Relationship and probabilistic stratification of Eczema Area and Severity Index and objective Scoring Atopic Dermatitis severity scores for atopic dermatitis

Dear Editor, The Harmonising Outcome Measures for Eczema initiative recommended the Eczema Area and Severity Index (EASI) as the core outcome instrument for measuring the clinical signs of atopic dermatitis (AD).1 However, EASI may not have been used in previous clinical trials, and other scores – such as Scoring Atopic Dermatitis (SCORAD), the objective component of SCORAD (oSCORAD) and the Investigator's Global Assessment (IGA) – remain widely used. It is useful to establish a method to convert these scores into EASI scores to compare the results from different studies effectively. Indeed, EASI and oSCORAD have been found to be strongly correlated (Spearman coefficient r = 0·92),2 suggesting the possibility to find a relationship between the two scores. We analysed the raw data from a published study that assessed multiple AD severity scores (SCORAD, oSCORAD, EASI and IGA) for each of 43 adult patients with AD at weeks 0, 2, 4, 8, 12 and 24 of a systemic therapy3 (238 observations in total). Our data analysis confirmed that SCORAD, oSCORAD, EASI and IGA are strongly (Spearman) correlated and their relationship could be described as a monotonic function (Fig. 1a). Conversion between objective Scoring Atopic Dermatitis (oSCORAD), Eczema Area and Severity Index (EASI) and Investigator's Global Assessment (IGA) derived from the published data3 on multiple atopic dermatitis (AD) severity scores (SCORAD, oSCORAD, EASI and IGA) for each of 43 adults patients with AD at weeks 0, 2, 4, 8, 12 and 24 of a systemic therapy (238 observations in total). The data cover the entire range of severity (EASI 0–61·4, SCORAD 1–82 and IGA 0–5). All the analyses were performed using R 3·4·0 (R Foundation for Statistical Computing, Vienna, Austria). (a) Relationships between SCORAD, oSCORAD, EASI and IGA. (b) Smoothed isotonic fit between oSCORAD and EASI. (c) Cross‐validated predictions of EASI by oSCORAD (y‐axis) against the corresponding, measured values of EASI. The solid and dashed lines represent a perfect prediction and the area within one minimal clinically important difference of the perfect prediction (accuracy area), respectively. (d) Probabilistic fit between EASI and IGA. (e) Stratification of oSCORAD and EASI. The upper values correspond to the default stratification where the cut‐off is 0·5 between each category. The lower values correspond to a personalized stratification where we value sensitivity more than precision with a cut‐off of 0·1 between each category. We first derived a relationship between EASI and oSCORAD, while discarding SCORAD, as SCORAD is similar to oSCORAD but contains subjective terms not measured in EASI. EASI and oSCORAD have a Spearman correlation coefficient r = 0·93, but a linear fit does not capture their relationship appropriately (Fig. 1a). To capture the nonlinear relationship, we applied an isotonic regression and smoothed the fit with locally weighted scatterplot smoothing (Fig. 1b). The identified regression was validated using a repeated k‐fold cross‐validation (k = 4, 25 repeats). We split our dataset for the training set and the testing set to have different patients, to avoid an overestimation of the performance due to correlations of the AD severity scores within a same patient. The root‐mean‐square error for the cross‐validation was 4·77 and the accuracy, measured by the ratio of the EASI values predicted from oSCORAD being within a minimal clinically important difference (6·6 for EASI),4 was 83% (Fig. 1c). Our data analysis (Fig. 1a) also demonstrated overlaps of EASI values between different categories of IGA, suggesting a probabilistic outcome. IGA is often used to classify patients with AD into different severity categories in an anchor‐based approach.5 6 The cut‐off values of EASI separating the categories are determined to optimize a classification metric. However, condensing EASI, which ranges from 0 to 72, into IGA, with usually four to six poorly defined categories,7 in a deterministic manner results in information loss. We therefore built a probabilistic fit of EASI to p‐IGA, a probabilistic IGA represented by a continuous score (Fig. 1d), by summing up the predicted probabilities of five logistic regression models, each of which classifies IGA between 0 and 1, 1 and 2 etc. For example, p‐IGA = 2·2 means a 20% probability of IGA = 3 and 80% probability of IGA = 2. In conclusion, we have proposed methods for deriving a relationship between EASI and oSCORAD, as well as a probabilistic stratification of EASI based on IGA. The proposed methods will help interpret results of clinical trials coherently in terms of EASI, while we hope that the core outcome set for AD will be used in all future AD trials to harmonize outcomes. The derived relationship between EASI and oSCORAD is valid only for a given accuracy (83%) evaluated by cross‐validation, assuming that the data used here are taken from representative patients with AD. The proposed computational methods are available as an app (www.bg.ic.ac.uk/research/r.tanaka/adconversionapp.html) to convert and visualize oSCORAD and severity strata by IGA in light of EASI. They can be further applied to data from different populations of patients with AD to confirm the preliminary results described here. Funding sources: the British Skin Foundation, grant 021/SG/17. Conflicts of interest: none to declare.