Why was the study undertaken?This analysis aimed to generate severity thresholds for clinician‐ and patient‐reported outcome measures of atopic dermatitis severity using a large, diverse clinical trial dataset of approximately 30,000 visits by 2822 adults with moderate‐to‐severe atopic dermatitis.What does this study add?Severity thresholds were evaluated using novel statistical methodologies (frequency cross‐over analysis and means based on spline regression), with the largest data set to date, and Sleep Loss VAS severity bands were assessed for the first time. Potential threshold bands demonstrated high goodness‐of‐fit, and most demonstrated high concordance.What are the implications of this study for disease understanding and/or clinical care?This severity band analysis of clinician‐ and patient‐reported outcomes using a large global data set of adults with moderate‐to‐severe atopic dermatitis may add to the understanding of atopic dermatitis disease severity and stratification in a clinical trial setting.
INTRODUCTION
Atopic dermatitis (AD) is a multifaceted disease characterized by chronic skin inflammation, pruritus and reduced quality of life (QoL).1, 2, 3 Various clinician‐reported outcomes (ClinROs) and patient‐reported outcomes (PROs) are used to assess different aspects of AD. The Harmonising Outcome Measures for Eczema (HOME) initiative, a multi‐stakeholder group, recommends utilizing the Eczema Area and Severity Index (EASI),4 Dermatology Life Quality Index (DLQI),5 Patient‐Oriented Eczema Measure (POEM) and Numeric Rating Scale (NRS)‐itch and Atopic Dermatitis Control Tool (ADCT) or Recap of Atopic Eczema (RECAP) instruments in all AD clinical trials.6 The SCORing Atopic Dermatitis (SCORAD) index developed by the European Task Force on Atopic Dermatitis is another widely used and validated tool that includes objective and subjective measures to evaluate disease severity and extent.7
Severity bands enable patients and physicians to translate the numeric measures of AD severity instruments into understandable and meaningful tools. Thresholds for severity bands have been derived from data obtained in a variety of settings, including clinical trials,8 outpatient clinical practice9, 10, 11, 12, 13, 14, 15 and cross‐sectional surveys16 (Table S1). Many of these analyses were based on limited sample sizes and geographic regions.
Moreover, multiple sets of severity bands have been published for some measures, including SCORAD,8, 9 objective SCORAD (oSCORAD),8, 9, 10 EASI,9, 14 DLQI11, 15, 16 and POEM.9, 13, 16 The generation of severity band thresholds using larger, more diverse data sets may provide additional insights into clinical and QoL outcomes in patients with AD.
Herein we report an analysis of severity bands for several AD ClinROs and PROs using a substantial, pooled data set of global, randomized, placebo‐controlled clinical trials of dupilumab in adults with moderate‐to‐severe AD.
METHODS
Study population
We used data collected during 2013–2016 from five double‐blind, randomized, placebo‐controlled trials of dupilumab in adults with moderate‐to‐severe AD (ClinicalTrials.gov Identifiers: NCT01859988, NCT02277743, NCT02277769, NCT02260986 and NCT02755649; Table S2).17, 18, 19, 20 Efficacy assessment data from all available visits were pooled across all patients who received dupilumab or placebo. At each visit, disease severity data were collected from clinicians and self‐reported by patients using four ClinRO and four PRO instruments: total SCORAD (range: 0–103), oSCORAD (comprising the clinician‐assessed extent and severity components of SCORAD; 0–83), EASI (0–72), body surface area (BSA; 0%–100%), DLQI (0–30), POEM (0–28), SCORAD Pruritus VAS (0–10) and SCORAD Sleep Loss VAS (0–10). Higher scores represent greater severity for all instruments.10, 21, 22, 23, 24, 25, 26 The trials were conducted following the ethical principles derived from the Declaration of Helsinki, the International Conference on Harmonisation Good Clinical Practice guideline and applicable local regulatory requirements.
Written informed consent was obtained from all patients prior to enrolment and to any study procedure.
Outcome measures and severity band threshold analysis
Statistical analyses were performed using SAS version 9.4 (SAS Institute) and R version 4.2.2 (The R Foundation for Statistical Computing). All analyses were anchor‐based. A summary of the methodology can be found in Figure S1.
