Artificial intelligence (AI)-based tools could support the diagnosis, monitoring and management of atopic dermatitis (AD), but limited external validation, unrepresentative datasets, and a lack of testing within clinical pathways currently restrict their translation into practice, according to a recent scoping review.
AI-driven tools are being developed to diagnose AD, assess disease severity, monitor symptoms and predict disease activity. But how ready these tools are for clinical use remains uncertain.
As such, a scoping reviewpublished in the journal Clinical and Experimental Dermatologymapped the available research on AI-driven digital tools for AD, assessed study methodologies and identified barriers to their real-world implementation.
The authors searched the MEDLINE, Embase, Web of Science and Scopus databases from their inception to December 2024. Eligible publications included primary research evaluating diagnostic aids, symptom-tracking applications, predictive models,AI-supported teledermatology systemsand language-based tools.
Two reviewers independently screened the records and extracted data, with 52 studies ultimately included. As some tools had multiple functions, they were assigned to multiple categories.
Diagnostic performance and AI-driven tools in AD
Diagnostic tools were the most frequently studied application, while predictive models were also prevalent. Many diagnostic systems used convolutional neural networks to analyse clinical or dermoscopic images and distinguish AD from other skin conditions or healthy skin.
Reported diagnostic accuracy frequently exceeded 90%, although the performance measures, datasets and comparators varied between studies. Training datasets ranged from 174 to 309,000 images, and several studies did not report the sizes of their training and test datasets separately.
One deep-learning severity system achieved accuracy exceeding 85% when assessing individual components of the Eczema Area and Severity Index.
However, laboratory performance did not always carry over into real-world settings. One sensorised glove identified scratching with 94% accuracy under laboratory conditions, but sensitivity fell to 40% and precision to 6% in real-world or paediatric environments.
Among the teledermatology studies, one system was evaluated using more than 1,000 retrospective cases. AI assistance increased agreement with a dermatologist panel from 48% to 58% among primary care physicians and from 46% to 58% among nurse practitioners.
Only two studies evaluated natural language processing or large language model applications. These included a system that extracted AD-related information from electronic health records and an image-and-language model that generated diagnostic suggestions and treatment recommendations.
Limited external validation and representative datasets
Although 96% of studies reported methodological transparency and data partitioning, only 21% externally validated their models, with the same proportion making their code available. Just 35% evaluated a model at its intended stage of the care pathway.
Only 38% used more than one dermatologist or clinical expert to label
Skin colour was reported by 27% of studies and fewer than half included basic demographic information such as age, sex or disease severity.
The authors highlighted evidence of reduced diagnostic performance in darker skin tones and noted that two teledermatology studies did not represent the full range of skin tones, including Fitzpatrick skin types V and VI.
As a scoping review, the study did not formally assess publication bias. Non-English publications and conference abstracts were excluded, and the findings reflect literature published only up to December 2024. Heterogeneity in methods and performance measures also prevented direct comparisons between tools.
The authors concluded that AI applications are intended to complement rather than replace clinical expertise. They called for external validation, more representative datasets and evaluation within clinical settings before these technologies are implemented more widely in AD care.
Reference
Yip A et al.AI-driven digital tools for atopic dermatitis: a scoping review. Clin Exp Dermatol 2026;11 August: llag353.
