Artificial intelligence—enabled precision medicine for inflammatory skin diseases

Abstract

Recent advances in artificial intelligence (AI) and multimodal data collection are revolutionizing dermatology. Generative AI and machine-learning approaches offer opportunities to enhance the diagnosis and treatment of autoimmune and inflammatory skin diseases, including atopic dermatitis, psoriasis, hidradenitis suppurativa, vitiligo, alopecia areata, and rheumatic skin disease. This review examines the current landscape of AI applications for inflammatory skin diseases and explores how generative AI and machine-learning methods can advance the field through deep phenotyping, characterization of disease heterogeneity, drug discovery, precision medicine, and delivery of clinical care. We discuss the promises and challenges of these technologies and present a vision for their integration into clinical practice.

INTRODUCTION

The increasing availability of comprehensive patient information, spanning multimodal imaging, multiomic profiles, and real-world clinical data is driving a significant transformation in health care. Artificial intelligence (AI)—based computational methods have emerged as a powerful tool to analyze these complex datasets and extract clinically meaningful insights, enabling advances in disease diagnosis, treatment planning, and clinical decision support (Khera et al, 2023; Saab et al, 2025; Tu et al, 2025). These technological advances are particularly timely for inflammatory skin diseases, as our understanding of the immune system and inflammatory pathways continues to evolve and enable targeted therapies (Hawkes et al, 2017; van der Schaft et al, 2019). Given the complex pathophysiology and heterogeneity of inflammatory skin diseases, AI offers vast opportunities for applications ranging from understanding precise disease mechanisms to aiding in patient communication.

As a brief primer, AI refers to the general paradigm of computational systems to mimic human intelligence, whereas machine learning (ML), a subset of AI, refers to approaches for computer programs to learn from data for specific tasks. Generative AI encompasses approaches that can create or simulate information. Various learning paradigms exist, with the most common being supervised learning, in which a model learns from data that have been previously labeled with a known outcome (eg, diagnosing psoriasis, classifying atopic dermatitis [AD] severity). In contrast, the unsupervised learning paradigm enables models to discover structures or patterns in the data, enabling identification of novel disease subtypes based on molecular or clinical profiles without pre-conceived classifications (eg, identifying rosacea morphological subtypes). Intermediate approaches that integrate features of both supervised and unsupervised learning include semi-supervised (using small amounts of labeled data and large amounts of unlabeled data) and self-supervised (learning from unlabeled data with learned labels) learning. Finally, the reinforcement learning paradigm enables models to learn from trial and error (eg, optimizing alopecia areata [AA] patient education on the basis of feedback).

These paradigms are executed by various model architectures. Dermatology, with its strong reliance on visual cues and images, has flourished along with advances in computer vision and image analysis, employing architectures for image-based tasks. For years, convolutional neural networks (CNNs) have been a standard choice for skin diseases, particularly with application of skin cancer diagnosis. A prominent demonstration of this was in 2017, when deep CNNs achieved performance comparable with that of dermatologists in classifying images of skin lesions (Esteva et al, 2017). A more recent architecture, the transformer, enables incorporation of both local detail and global context in data and can be utilized as generative AI to create synthetic content mimicking the data on which they are trained. Transformer models are based on the idea of “self-attention,” in which they learn context and meaning by tracking relationships in sequential data such as words in a sentence. These transformers power current state-of-the-art models for images and text, including large language models (LLMs) and vision transformers (ViTs). A foundation model refers to a model trained on a large dataset that can then be adapted for a specific task. Some recent relevant examples include Articulate Medical Intelligence Explorer (AIME) for differential diagnosis (McDuff et al, 2025), MONET for skin image annotation (Kim et al, 2024), and PanDerm for various applications from lesion segmentation to melanoma prognostication (Yan et al, 2025). Recently, there is a rise of agentic AI—which incorporates multiple autonomous systems capable of executing tasks independently—that enables greater task complexities (Zou and Topol, 2025). Although many existing tasks involve applications in skin cancer and melanoma, there is currently an evolving landscape for applications in inflammatory dermatology and beyond, with potential for personalized pathway targeting and management. Further multimodal data and applications include dermatopathology classification, teledermatology triaging, and beyond (Cazzato and Rongioletti, 2024; Muhaba et al, 2022; Soenksen et al, 2021).

In this paper, we comprehensively review current applications of ML and AI in inflammatory skin disorders (Figure 1). We first review broad applications in inflammatory skin diseases, including disease classification and prediction, deep phenotyping, biomarker discovery, therapeutic design, and clinical care (Figure 1). Then, we discuss disease-specific applications, including but not limited to AD, psoriasis, acne, rosacea, hidradenitis suppurativa (HS), vitiligo, AA, cutaneous lupus, dermatomyositis (DM), and morphea/scleroderma (Table 1 and Supplementary Table S1). Finally, we explore present challenges and future opportunities for future diagnostic and therapeutic development and their potential for clinical implementation.

Figure 1. Landscape of current and future applications of AI to advance precision medicine for inflammatory skin diseases.

This figure summarizes the landscape of AI in inflammatory skin disease, starting from inflammatory dermatoses at the top. Data from external and internal factors encompassing various modalities are shown in the subsequent row. The raw data may undergo preprocessing, such as segmentation and quality control, to prepare for subsequent tasks. The left column reveals the range of AI approaches and algorithms for various applications in the figure discussed in this review. Classification applications utilize known labels or outcomes for a specific task, such as diagnosis, prognosis, or prediction of disease severity. Exploratory tasks may include using unsupervised learning to identify features (or an attribute) that are directly or indirectly quantified (eg, latent features). This may then lead to iterative improvement in a supervised ML task or aid in discovery tasks (eg, cell surface marker associations, cytokine associations) and precision medicine approaches (eg, patient subtyping based on receptors). In silico means through computer modeling or simulation. The bottom rows show how generative AI can integrate with research and clinical care to improve patient phenotyping and development of new treatments for inflammatory skin disease. AI, artificial intelligence; EASI, Eczema Area and Severity Index; ML, machine learning; SALT, Severity of Alopecia Tool.

Table 1.

AI-Enabled Advancements across the Spectrum of Inflammatory Skin Diseases

Diseases	Application
	Classification and Diagnosis	Phenotyping, Subtyping, and Treatment Personalization	AI for Workflows and Patient Care
Atopic dermatitis	13	16	14
Psoriasis	14	14	13
Rosacea	4	10	3
Acne	4	3	3
Hidradenitis suppurativa	3	3	0
Alopecia areata	3	4	0
Vitiligo	2	2	0
Cutaneous lupus erythematosus	3	5	0
Morphea or scleroderma	1	4	1
Dermatomyositis	1	3	0
Other or multiple	23	23	20

Abbreviation: AI, artificial intelligence.