To identify the most appropriate anchor for each instrument, Spearman correlation coefficients were calculated to determine the strength of association between each instrument and two anchors: Investigator's Global Assessment (IGA) and Patient Global Assessment of Disease Status (PGADS). IGA results are scored on a scale of clear (0), almost clear (1), mild (2), moderate (3) and severe (4). For PGADS, patients are asked ‘Considering all the ways in which your eczema affects you, indicate how well you are doing’, and they can choose from the following responses: poor (1), fair (2), good (3), very good (4) or excellent (5).
Strength of correlation was defined as negligible (r = ±0.00–0.30), low (±0.31–0.50), moderate (±0.51–0.70), high (±0.71–0.90) and very high (±0.91–1.00).27 Correlations of ClinROs were higher with IGA than with PGADS, while correlations of PROs were higher with PGADS than with IGA (Figure 1); therefore, the ClinROs were anchored to IGA and the PROs to PGADS.
Means (using spline regressions), frequencies, medians and modes (using a previously published methodology)9 of the anchors calculated for each integer value of the instrument were utilized to determine severity band thresholds. Spline regression analyses quantify relationships in a segmental, or piecewise, manner, which are then linked from one ‘segment’, or regression range, to the next. The points at which the piecewise regression lines join (‘knots’) are determined by an algorithm that optimizes the fit of the regression model and, when rounded to the nearest whole numbers, represent candidate severity band thresholds; knots for each outcome measure are reported in Table S3.
Bootstrap methods were applied to ensure robust estimates of the knots and their standard errors. Frequency cross‐over evaluations were performed, in which the threshold of the first severity band was set to 0 and each successive severity band threshold was set at the value of the outcome measure at which the maximum anchor frequency changed category. Finally, modified versions of previously published severity bands (e.g.
Chopra et al.9) were evaluated.
Outcome measure severity bands were designated in order of severity into five categories: clear/almost clear/mild/moderate/severe (total SCORAD, oSCORAD, EASI, BSA), excellent/very good/good/fair/poor (DLQI, POEM) or none/mild/moderate/severe/very severe (SCORAD Pruritus VAS, SCORAD Sleep Loss VAS).
Two analyses were performed using (1) all patient visits and (2) excluding patients' first visit, which allowed them entry into their respective clinical trial. Due to the relatively low correlation between anchors and endpoints at first visit, this timepoint was removed from the analysis. Results showing the analyses using all timepoints can be found in Table S3.
Candidate bands were compared empirically by evaluating R2 and weighted kappa (κ)‐coefficients. R2 values (range: 0–1) were used to assess goodness‐of‐fit, with higher values indicating better fit. To measure concordance between candidate severity bands and their respective anchors, weighted κ‐coefficients (range: −1.00 to 1.00) were calculated, based on Cohen's κ.28 Weighted κ‐coefficients were calculated using quadratic weighting; additional weighting analyses using equal and linear weights are reported in Table S5.
Concordance strengths for the weighted κ‐coefficients were defined as no agreement (0), slight agreement (0.01–0.20), fair agreement (0.21–0.40), moderate agreement (0.41–0.60), substantial agreement (0.61–0.80) and almost perfect to perfect agreement (0.81–1.00).28
RESULTS
Approximately 30,000 visits by 2822 patients were included in this analysis. The number of visits varied between severity instruments because of missing data for the various endpoints/visits. Baseline demographics and disease characteristics have been previously published.17, 18, 19, 20
ClinROs severity bands
The calculated severity band thresholds for ClinROs are shown in Figure 2. A detailed explanation of the different methods we used to calculate thresholds for severity bands (including knots and 95% confidence intervals from bootstrapped spline regression analysis) using total SCORAD as an example can be found in the Supplement. For total SCORAD, the frequencies, medians and modes identified the following severity bands with high goodness‐of‐fit and almost perfect agreement: 0 to <5 (clear), 5 to <18 (almost clear), 18 to <37 (mild), 37 to <69 (moderate) and 69 to 103 (severe) (R2: 0.997; quadratic weighted κ: 0.833) (Figure 2a, Figure S2b,c).
Analysis using means identified slightly different severity bands for all categories except Mild, with similarly high rates of fit and agreement: 0 to <7 (clear), 7 to <21 (almost clear), 21 to <38 (mild), 38 to <71 (moderate) and 71 to 103 (severe) (R2: 0.997; κ: 0.826; Figure 2a; Table S5, Figure S2a). When evaluating thresholds using modified versions of previously published bands, severity bands with higher thresholds for mild and moderate and wider almost clear and mild bands were considered (Figure S3a). All six methods used to determine severity bands for oSCORAD resulted in high R2 and substantial agreement (Figure 2b, Table S5).
