypes. Regarding phenotypes based on disease trajectories, we ideally should use the same definitions, e.g. early-onset disease, by using uniform age cut-offs, when using non-data-driven approaches. However, first, we should get a clearer picture of the predictive ability of such cut-offs. As for morphology, it would be desirable to develop (diagnostic) criteria for morphological phenotypes, as current diagnostic criteria for AD do not facilitate the identification of these or phenotypes in general.^{199, 200} The current heterogeneity in phenotyping of AD has demonstrated a need for international harmonization. More research using unbiased data-driven approaches in well-defined, population-based settings should be considered to allow for the identification of phenotypes that are not obvious to clinical observation. Selection of appropriate data-driven techniques should be guided by the nature of the dataset, e.g. whether it is cross-sectional or longitudinal, and by the types of input available (disease activity, severity, clinical presentation etc.). To date, most cross-sectional data-driven techniques have been from the family of cluster analysis, and longitudinal data techniques have been from the family of mixture models such as latent class analysis. Phenotypes identified by a wide range of cross-sectional data may be more richly characterized than phenotypes identified by a smaller range of fewer but longitudinally collected data, but their interpretation may be different. For example, cross-sectional phenotypes may describe clinical AD presentation well but may be less suitable to track the persistence or resolution characteristics that longitudinal phenotypes characterize and vice versa. Whatever the method, the resulting phenotypes should be interpreted in context of the demographic characteristics (e.g. age, sex, ethnicity and geographical region) of the population represented by the sample used to derive them, i.e. not assumed to be applicable to populations not included in the sample. Phenotype studies should also be replicated in independent populations to investigate the stability of the identified phenotypes. In addition, it would be of interest to investigate phenotypes based on allergic comorbidities, since we apprehend AD as part of a larger group of diseases with TH2 inflammation skewing. Lastly, the identification of clinically meaningful phenotypes in the context of treatment outcome should be pursued, by investigating therapeutic effectiveness and safety in patients stratified according to phenotype.