Presented are numbers of reviewed studies grouped by skin disorder in rows and grouped by application in columns. Studies are outlined in Supplementary Table S1.

METHODS

To identify relevant articles for this review, we conducted a comprehensive literature search of PubMed and Google Scholar, which included peer-reviewed manuscripts, preprints, and conference papers. The search was supplemented by reviewing references of identified articles and following “similar article” suggestions. The search utilized keywords including a combination of AI paradigms or models (eg, ML, AI, generative AI, foundation model, deep learning), applications (eg, prediction, classification, subtyping, clustering, phenotyping, biomarker discovery, drug repurposing), and inflammatory skin diseases (eg, AD, HS, vitiligo, AA, cutaneous lupus, etc). Final inclusion in this manuscript was based on a manual evaluation by the authors of each individual publication’s quality and relevance to the topic.

AI APPLICATIONS IN INFLAMMATORY DERMATOSES

Classifying lesions and diagnosing inflammatory skin conditions

Although early AI applications and advancements in dermatology have primarily focused on melanoma detection (Choy et al, 2023; Esteva et al, 2017; Young et al, 2021), recent developments have expanded to encompass a greater diversity of skin diseases, including lesions of inflammatory skin disorders such as acne, psoriasis, eczema, rosacea, vitiligo, and other chronic inflammatory conditions. Multiple publications have focused on single-disease or multidisease classification on the basis of clinical images, dermatoscope images, and teledermatological images (Huang et al, 2021a; Li Pomi et al, 2024; Liu et al, 2020; Muñ oz-López et al, 2021; Prabhu et al, 2018¹; Wang et al, 2021b; Zhu et al, 2021b) (Table 1 and Supplementary Table S1 [first column]). Classification for both low- and high-level processing tasks may be utilized in the approach to inflammatory skin lesion identification (Choy et al, 2023; Li et al, 2023b). Low-level application includes processing of raw images or segmentation of skin lesions to highlight areas of interest for further analytics. As an example, Czajkowska et al (2022, 2021) demonstrated an automated framework for segmenting the epidermis from high-frequency ultrasound images in inflamed skin). High-level classification includes determining a diagnosis or directly predicting management, with an example from Abhishek et al (2021) demonstrating better performance with direct management prediction than inferring management through the prediction of diagnosis first. Beyond traditional images, the scope of diagnostic classification has expanded to accommodate a greater variety of input data sources, such as histological slides (Hosny et al, 2020; Muhaba et al, 2022) to nonimage data such as gene expression profile sequencing from tape strips, Raman spectroscopy, and electronic health records (EHRs) (Greenfield et al, 2023; Gustafson et al, 2017; He et al, 2021; Jiang et al, 2022) (Figure 1). Generative AI may be utilized for image enhancement and data augmentation to address dataset limitations such as low sample size and artifacts such as hair and ink markings.

Advances in diagnosis include the fine tuning of existing models to enable efficient learning on smaller datasets. By leveraging baseline knowledge from models pretrained on large datasets (such as ImageNet), fine-tuned models do not have to learn basic visual features from scratch (such as edges, textures, or shapes). Often, lower-level tasks such as edge detection and segmentation are incorporated within deep learning architectures. Hosny et al (2020) demonstrated the specialization of AlexNet for classifying 7 predefined skin lesion types such as melanoma, dermatofibroma, and vascular lesions. This approach to AI specialization is valuable for future approaches in inflammatory dermatoses given less need for extensive new datasets or computational resources.

A particularly useful development is the emergence of interpretable AI that allows for clinicians to scrutinize how models are making decisions. Approaches include those that probe into attention networks and gradient-based saliency maps, such as class activation mapping (CAM), to highlight and enhance our understanding of specific features such as morphology and pigmentation. Mohan et al (2025) showed improved performance of self-supervised transformer approaches such as DINOV2 in classifying a wide array of skin diseases and utilized techniques such as Grad-CAM and SHAP for model interpretation. Corbin and Marques (2023) utilized Grad-CAM to assess skin-type bias and showed improved fairness metrics with joint regularization and transformer-based synthetic data blending. This integration of transparency is crucial for building clinical trust and can also offer insights into disease morphology from the model’s perspective. These classification and prediction approaches are particularly useful for primary care referral decisions or prognostication for clinical decision support for both general practitioners and dermatologists.

Phenotyping, subtyping, and treatment personalization

In addition to improving diagnosis, ML methods can serve to expand our understanding of inflammatory skin diseases, enabling applications such as disease characterization, biological hypothesis generation, and precision treatment approaches (Félix Garza et al, 2019; Tang et al, 2024). Biomarker identification and mechanistic insights can be elucidated with evolving computational approaches that integrate diverse data types, such as genomics, transcriptomics, cell surface markers, and clinical phenotypes, to enable the understanding of patterns and relationships across modalities (Clayton et al, 2021; Jiang et al, 2022; Wang et al, 2024c). This can be valuable in complex conditions such as systemic lupus erythematosus (SLE), scleroderma, morphea, and other autoimmune disorders, where target inflammatory pathways and subtypes have been identified that may reflect underlying patient or molecular heterogeneity and can help inform targeted treatment or prognosis (Andreoletti et al, 2021; Choi et al, 2023; Cutts et al, 2024). Clustering algorithms at either the data level or patient level can enable the characterization of disease subtypes based on variations in molecular mechanisms, treatment response, and prognosis (Table 1 and Supplementary Table S1 [second column] and Figure 1) (Berna et al, 2020; Cazzaniga et al, 2021; Neely et al, 2022). On a broader level, integration with population-level information can elucidate risks from external exposures (eg, pollution, humidity) as well additive risks due to comorbidities (eg, diabetes mellitus, mental health disorders). For example, there have been studies that identify external or comorbid factors influencing irritant dermatitis and HS (Fortino et al, 2020; Hua et al, 2021; Papa et al, 2023). Currently, these methods enable researchers, epidemiologists, and clinicians to try improving the understanding of disease mechanisms.