Analysis of means yielded slightly lower mild and moderate thresholds than those obtained from frequencies, medians and modes: 0 to <5 (clear), 5 to <18 (almost clear), 18 to <37 (mild), 37 to <69 (moderate) and 69 to 83 (severe) (R2: 0.983; κ: 0.751). Severity bands with higher thresholds for mild and moderate and wider almost clear and mild bands were considered when evaluating thresholds using modified versions of previously published bands (Figures S3a and S4b).
For EASI, the frequencies, medians and modes identified the following severity bands with high goodness‐of‐fit and almost perfect agreement: 0 (clear), 1 to <3 (almost clear), 3 to <11 (mild), 11 to <37 or 38 (moderate) and 37 or 38 to 72 (severe) (R2: 0.991–0.992; κ: 0.805–0.806) (Figure 2c, Figure S5b). In contrast, analysis using means identified slightly different severity band thresholds for most categories and included a much larger moderate category (18 to <52; Figure 2c, Figure S5a,b). Goodness‐of‐fit was higher for this method (R2: 0.996) and agreement was substantial (κ = 0.678).
Evaluation of severity bands using versions modified from previously reported all yielded high R2 values and substantial agreement (Figure S3c).
BSA severity bands calculated using frequencies, medians and modes were quite similar to each other: 0 (clear), 1% to <5% or 6% (almost clear), 5% or 6% to <23% (mild), 23% to <74% or 80% (moderate) and 74% or 80% to 100 (severe) (R2: 0.983–0.984; κ: 0.747–0.753) (Figure 2d, Figure S6b,c). Analysis using means identified slightly different thresholds, especially for the mild and moderate severity bands, with similar agreement and fit: 0 to <2% (clear), 2% to <13% (almost clear), 13% to <29% (mild), 29% to <81% (moderate) and 81% to 100% (severe) (R2: 0.983; κ: 0.707; Figure 2d, Figure S6a). Evaluation of severity bands using a version modified from previously reported yielded high R2 values and substantial agreement (Figure S3c).
PROs severity bands
The calculated severity band thresholds for PROs are shown in Figure 3. For DLQI, the medians and modes identified generally similar severity bands with high goodness‐of‐fit and moderate agreement: 0 (excellent), 1 to <3 (very good), 3 to <10 (good), 10 to <22 or 23 (fair) and 22 or 23 to 30 (poor) (R2: 0.995; κ: 0.588 and 0.590) (Figure 3a, Figure S7b). Notably, wider bands for very good (2 to <5 and 1 to <6) were detected using frequencies and means, respectively (Figure 3a, Figure S7a,c).
Modified versions of severity bands reported previously were similar to those determined by frequencies and means; all had high R2 values and moderate agreement (Figure S8a).
For POEM, severity bands identified by means had the highest κ statistics and equal R2 values compared with analyses by frequencies, medians and modes; these bands were 0 (excellent), 1 < 6 (very good), 6 to <9 (good), 9 to <20 (fair) and 20 to 28 (poor) (R2: 0.994; κ: 0.605) (Figure 3b, Figure S9a–c). Analysis of severity bands modified from those previously reported all resulted in high goodness‐of‐fit and moderate agreement (Figure S8b).
Severity band thresholds for SCORAD Pruritus VAS varied among analyses using frequencies, means, medians and modes (Figure 3c, Table S5, Figure S10a–c). For example, the bands determined by means were 0 to <2 (none), 2 to <4 (mild), 4 to <7 (moderate), 7 < 9 (severe) and 9 to 10 (very severe) (R2: 0.999; κ: 0.422). Analysis by medians was unable to distinguish between moderate and severe, resulting in only four severity categories: 0 < 2 (none), 2 to <5 (mild), 5 to <10 (moderate/severe) and 10 (very severe) (R2: 0.995; κ: 0.340).
For SCORAD Sleep Loss VAS, analysis by means produced the highest agreement of all methods: 0 (none), 1 to <5 (mild), 5 to <7 (moderate), 7 (severe) and 8 to10 (very severe) (R2: 0.993; κ: 0.254) (Figure 3d, Figure S11a–c). Analyses with frequencies, medians and modes were unable to distinguish between severe and very severe, resulting in only four severity categories (none, mild, moderate and severe/very severe; R2: 0.986; κ: 0.232). When using results from analysis of Pruritus VAS means, similar severity band thresholds to overall means were produced.
