P008 Can artificial intelligence-powered EczemaNet facilitate whole-person remote eczema assessments using digital images?

Abstract An automated remote assessment tool using digital images could standardize evaluations of atopic eczema. EczemaNet (www.eczemanet.co.uk) is such a tool developed for this objective, demonstrating reliability for assessing eczema signs from individual images. However, the optimal method for combining severity assessments from multiple images into a whole-person severity score remains unclear. This study explored different approaches for integrating data from multiple images into a whole-person severity score. The EczemaNet study (IRAS 325468) recruited children and adolescents aged < 18 years with eczema from three UK National Health Service trusts. To train the tool, trained research staff took digital images of eczema-affected skin in four body regions (head and neck, body, arms, and legs) and graded each digital image on a 0–3 scale for four clinical signs (erythema, oedema, lichenification, excoriation) based on in-person assessments. Assessors also collected eczema extent using an online manikin with 45 body regions, Eczema Area and Severity Index (EASI) and validated Investigator’s Global Assessment (vIGA). Several approaches to integrate individual image scores with extent data to calculate a ‘per person’ severity score were then evaluated. These included approaches that mimicked EASI (n = 2) and approaches that mimicked IGA (n = 3). These approaches were correlated with EASI using Spearman’s correlation and compared with vIGA using weighted kappa statistics. Exploratory analysis (from 52 patients; 920 images with a median of 15 images per person) was used to determine the most promising thresholds for calculating the per person scores and establish which approaches were most likely to be comparable with EASI and vIGA. In the subsequent validation dataset (152 patients; 2342 images in total with a median 11 images per person), good correlation with EASI was demonstrated for all approaches. Approaches that mimicked EASI had correlation coefficients of 0.93 and 0.92, respectively, while the vIGA-mimicking approaches had correlation coefficients of 0.92, 0.82 and 0.82, respectively. Similarly, all approaches had substantial agreement with vIGA. Approaches that mimicked EASI had kappa statistics of 0.72 and 0.73, respectively, while the IGA-mimicking approaches had kappa statistics of 0.71, 0.72 and 0.72. Our study suggests that it is possible to create a per person eczema severity assessment using clinical images combined with a measure of extent that has good correlation with EASI and agreement with vIGA. Additional feasibility work and validation work are required.