Efforts on treatment are also underway to study the integration of protein—drug interaction networks that can further accelerate therapeutic target identification and drug repurposing efforts. For instance, Patrick et al (2018a) used a word-embedding-based ML approach to identify candidate drugs for AD and psoriasis on the basis of enriched genes from lesional skin transcriptomic data. Mathematical models of biological interactions have also been employed to increase the understanding of biological processes and drug response (Barraza et al, 2024; Kardynska et al, 2023). These approaches analyze drug response pathways to predict treatment efficacy for specific disease subtypes, whereas transcriptomic analyses can provide deeper insight into disease mechanisms as another approach to identify targets for drug development (Yamanaka et al, 2023) (Table 1 and Supplementary Table S1 [second column]). Moreover, AI approaches are also investigated at the stage of clinical trial design and treatment selection through patient stratification based on disease severity, subtypes, or predicted responses (Huang et al, 2023; Hurault et al, 2020). The potential for generative AI has also been proposed for applications such as biomarker identification and drug design (Shanehsazzadeh et al, 2023²; Ying et al, 2024³). The generative models enabling these advancements vary widely, encompassing approaches from foundational physical equations and probability distributions to sophisticated neural network architectures such as transformers (including LLMs), generative adversarial networks, and diffusion models (Figure 1 [left column]). The integration of various learning paradigms and diverse data types enables the discovery of complex relationships among biological mechanisms, environmental exposures, and clinical manifestations. Although exploratory, the potential to identify targeted therapeutic candidates (eg, IL-4 inhibitors, Jak inhibitors) can support the advancement of precision medicine strategies that target disease-relevant pathways in inflammatory skin diseases. These applications are still at the beginning stages, with gaps and opportunities for further expansion.

AI as a clinical aide for physicians and patients

AI is beginning to reshape clinical workflows for clinicians and care delivery for patients, with new uses in dermatology emerging. Across medical specialties, the development of AI scribes and clinical summarizers can enhance clinical efficiency by automating documentation and supporting workflows (Miao et al, 2025; Williams et al, 2024). The accelerated adoption of telehealth after the COVID-19 pandemic also enhanced teledermatology advancements, spurring investigations into how generative AI can aid with virtual consultations, enhance remote image assessment, and facilitate accurate documentation (Shapiro and Lyakhovitsky, 2024; Vodrahalli et al, 2023, 2021) (Figure 1). This technological advancement holds significant promise for expanding access to dermatological expertise, especially for patients in underserved or remote areas (Muñoz-López et al, 2021).

AI is also helping clinicians with image interpretation and clinical reasoning. ViT models are also being developed to generate explanations or interpretations of dermatological images directly (Lin et al, 2025; Mohan et al, 2025; Zhou et al, 2024a). For example, SkinGEN from Lin et al (2025) proposes a framework for combining vision language models with generative AI to visually explain dermatological diagnoses. Kim et al (2024) introduced an image—text foundation model, MONET (medical concept retriever), trained on dermatology image—text pairs from literature with capability of annotating both clinical and dermatoscope images. Yan et al (2025) introduced PanDerm, a multimodal vision foundation model with various abilities, including providing a differential diagnosis that incorporated inflammatory skin diseases and predicting disease progression. Recently, AI has been shown to be capable of clinical reasoning and able to provide differential diagnosis that is noninferior to that of expert dermatologists and with incorporation of visual data, as seen from Google’s AIME that can analyze smartphone skin images (Saab et al, 2025⁴). Although AI applications to aid clinicians in dermatology care are promising, applications in inflammatory skin diseases remain an area with significant opportunity (Table 1 and Supplementary Table S1 [right column]).

In terms of applications in patient education, generative AI can improve communication by aiding in translation or creating accessible explanations of medical information, such as pertaining to chronic management of inflammatory dermatoses or side effects from steroids (Aydin et al, 2024; Johri et al, 2025). AI-generated translations of medical notes and reports aim to make complex information more understandable for patients. For example, Zhang et al (2024b) assessed ChatGPT-4’s ability to translate hypothetical dermatopathology reports into patient-friendly language. Although physician raters found the translations generally favorable, they also noted occasional errors for longer, more complex reports (Zhang et al, 2024b). One study also evaluated the utility of LLMs in answering questions from patients, which can be useful to guide patients because there is an increasing prevalence of LLM use for health information among lay people (Mendel et al, 2025). These studies are still in the beginning stages, and dedicated research and development will be needed to harness its potential for complex challenges provided by inflammatory skin diseases.

DISEASE-SPECIFIC APPLICATIONS OF AI

AD

Classification and diagnosis.

Recent approaches extend image-based CNN models (Hammad et al, 2023; Wu et al, 2020), with newer applications emerging that synthesize multimodal data to advance the understanding of AD. For instance, a supervised ML classifier for AD based on gut transcriptomics and microbiome identified 35 predictive genes (including GRP1, CCL22, TTC27) and 50 microbiota features (such as Akkermanisia spp, Verrucomicrobia spp, Propionibacterium spp) that help distinguish patients with AD from healthy controls (Jiang et al, 2022). AD classification models with microarrays identify genes related to epidermal development, keratinocyte function, immune responses, and pyroptosis (Ghosh et al, 2015; Wu et al, 2023). Besides distinguishing AD from healthy skin, classification models have been applied to differentiating AD from psoriasis. Liu et al (2022b) used single-cell RNA sequencing and cellular indexing of transcriptomes and epitopes by sequencing on CD45+ immune cells, highlighting genes linked to T helper 2/T helper 17 balance, especially in T-resident memory cells. He et al (2021) similarly utilized RNA sequencing from tape strips to distinguish between AD, psoriasis, and controls, identifying improved classification with dendritic and T-cell markers (CD3, ITGAX/CD11c, CD83), downregulated genes involved with terminal differentiation (FLG2, LCE5A), genes with T helper 2 skewed in AD tissue (IL13, CCL17/TARC, CCL18) or T helper 17/T helper 1—related expression in psoriasis (IL-17A, IL-36A, IFN-γ, CXCL9), and expression of nitric oxide synthase 2 and inducible nitric oxide synthase (He et al, 2021). Overall, the models confirm a skewed T helper 2 profile in AD versus a T helper 17/T helper 1 profile in psoriasis. These studies show how ML models can illuminate novel biological pathways involved in the pathogenesis of AD and how they are similar or distinct from other inflammatory skin diseases.