All bands had κ coefficients that suggest fair agreement between PGADS severity bands and those identified in this analysis. The κ coefficients for SCORAD Sleep Loss VAS were the lowest compared with all other outcomes evaluated in this analysis.
DISCUSSION
In this study, severity bands for ClinRO and PRO instruments were assessed using a global clinical trial database of adults with AD. Across all measures, the five severity bands demonstrated high goodness‐of‐fit and mostly high concordance with clinical and QoL outcome measurements. Thresholds for severity bands identified using modified versions of previously reported methods were similar to or not as good as those identified by frequencies, medians and modes.
The present analyses are based on larger population sizes (2822 vs. 1 patients), greater geographic reach (including North America, Europe and the Asia Pacific region) and greater diversity of investigators, patients and disease severity levels than populations used for previous analyses.8, 9, 10, 11, 12, 13, 14 ClinROs measures were anchored to IGA, while PROs were anchored to PGADS. This is consistent with US Food and Drug Administration recommendations that patient‐reported measures should be used as anchors for PROs,29 with anchor selections from previously published bands, and with related analyses comparing relative changes in anchors and severity measures.9, 11, 13, 14, 15, 16, 30, 31
Previously published methodologies constructed potential severity bands based on unit changes in the mean, median and mode of the anchor calculated for each integer value of the instrument.9 Histograms developed during our implementation of these methodologies indicated that limiting our evaluation to one‐unit changes in the means excluded potentially meaningful information. Therefore, we utilized two additional methodologies to construct candidate severity bands: frequency cross‐over analyses and a more comprehensive means‐based methodology. Specifically, we calculated the mean of the anchor at all integer levels of the instrument and performed spline regression analysis, in which the independent variable (x) was the set of instrument integer levels and the dependent variable (y) was the corresponding set of mean anchor values.
The use of spline regressions with continuous means provides more severity threshold options compared with evaluating unit changes in means. Furthermore, unlike linear regression, using spline regression accounts for the possibility that the relationship between the measurement and anchors may be non‐linear. Therefore, our new approach using spline regression analyses to delineate severity bands in AD assessment instruments can contribute additional insights on the relationship between anchor and outcome measure.
This study utilized five disease severity bands to stratify outcomes into comprehensive and clinically meaningful categories. This is different from previously published severity strata, in which three to four categories are more common. The bands suggested for PRO outcomes in this analysis showed reasonable agreement with those previously published,11, 13 while results for ClinROs identified severity strata that differed from those previously published.8, 9, 12, 14 These differences may be derived from the different banding methodologies.
They could also reflect the different settings and populations in these studies. Investigators assessing patients with moderate‐to‐severe AD in clinical trials, as in this analysis, may perform assessments differently than clinicians assessing patients with AD with varied disease severities in a clinical setting. Interestingly, patients seem to provide similar assessments of their disease in both settings.
In addition, we present SCORAD Sleep Loss VAS severity threshold data for the first time. There are currently no published articles, suggesting severity bands for this outcome measure. The subjective effect of AD on sleeplessness as scored by the patient is an important measure that is accounted for in the total SCORAD instrument.
Up to 90% of adults with AD report sleep disturbances, however quantifying sleep quality remains a challenge.32 The exploratory SCORAD Sleep Loss VAS thresholds identified in this analysis demonstrated only fair agreement with PGADS, with the lowest κ coefficients for any outcome measure analysed. Further analyses are needed to identify threshold bands that have higher concordance for this severity measure.
The use of a data set derived from a population of patients in the clinical trial setting, rather than in a community setting, provides both strengths and limitations to this analysis. One strength is the use of a large data set of approximately 30,000 visits from approximately 2800 adult patients across 31 countries. Given the large, diverse population included in this analysis, the variety and thresholds of the severity bands identified here may be more representative than those from previous smaller studies.
Due to the inclusion criteria of the trials, which required moderate‐to‐severe disease at baseline, there were a high number of data points in the severe band categories. To address this, data from the first visit were excluded from the analysis. Data were pooled from all patients regardless of whether they received placebo or study drug, and all visits were evaluated independently.
The use of repeated visits from the same pool of patients could have contributed to the almost perfect goodness‐of‐fit observed for a majority of outcome instruments evaluated. In addition, patient assessments were conducted at regular intervals regardless of disease severity, as required in clinical trials compared with ad hoc visits, which are more common in clinical practice. While this may reduce data collection bias, these data may exhibit some variance compared to real‐world results and care should be taken in utilizing these thresholds in clinical settings until further validation.
