AI08 Validation of the EczemaNet artificial intelligence-based eczema severity tool in a paediatric population

Abstract

Accurate assessment of disease severity is essential for monitoring atopic eczema and guiding triage and treatment decisions. Traditional clinician-based scoring systems, such as Eczema Area and Severity Index (EASI) and Investigator Global Assessment (IGA), are resource intensive and require face-to-face assessment, limiting scalability for remote monitoring. EczemaNet is an artificial intelligence (AI)-based eczema severity assessment tool to generate severity scores that correspond to EASI and IGA from images. EczemaNet was trained using data collected from nearly 300 children with eczema. The aim of the study was to assess the consistency between EczemaNet-generated and clinician-­assessed EASI and IGA scores in children. We conducted a validation study for EczemaNet by collecting 1979 clinician-taken images from 166 children attending secondary care paediatric allergy and dermatology clinics. We assessed agreement between EczemaNet-generated scores and clinician-assessed EASI using intraclass correlation coefficients. Agreement with IGA and EASI severity strata was assessed using weighted kappa (κw), and correlation using Spearman’s coefficient. Overall, 56.6% of patients (n = 94) had darker skin tones (Fitzpatrick IV–VI). Over 80% of patients had sufficient data to generate an EASI-mimicking score, and > 90% had adequate data to generate an IGA-mimicking score. The AI-EASI method showed excellent agreement with clinical EASI scores (intraclass correlation coefficient = 0.90), good agreement with IGA (κw = 0.73) and very good agreement with EASI severity strata (κw = 0.88). The AI-IGA method showed comparable performance, with excellent correlation with clinician-­assessed EASI scores (r = 0.95) and good agreement with IGA (κw = 0.72). Sensitivity analyses suggested that Fitzpatrick skin type did not significantly impact model performance; however, agreement was reduced in patients with more severe disease. Using ­clinician-provided images, EczemaNet demonstrates excellent agreement with EASI and good agreement with IGA in a real-world paediatric population. These findings support its potential as a scalable tool for remote disease monitoring, clinical decision support and research, pending improved performance in cases of severe disease.