Models trained on EHR data, which include least absolute shrinkage and selection operator (LASSO) logistic regression and natural language processing or transformer-based patient representation with XGBoost model, have successfully classified AD (Gustafson et al, 2017; Wang et al, 2024b). Other promising modalities for classification include handheld confocal Raman spectroscopy, with one study classifying patients with AD with spectroscopic signatures that depict stratum corneum proteins and lipids (Dev et al, 2022) and another that can assess AD severity on the basis of measured water, ceramide, and urocanic acid content (Ho et al, 2020). With regard to environment-related risk, Huang et al (2021b) conducted a 14-year follow-up birth cohort study involving 1439 mother—infant pairs with models trained on combined demographic, environmental (eg, air pollutant), and allergen data and noted prenatal exposure to nitrogen dioxide as a significant predictor of prenatal AD.

Phenotyping, subtyping, and treatment personalization.

Approaches to identify biomarkers and subtypes of AD are performed with the goal of analyzing personalized treatment approaches and responses. AD severity models include a Bayesian inference model using prior AD flares and treatment response (Hurault et al, 2020) and a gradient boosting and regression model from clinical data, identifying severity-associated factors such as later onset, total IgE levels, male sex, and atopic stigmata (Maintz et al, 2021). Wearable biosensors can assess AD severity and treatment response by integrating various technologies: actinography to quantify itch, sensors for physical properties such as skin capacitance and temperature, and surface imaging and chemical sensors that capture inflammatory markers such as changes in vascularity and pH (Greenfield et al, 2023; Khan et al, 2024; Yang et al, 2025).

Besides AD severity phenotypes, unsupervised clustering has enabled identification of subphenotypes that associate with disease clearance and enable personalization of treatment and prediction of disease outcomes. Berna et al (2020) clustered baseline clinical data from over 8000 children in the Pediatric Eczema Elective Registry and identified 5 distinct AD subphenotypes, defined by factors such as race, age of onset, and allergy profiles, with statistically significant differences in long-term disease clearance. Another study trained models on EHR and claims data to predict dupilumab nonresponse, identifying factors such as concomitant ibuprofen use and higher comorbidity scores as predictors for nonresponse at 6 months, offering potential flags for closer monitoring or alternative management approaches (Wu et al, 2022a).

Further efforts at treatment personalization involve targeted drug prediction and development. For instance, Clayton et al (2021) clustered 406 skin transcriptome samples from patients with AD, identifying 3 unique keratinocyte activation programs and therapy-dependent modifications. For instance, they identified that cyclosporine primarily affected keratinocytes responding to IL-17A, whereas dupilumab specifically reversed the IL-4— and IL-13—responsive keratinocyte profile (Clayton et al, 2021). With regard to targeted drug development, Wang et al (2022) combined network analysis, deep learning, and molecular simulation, confirming TNFα and IL-4 as key disease targets and further identifying caffeoyl malic acid as a dual target inhibitor for the treatment of AD. A couple of other efforts in AD include computational approaches for drug repositioning, particularly targeting pathways such as Jak—signal transducer and activator of transcription (STAT), immune factors, or inflammatory cytokines beyond IL-4/IL-13 (Prakash et al, 2020; Prasannanjaneyulu et al, 2022). These studies have led to the clinical development of upcoming therapies, including biologics targeting IL-31 and OX-40/OX-40L, many of which are undergoing clinical trials.

AI as a clinical aide for physicians and patients.

As the daily management of chronic inflammatory skin diseases extends beyond the clinic, LLMs are increasingly playing a role in medical care. Two studies assessed LLM responses to common questions around AD, overall finding answers to be mostly acceptable but otherwise incomplete or deviated from evidence-based medicine (Lakdawala et al, 2023;Sulejmani et al, 2024).

Summary.

AI applications in AD are moving beyond image-based classification to incorporate multimodal data for more nuanced diagnosis and analysis. Initial applications leverage supervised learning paradigms with image-based CNN models for lesion identification, diagnosis, and severity determination. Subsequent models aim to diagnose AD with data from transcriptome, microbiome, clinical data, or spectroscopy. A few models that are specifically trained to distinguish AD from psoriasis with gene expression highlight T helper 2 skew in AD. Unsupervised learning paradigms aim to stratify patients on the basis of severity or treatment response. In clinical care, generative models are beginning to find use as patient-facing educational aides.

Psoriasis vulgaris

Classification and diagnosis.

Multiple image-based applications have been applied to psoriasis that range from segmentation to diagnosis. As an example low-level task, Shrivastava et al (2017) developed a Bayesian model for segmenting psoriatic lesions from clinical images by classifying each pixel, followed by secondary extraction of color, entropy, and light intensity to determine psoriasis severity. On a higher level, multiple studies extend deep learning CNN architectures to distinguish psoriasis from controls (Aijaz et al, 2022; Chakraborty et al, 2024; Zhao et al, 2020), determine PASI (Huang et al, 2023; Okamoto et al, 2022; Schaap et al, 2022; Shrivastava et al, 2016), or predict severity changes over time (Moon et al, 2024). Applications also include distinguishing psoriasis from potential mimics, such as AD, using images (Wu et al, 2020) and gene expression datasets (He et al, 2021; Liu et al, 2022b; Seremet et al, 2024) or distinguishing scalp psoriasis from seborrheic dermatitis using dermoscopic images (Yu et al, 2022). Beyond diagnostics, ML applications in psoriasis care include prediction of complications such as psoriatic arthritis, plaque burden, or treatment response (Yu et al, 2020). Patrick et al (2018b) utilized genetic markers to predict the risk of developing psoriatic arthritis versus cutaneous-only psoriasis. Another prospective study on 158 patients with psoriatic arthritis demonstrated that random forest models can predict minimal disease activity at subsequent visits with strongest predictors of global pain, Psoriatic Arthritis Impact of Disease (PsAID) score, patient global assessment of disease, and Health Assessment Questionnaire Disability Index (Queiro et al, 2022).

Phenotyping, subtyping, treatment personalization.