Finally, this study evaluated disease severity in adults only and thus should not be extrapolated to paediatric patients. Future studies to support our methods and results, include the following: expanding these results to paediatric patients; validating these thresholds using data from external, real‐world cohorts; focusing on single observations of individual patients; and including patients with comorbidities or those treated with other systemic or topical drugs.
This robust analysis, utilizing a large clinical trial data set of adult patients with moderate‐to‐severe AD, demonstrated the high concordance and goodness‐of‐fit of potential disease severity thresholds across clinician‐ and patient‐reported outcome measures. The use of frequency cross‐over analysis and spline regressions to determine thresholds and the identification of severity thresholds for SCORAD Sleep Loss VAS are novel methodologies. This analysis may aid in the interpretation of AD disease severity and stratification in adult clinical trial settings.
AUTHOR CONTRIBUTIONS
All authors had access to study data and take full responsibility for the integrity of the data analysis and publication. Concept and design: Andreas Wollenberg, Eric L. Simpson, Yael A.
Leshem, Ana B. Rossi and Amy H. Praestgaard.
Acquisition, analysis or interpretation of data: Andreas Wollenberg, Eric L. Simpson, Yael A. Leshem, Alain Taïeb, Norito Katoh, Jingdong Chao, Ana B.
Rossi and Amy H. Praestgaard. Drafting of manuscript: Jingdong Chao, Ana B.
Rossi and Amy H. Praestgaard. Statistical analysis: Amy H.
Praestgaard.
FUNDING INFORMATION
Research sponsored by Sanofi and Regeneron Pharmaceuticals Inc. ClinicalTrials.gov Identifiers: NCT01859988, NCT02277743, NCT02277769, NCT02260986 and NCT02755649.
CONFLICT OF INTEREST STATEMENT
A. Wollenberg has served as an investigator for Eli Lilly, Galderma, LEO Pharma, Novartis, Pfizer, Regeneron Pharmaceuticals Inc., Sanofi and UCB; as a consultant for AbbVie, Aileens, Almirall, Anacor, Arena Pharmaceuticals, BMS, Eli Lilly, Galapagos, Galderma, GSK, LEO Pharma, L'Oréal, MSD, Novartis, Pfizer, Regeneron Pharmaceuticals Inc., Sandoz and Sanofi; and has received research grants from LEO Pharma and Pierre Fabre. E.L.
Simpson has received personal fees from AbbVie, Aclaris Therapeutics, Amgen, Arcutis, Astria Therapeutics, Attovia Therapeutics Inc., Bambusa Therapeutics Inc., Castle, CorEvitas, Dermira, Eli Lilly, Evomunne, FIDE, Impetus Healthcare, Incyte, Inmagene Biopharmaceuticals, Innovaderm Reche/Indero, Janssen, LectureLinx (LLX), LEO Pharma, Numab Therapeutics AG, Pfizer, Recludix Pharma, Regeneron Pharmaceuticals Inc., Roche Products Ltd., Sanofi‐Genzyme and Sitryx Therapeutics; and is reported to receive grants from (or serves as principal investigator for) AbbVie, Acrotech, Amgen, Arcutis, ASLAN, Castle, Dermavant, CorEvitas, Dermira, Eli Lilly, Incyte, Pfizer, Regeneron Pharmaceuticals Inc., Sanofi‐Genzyme, Target and VeriSkin. Y.A. Leshem has received honoraria as a consultant from AbbVie, Dexcel Pharma, Eli Lilly, Genentech, Pfizer, Regeneron Pharmaceuticals Inc. and Sanofi; an independent research grant from AbbVie; and has, without personal compensation, provided investigator services for AbbVie, Eli Lilly, Pfizer and Sanofi.
A. Taïeb has no relevant conflicts of interest to declare. N.
Katoh has received honoraria for serving as a lecturer for Maruho, Mitsubishi Tanabe Pharma, Sanofi and Taiho Pharma; and has received research grants from AbbVie, Eli Lilly, Kyowa Kirin, LEO Pharma, Maruho and Mitsubishi Tanabe Pharma. J.D. Chao is an employee and shareholder of Regeneron Pharmaceuticals Inc.
A.B. Rossi and A.H. Praestgaard are employees of Sanofi and may hold stock and/or stock options in the company.
ETHICAL APPROVAL
The local institutional review board or ethics committee at each study center oversaw trial conduct and documentation for all clinical trials from which data were utilized for this analysis.
ETHICS STATEMENT
All patients provided written informed consent before participating in the trial.
Supporting information