AI with molecular datasets is driving advancements in psoriasis phenotyping and treatment response. For biomarker discovery, Zhou et al (2024b) utilized LASSO and SVM-RFE ML models to identify 4 potential diagnostic biomarkers of psoriasis (UGGT1, CCNE1, MMP9, and ARHGEF28), with experimental validation supporting CCNE1 as a key candidate owing to its elevated expression in psoriatic epidermis and relevant immune cells (CD4+ T cells, neutrophils). In terms of molecular subtyping, multiple approaches utilize unsupervised learning paradigms to identify potential subtypes in psoriasis that may differentially respond to treatment. Zhang et al (2024a) clustered skin transcriptomes identifying 4 distinct psoriasis subtypes (immune activated, stromal, and intermediate) with different predicted treatment responses: the immune and intermediate subtypes favored methotrexate or anti—IL-12/23 therapies, whereas the stromal subtype appeared more responsive to anti-TNFα/anti—IL-17RA agents. Li et al (2023a) similarly identified 2 psoriasis subtypes with distinct immune infiltration profiles and potential candidate treatment: cluster 1 with higher activated CD4+ T cells and proposed glycogen synthase kinase inhibitor treatment and cluster 2 with higher plasmacytoid dendritic cell and neutrophils with proposed mTOR and phosphoinositide 3-kinase inhibition. Another study that clustered word embeddings identified top 5 drug repurposing candidates that include budesonide, hydroxychloroquine, leflunomide, mesalamine, and cyclophosphamide (Patrick et al, 2018a). Unsupervised learning paradigms also investigate psoriasis treatment—specific responses. Damiani et al (2020) utilized unsupervised artificial neural networks on baseline clinical and blood count data to predict responders to secukinumab, whereas Pournara et al (2021) utilized finite mixture models on patients with psoriatic arthritis to identify 7 clinical clusters that show variable response trajectories to different secukinumab dosing. Tomalin et al (2020) developed models with bagging and ensemble methods on blood protein data to predict 12-week treatment response for tofacitinib and etanercept, finding predictive proteins, including IL-17A/C.

AI as a clinical aide for physicians and patients.

Within clinical workflows, Klçoğlu et al (2025) assessed ChatGPT’s responses to frequently asked questions about psoriatic arthritis, finding that treatment-related answers scored lower than symptom-related ones. Other applications of generative AI in psoriasis are limited.

Summary.

Similar to AD, psoriasis vulgaris has been a focus of AI research, particularly with regard to lesion segmentation, disease classification, and assessment of disease severity (eg, PASI, PsAID). Most approaches utilize supervised learning paradigms with CNN deep learning models to analyze images. Recent advances focus on treatment personalization based on selection of targeted therapies (eg, anti—IL-12 vs anti—IL-17) or predicting therapeutic outcomes (eg, secukinumab response).

Rosacea

Classification and diagnosis.

A few studies developed image-based deep learning models to classify rosacea, such as RosNet (adapted from Inception-ResNet-v2) and FACES (an ensemble deep learning model) (Binol et al, 2020; Park et al, 2023). Some deep learning models are trained to differentiate rosacea from potential mimics (acne, seborrheic dermatitis, eczema) and classify rosacea subtypes (erythematotelangiectatic, papulopustular, and rhinophymatous), with one study highlighting dermoscopic features such as vascular polygons and yellow and red halos to differentiate rosacea from other facial inflammatory diseases (Ge et al, 2022; Zhao et al, 2021).

Phenotyping, subtyping, and treatment personalization.

Current approaches leveraging AI to understand rosacea pathogenesis includes investigating molecular relationships and shared diagnostic phenotypes. From genetics data, Gao et al (2025) utilized bidirectional 2-sample Mendelian randomization analysis, finding a potential causal link between rosacea and cardiovascular disease, whereas Deng et al (2023) identified variants in LRRC3, SH3PXD2A, and SLC26A8, with subsequent functional analysis supporting a neurogenic inflammation mechanism in rosacea. Wang et al (2024a) investigated immune mechanisms linking rheumatoid arthritis and rosacea, identifying and validating CXCL19 and CCL27 as shared biomarkers.

Unsupervised clustering methods have been utilized to identify molecular endotypes of rosacea. From transcriptomics data, Mao and Li (2024) identified 2 rosacea subtypes (neurogenic and inflammatory), and a subsequent linear regression model on 4 genes (PNPLA3, CUX2, PLIN2, HMGCR) can diagnose neurogenic rosacea with an area under the curve (AUC) of 0.90 on a validation cohort. With respect to drug design for rosacea, Barraza et al (2024) identified downregulated SKAP2 and upregulated S100A7A as potential disease-related targets, which they then confirmed through computer simulations, showing stable binding with existing rosacea drugs (eg, isotretinoin, permethrin) as well as possible repositioned drugs (eg, dasatinib, lovastatin).

AI as a clinical aide for physicians and patients.

Unsupervised learning paradigms have been utilized to understand patient topic categories from online forums (Rajalingam et al, 2023) as well as clinical practice patterns for medication prescriptions (Nicholas et al, 2024). No related generative AI publications were found for rosacea.

Summary.

AI applications in rosacea have primarily focused on diagnosis and distinguishing from clinical mimics, particularly through deep learning models for image analysis. Recent work includes the use of unsupervised clustering paradigms to explore the molecular heterogeneity of rosacea for treatment personalization. This demonstrates how ML models are able to identify new clinical features to improve clinical diagnosis of rosacea, identify new drug targets, and reposition existing drugs for the treatment of rosacea.

Acne vulgaris

Classification and diagnosis.

Acne lesion identification is sometimes included in multiclass diagnostic models. For example, Karthik et al (2022) utilized a modified attention-based CNN to differentiate acne from psoriasis, actinic keratosis, and melanoma.

Phenotyping, subtyping, and treatment personalization.

Current approaches are mainly aimed at assessing acne severity, with approaches that include an ensemble pruning framework (Liu et al, 2022a) and transfer learning with an image-based CNN model (Yang et al, 2021).

AI as a clinical aide for physicians and patients.

Lakdawala et al (2023) investigated ChatGPT’s responses to patient questions around acne, identifying lower accuracy responses concerning acne treatment. Other applications of generative AI in acne are overall limited.

Summary.

Compared with that in other inflammatory dermatoses, AI applications in acne vulgaris are primarily focused on severity assessment. Diagnostic and patient support applications are limited.

HS

Classification and diagnosis.

Within the last few years, ML has been utilized to assist in the diagnosis of HS. From claims data, Ali et al (2025) identified HS with XGBoost model and identified top features that include the use of sulfamethoxazole/trimethoprim, use of clindamycin phosphate, and having an obesity diagnosis. Kirby et al (2024) trained boosting algorithms to diagnose HS and distinguish it from mimics such as cutaneous abscesses and cellulitis, finding the strongest predictive features of age, sex, and risk factors (obesity, skin infections, skin infection treatments). For severity determination, Wiala et al (2024) automated smartphone image classification of HS using CNNs, with AUC up to 0.89 and noting that classification bias was caused by the presence of tattoos and excessive hirsutism.

Phenotyping, subtyping, and treatment personalization.

Cazzaniga et al (2021) performed cluster analysis on clinical data of 1100 patients with HS and identified distinct early-onset and late-onset clusters, with the latter group having significantly lower ORs of factors such as current smoking, neck involvement, and family history of the disease. González-Manso et al (2021) also utilized a clustering approach on select clinical, molecular, and genetic data to propose 2 distinct endotypes of HS of similar disease severity: cluster 1 represented nonobese males with nodular lesions posteriorly, higher serum IL-10, and presence of gamma-secretase allelic variants, whereas cluster 2 represented individuals with obesity with anterior lesions, more sinus tracts and abscesses, and higher serum IL-1, CRP, IL-17, and IL-6 concentrations. Nevertheless, there is still a significant opportunity to improve subphenotyping of HS heterogeneity to understand its pathogenesis as well as develop novel precision therapies for HS.

AI as a clinical aide for physicians and patients.

No related generative AI publications were found for HS. Given the chronic and devastating nature of the disease, there is a large opportunity for applications in chronic management and daily treatment optimization.

Summary.

Given the clinical heterogeneity and negative impact of diagnostic delays in HS, AI is emerging as a tool for earlier identification and patient stratification, with potential to improve disease understanding for treatment personalization.

AA

Classification and diagnosis.

Limited studies evaluating AA severity and progression risk exist. Lee et al (2020) developed AloNet, a deep learning framework to segment images of hair loss for AA and calculate Severity of Alopecia Tool (SALT) scores from standardized photographs. Zhang and Nie (2022) analyzed gene expression to predict progression to alopecia universalis, with XGBoost performing best for progression risk and identifying 4 key genes associated with disease progression (CD28, HOXC13, KRTAP1-3, GPRC5D).

Phenotyping, subtyping, and treatment personalization.

There are a few example applications of phenotyping AA and developing treatments. Xiong et al (2023) analyzed gene expression data from transcriptomic AA datasets to identify diagnostic markers for severe AA, identifying 4 immune-monitoring genes with diagnostic effectiveness (LGR5, SHISA2, HOXC13, and S100A3). Zhang and Nie (2022) evaluated key AA progression genes with the Connectivity Map database to identify azacitidine and anisomycin as possible repurposed therapies for AA. Chen et al (2021) analyzed transcriptomics from clinical trial biopsies using regulatory network analysis to define distinct drug mechanisms of response for treatments such as ruxolitinib and tofacitinib. By matching pretreatment biopsy profiles to these inferred networks, the authors proposed that these molecular profiles can guide patient—drug matching in clinical trials (Chen et al, 2021). Generative AI paradigms have also been applied to drug discovery. Specifically, the drug discovery company Absci has demonstrated the potential for de novo antibody design using generative AI. They developed an antibody (ABS-201) targeting the prolactin receptor for androgenic alopecia, which showed improved hair regrowth in mouse models compared with minoxidil, although these developments are still in early stages (Absci, 2025).

AI as a clinical aide for physicians and patients.

No high-quality studies were identified.

Summary.

AI applications are primarily focused on automating severity assessment (eg, SALT) or predicting progression to alopecia totalis or universalis and the identification of novel biomarkers of disease prognosis and progression. Deep learning and generative AI models also show promise in early therapeutic development for AA, but there still remains significant opportunity for research in this area.

Vitiligo

Classification and diagnosis.

Existing applications focus on image-based lesion detection and disease classification. Guo et al (2022) developed a hybrid AI model using YOLOv3 for lesion detection and UNet++ for segmentation, with better performance on Fitzpatrick skin types III/IV. Zhong et al (2024) compared Resnet and Swin Transformer deep learning models on dermoscopy images, with the latter model achieving better vitiligo classification performance (AUC = 0.94). Hillmer et al (2024) developed a CNN to quantify facial vitiligo severity with good physician agreement (interclass correlation coefficient = 0.67—0.70), with ability to track evolution over time.

Phenotyping, subtyping, and treatment personalization.

Limited studies exist in vitiligo. A promising study from Wang et al (2021a) investigated potential therapeutic targets using transcriptomics data and network analysis, identifying kaempferide as a candidate drug, with proteomic profiles suggesting potential multitarget mechanism involving p38 MAPK pathway inhibition (through CDK1/PBK) and cellular reduction—oxidation homeostasis modulation to promote melanogenesis.

AI as a clinical aide for physicians and patients.

No high-quality studies were identified.

Summary.

AI applications have primarily focused on image-based lesion segmentation and severity assessment. Early work exploring molecular data to identify potential therapeutic interventions is emerging.

Rheumatic skin diseases

CLE: classification and diagnosis.

Many AI applications in lupus primarily focus on SLE diagnosis, severity classification, or risk prediction of complications such as lupus nephritis (Ceccarelli et al, 2017; Li et al, 2022; Ma et al, 2022; Yaung et al, 2023). For example, Nair et al (2025) profiled medical records with generative AI to clinically classify SLE on the basis of American College of Rheumatology (ACR) 1997 criteria with high concurrence for the “discoid rash” criteria. A few studies specifically aim to diagnose CLE with CNN-based algorithms to classify CLE subtypes (acute CLE, subacute cutaneous lupus, discoid lupus) or differentiate from clinical mimics such as AD, psoriasis, and erythema annulare centrifugum (Li et al, 2024; Wu et al, 2021). One notable study utilized transcriptomics to classify CLE, identifying associated genes related to type 1 IFN—mediated inflammation (TLR3, MYD88, IRF7), STAT2 and PRF1, or IFN stimulation (such as CXCL9, CCL8, ISG15) (Seremet et al, 2024).

Phenotyping, subtyping, and treatment personalization.

Unsupervised and supervised learning paradigms have been utilized to deconstruct CLE heterogeneity or advance its understanding relative to SLE. Among studies on lesional skin, Lee et al (2025⁵) leveraged transcriptomics to phenotype the immune heterogeneity among various SLE rashes, identifying, for example, upregulation of type 1 IFN, TNF-alpha, IL6—Jak—STAT6 pathways for subacute CLE. From peripheral blood, Tao et al (2025) combined tandem mass tag proteomics with transcriptomics, identifying IFI44 and EPSTI1 as key serum markers that can differentiate subacute CLE from SLE (with AUC = 0.85—0.89), whereas Zhu et al (2021a) utilized whole blood RNA and unsupervised k-means clustering, identifying 6 patient clusters: 2 showing elevated inflammation modules characterized by IFN, neutrophil, and cell death signatures, whereas the other 3 clusters exhibited predominant T-cell signatures.

AI as a clinical aide for physicians and patients.

Nair et al (2025) profiled medical records with generative AI to classify SLE on the basis of ACR 1997 criteria. Other AI applications in CLE clinical care are limited.

Summary.

AI studies in CLE aim to classify disease, differentiate from SLE subtypes, and advance the understanding of disease heterogeneity. More work is needed in the area of AI applications for phenotyping, subtyping, and drug design for lupus because limited Food and Drug Administration (FDA)—approved therapies exist. Given large amounts of available multiomics and clinical data, this represents an opportunity for AI-based computational drug repurposing and biomarker discovery.

DM

Classification and diagnosis.

Limited approaches exist to diagnose DM. In one cross-sectional study, Xue et al (2023) trained a random forest model on clinical and laboratory variables to predict anti-MDA5 antibody status with AUC >0.97, finding association with high rates of fever, alopecia, periungual telangiectasia, Gottron’s sign, interstitial lung disease, and arthritis.

Phenotyping, subtyping, and treatment personalization.

Current AI applications leverage supervised or unsupervised learning paradigms to advance biomarker or subtype discovery. Wang et al (2024c) utilized gene expression data and weighted gene coexpression network analysis to pinpoint ISG15 as a DM biomarker, validating its biomarker potential in biopsied skin tissues. Neely et al (2022) utilized unsupervised learning methods on single-cell RNA and protein expression in blood in patients with juvenile DM, finding an inflammatory and antigen-presenting subphenotype in CD16+ monocytes and expanded transitional B-cell population as well as pan-cell—type IFN gene signature that correlated with disease activity in cytotoxic cell types. Additional ML studies into DM subtypes investigate phenotypic clustering on the basis of antibody status (eg, anti-MDA5) or predicting rapidly progressive interstitial lung disease (Koopman et al, 2024).

AI as a clinical aide for physicians and patients.

No high-quality studies were identified.

Summary.

AI applications in DM primarily focus on antibody prediction, biomarker identification, or subtype identification but, overall, had few studies focused on DM. There is a significant opportunity for AI approaches to improve DM disease classification, subphenotyping, and drug discovery.

Morphea and scleroderma

Classification and diagnosis.

Supervised learning in morphea/scleroderma have been applied to clinical data. Choi et al (2023) trained a random forest model, distinguishing 181 patients with localized morphea, generalized morphea, and systemic sclerosis (SSc), with top differentiating features (through SHAP) including anticentromere autoantibodies, anti—scl-70 autoantibodies, rheumatoid factor, histologic sclerosis, and noncutaneous manifestations. Jamian et al (2019) utilized EHRs to identify SSc, finding that the best algorithms combined clinical and billing codes, with top features including positive antinuclear antibody titers and presence of Raynaud’s phenomenon keyword.

Phenotyping, subtyping, and treatment personalization.

Phenotyping approaches with skin biopsies and gene expression data aim to advance disease understanding. Martínez et al (2022) analyzed lesional and nonlesional skin samples from various inflammatory skin diseases (CLE, psoriasis, AD, and SSc), revealing shared inflammatory profiles (eg, IFN, IL-12 complex) that distinguish from control skin as well as distinguishing features specific for each disease (eg, lack of IL-21 complex signature and increased TGFβ fibroblast signature in SSc). Franks et al (2019) developed an elastic net ML classifier using gene expression from SSc skin biopsies to assign to 4 intrinsic molecular subsets (inflammatory, fibroproliferative, normal like, and limited), with accuracy of 85%.

AI as a clinical aide for physicians and patients.

No high-quality studies were identified.

Summary.

AI applications are mainly aimed at research advancements, especially with molecular phenotyping.

Other inflammatory skin diseases

Classification and diagnosis.

Beyond the previously highlighted inflammatory dermatoses, classification approaches are being extended to other related conditions such as autoimmune bullous disorders and vasculitides. For autoimmune bullous disorders, Capurro et al (2024) compared models for automatic classification of direct immunofluorescence patterns, finding that the Swin Transformer generative model performed the best across the pattern classes (intercellular, linear, negative). Among vasculitides, an AI model achieved high diagnostic performance for predicting Kawasaki disease and intravenous Ig resistance from health data (Omar et al, 2025), whereas a hierarchical attention network model successfully predicted anti-neutrophil cytoplasmic antibody—associated vasculitis with AUC >0.9 from clinical records (Wang et al, 2025).

Phenotyping, subtyping, and treatment personalization.

A few studies that leverage AI to advance disease understanding exist. Hu et al (2024) phenotyped single-cell RNA and VDJ sequences on skin lesions from patients with bullous pemphigoid (BP), pemphigus vulgaris (PV), and controls, finding expanded CD8+ T-resident memory cells in PV and CD8+ T effector memory cells and CD4+ T regulatory cells in BP. Fortino et al (2020) phenotyped allergic contact dermatitis (ACD) and irritant contact dermatitis (ICD) with skin biopsy microarray gene expression modules, identifying biomarker genes such as ADAM8 and CD47 important in ACD and RRM2 important in ICD. Bettuzzi et al (2022) utilized factor analysis and hierarchical clustering on baseline clinical and biological data for 113 patients with epidermal necrolysis, finding 3 clusters: 1 cluster of sicker patients with high mortality and 2 clusters of younger patients.

AI as a clinical aide for physicians and patients.

Limited AI applications exist.

Summary.

Isolated studies aim to leverage AI to classify inflammatory dermatoses or advance disease understanding with immunophenotyping. Newer AI models, including deep learning and generative AI applications, are underutilized (Omar et al, 2025).

CHALLENGES, OPPORTUNITIES, AND FUTURE DIRECTIONS

AI is rapidly transforming dermatology by providing novel tools to address the complex diagnostic and therapeutic challenges associated with skin diseases. Although early AI tools were predominantly focused on skin cancer detection, more recent advances in AI have expanded considerably to focus on inflammatory and autoimmune skin diseases. In this review, we discussed applications of AI across the spectrum of basic, translational, and clinical care, including disease classification/diagnosis, phenotyping, subtyping, and patient education. We outlined in depth disease-specific advances driven by AI/ML in AD, psoriasis vulgaris, HS, vitiligo, AA, and rheumatic skin diseases. However, there remains significant gaps and unmet needs in the field.

The predominant application of AI for inflammatory skin disease lies in disease classification and diagnosis because the first models were primarily trained on imaging data. With the emergence of large-scale multiomics and clinical data, subsequent studies within the last ~5 years have now shifted to center on disease phenotyping, subtyping, precision medicine approaches, and drug discovery/repurposing. Fewest studies exist on generative AI applications in inflammatory skin disease, likely owing to the more recent development of transformer-based models. The primary focus of most of the AI-based studies are on AD and psoriasis, which is expected because they are 2 of the most common inflammatory skin diseases. Significantly fewer high-quality studies exist for acne, rosacea, AA, vitiligo, and HS. The fewest studies focused on rheumatologic skin disease (eg, lupus, DM, morphea/scleroderma, vasculitis) and autoimmune blistering disorders. Therefore, much more work is needed in the areas of AI applications for phenotyping, subtyping, and drug discovery for HS, vitiligo, and rheumatologic skin diseases because limited FDA-approved therapies exist. Given large amounts of available multiomics and clinical data, this represents an opportunity for AI-based computational drug repurposing and biomarker discovery.

The emergence of multimodal data integration, ranging from clinical images and histopathology to transcriptomic and genomic profiles, has enabled more refined classification, disease subphenotyping, and identification of novel biomarkers and therapeutic targets (Figure 1). Looking ahead, the convergence of multimodal data analysis offers unprecedented opportunities for precision medicine in dermatology. By moving beyond images and integrating data such as genetic profiles, transcriptomics, exposomics, clinical information, and patient outcomes, AI tools can be utilized to match patients with optimal therapies on the basis of individual disease pathways and subphenotypes (Patrick et al, 2018a; Wongvibulsin et al, 2022). This approach could significantly reduce the trial-and-error nature of current treatment selection for possibly poorly defined diagnoses of inflammatory skin conditions. There is also an unmet need and opportunity for approaches to distinguish inflammatory conditions from infectious and malignant processes, particularly in cases where clinical presentations may overlap (Ujiie et al, 2022). In addition, AI-powered analysis of population-level data could identify communities at elevated risk for developing inflammatory conditions, enabling earlier interventions and preventive strategies.

The choice of AI paradigm and approach is influenced by the specific application and data available. Supervised learning remains the approach for well-defined classification tasks, such as diagnosis, where the label is determined to be of good quality. Trained models can provide value to non-dermatologists or clinicians in primary care settings where there may be less familiarity with inflammatory dermatoses. The supervised learning paradigm may be difficult if ground truth is not certain, such as in the unclear clinical diagnosis for some rheumatologic conditions. In those cases, various AI paradigms are utilized, including unsupervised learning approaches, with the aim to understand molecular endotypes or help with discovery-oriented tasks for researchers. Recent advances in transformers and large-scale foundation models may enable more complex applications, such as generating a differential diagnosis or summarizing patient history for clinicians or providing answers to patient questions, although these complex tasks present challenges with greater use of computational resources, lack of interpretability, and lack of ground truth for quality evaluation, which will require extensive monitoring and optimization before full clinical adoption.

The integration of generative AI and ML in dermatology presents both significant opportunities and important challenges for basic scientists and clinicians. Several key challenges persist. AI applications in inflammatory skin diseases beyond psoriasis and AD remain relatively understudied. Algorithmic bias, particularly related to the underrepresentation of skin of color, continues to limit the generalizability of AI models. Although some efforts aim to leverage generative AI to simulate various skin tones (Munia and Al Zubaer Imran, 2025⁶), these approaches also risk exacerbating health inequities and magnifying institutional racism (Groh et al, 2024; Grzybowski et al, 2024; Mikołajczyk et al, 2022⁷). A few studies propose methods to improve fairness across groups, such as model pruning (Wu et al, 2022b⁸). Misinformation is also rampant, with significant bias, inconsistent validation, and inadequate privacy protection (Wongvibulsin et al, 2024). Key technical hurdles include standardizing image quality across photo acquisition methods, addressing confounding factors and artifacts in images, and developing representative datasets that span diverse patient populations (Choy et al, 2023; Daneshjou et al, 2022), although preliminary models are evolving that can provide feedback on image quality during acquisition (Vodrahalli et al, 2023). Related to this, there has been significant discussion in the community regarding the potential harms of AI-based models, especially with respect to how hallucinations can lead to treatment recommendations and patient instructions that are factually incorrect or potentially detrimental/dangerous to both patients and the health system. In addition, the clinical translation of many AI tools remains in its infancy owing to limitations in data quality, external validation, and integration into existing workflows. Successful clinical implementation requires careful attention to workflow integration, provider education, and patient acceptance, ensuring that AI tools can enhance care delivery worldwide.

The future role of AI in dermatology will eventually include the integration of generative AI tools that support both routine clinical care and complex decision making. Generative AI models will enhance medical education, clinical simulation, and patient communication (Motolese et al, 2022; Zhang et al, 2024b), whereas improved analysis of immune mechanisms will enable the development of targeted treatments (Martínez et al, 2022; Seremet et al, 2024). Advancing the management of inflammatory skin disorders will require sustained collaboration among clinicians, researchers, pharmaceutical experts, and technology developers (Sengupta, 2023), with a strong focus on ethical considerations and equitable access to AI-enhanced care.

Supplementary Material

Supplementary material is linked to the online version of the paper at www.jidonline.org, and at 10.1016/j.jid.2025.10.596.

Abbreviations:

AA

alopecia areata

ACD

allergic contact dermatitis

ACR

American College of Rheumatology

AD

atopic dermatitis

AI

artificial intelligence

AIME

Articulate Medical Intelligence Explorer

AUC

area under the curve

BP

bullous pemphigoid

CAM

class activation mapping

CLE

cutaneous lupus erythematosus

CNN

convolutional neural network

DM

dermatomyositis

EHR

electronic health record

FDA

Food and Drug Administration

HS

hidradenitis suppurativa

ICD

irritant contact dermatitis

LASSO

least absolute shrinkage and selection operator

LLM

large language model

ML

machine learning

PsAID

Psoriatic Arthritis Impact of Disease

PV

pemphigus vulgaris

SALT

Severity of Alopecia Tool

SLE

systemic lupus erythematosus

SSc

systemic sclerosis

STAT

signal transducer and activator of transcription

ViT

vision transformer