Up-to-Date Progress, Future Challenges and Future of Artificial Intelligence in Dermatology: From Diagnosis, Drug Discovery and Clinical Pathway

Abstract

The advent of artificial intelligence (AI) heralds a transformative era in healthcare, with dermatology emerging as a frontier for innovation. AI has demonstrated remarkable potential in enhancing diagnostic accuracy, therapeutic decision-making, and personalized management of skin diseases, particularly through image-based analysis. This review aims to summarize the evolution of AI algorithms in adapting to the progress of dermatological practices, assess the integration of AI-driven diagnostic and therapeutic interventions in clinical settings, and discuss the potential of AI in enhancing personalized and precise management of dermatological conditions. We evaluate the integration of AI-driven tools in clinical practice, highlighting their role in reducing diagnostic errors, while addressing persistent challenges such as demographic bias in training data. Further, we discuss the promise of generative AI in medical education, where synthetic clinical vignettes generated by models such as Generative Pre-trained Transformer 4 (GPT-4) have achieved high expert ratings for scientific accuracy. By critically analyzing current research and applications, this review outlines a roadmap for overcoming barriers including data diversity limitations and real-world validation and calls for interdisciplinary collaboration to advance intelligent, equitable dermatological care.

Graphical abstract

Keywords

dermatology / artificial intelligence / machine learning / precision medicine / clinical applications

Introduction

Artificial intelligence (AI), rooted in Turing’s seminal question “Can machines think?”, has evolved into a transformative discipline integrating theories, methods, and applications to simulate human intelligence[1,2]. Since AlphaGo’s 2016 victory over Lee Sedol, AI has advanced across text-speech recognition, image analysis, generative models, and large language models (LLMs), increasingly matching or exceeding human performance[3]. Improved computing power and algorithms have driven AI from laboratory concepts to multi-field deployment, spanning neural networks to protein structure prediction[4], contributing to the Nobel Prize in Physics and Chemistry in 2024. In medicine, a data-intensive field, AI’s precision and processing speed address the high demands of diagnostic imaging and clinical decision support, ensuring efficient and accurate data extraction and analysis[5,6].

Skin diseases rank as the fourth most prevalent global health issue, often underestimated due to inadequate health education and resources[6]. The 2019 Global Burden of Disease Study highlights a persistent rise in their prevalence, affecting millions worldwide[7]. In China, 30,000 dermatologists serve 360 million patients, revealing a substantial shortage domestically[8]. Meanwhile, sub-Saharan Africa has 3 dermatologists for every million individuals, compared with 100 for every million in the U.S. and as few as 2 for every million in certain South American regions. These shortages impede timely diagnosis and treatment, notably in underserved and remote areas[9,10]. To address these global inequities, AI-driven solutions, such as teledermatology platforms and automated diagnostic systems, present scalable strategies to improve accessibility, precision, and mitigate healthcare obstacles, thereby advancing equitable dermatologic care.

Dermatology traditionally relies on visual assessment, pathological interpretation, and dermoscopic image evaluation for diagnosis, making it conducive to medical imaging AI applications[11,12]. The training of AI models benefits from extensive datasets such as the 129,000-lesion archive curated by Esteva et al.[13]. Google Inception v3, a skin cancer detection system, exhibited diagnostic accuracy on par with expert dermatologists. AI contributes to objectivity by quantitatively analyzing lesion boundaries, pigmentation patterns, and textural features[14,15]. Moreover, AI-powered telemedicine platforms have enhanced diagnostic accessibility in underserved regions[16]. To ensure consistent clinical application, the Checklist for Evaluation of Image-Based Artificial Intelligence Reports in Dermatology (CLEAR Derm) Consensus Guidelines underscores essential principles such as fairness, robustness, and patient safety in AI[17,18]. This places greater demands on clinicians when using AI in clinical practice.

AI research has evolved from analyzing images to integrating electronic health records (EHRs), laboratory data, and multi-omics information to establish comprehensive diagnostic frameworks. Including machine learning (ML), these techniques have been instrumental in identifying metabolomic biomarkers for early detection of systemic lupus erythematosus (SLE), delineating DNA methylation networks in psoriasis, and constructing personalized models for atopic dermatitis based on mitochondrial gene profiles[19–22]. AI-driven tools, such as random forest models for predicting drug responses, have significantly enhanced personalized therapy optimization[21,23,24]. Autonomous surgical robots have demonstrated remarkable precision levels, facilitating precise resections of skin tumors[25–27]. Natural language processing (NLP) systems such as Generative Pretrained Transformer 4 (GPT-4) analyze EHRs and patient narratives to offer tailored clinical recommendations, thereby augmenting diagnostic and treatment procedures[27–33].

AI comprises five key technical domains driven by functional requirements: (1) pattern recognition for feature-based classification[34]; (2) ML to mimic human learning and improve algorithmic performance[35]; (3) data mining for extracting diagnostic and decision-support insights[36]; (4) intelligent algorithms designed for specialized computational tasks; and (5) human–computer interaction facilitated by LLMs to integrate clinical data[37].

This review synthesizes advances across four AI paradigms: image recognition, large language models, multi-omics, and multimodal integration, within a unified dermatological framework. It examines the translational pathway from algorithm development to clinical deployment, addressing regulatory, ethical, and implementation challenges. By bridging these domains, this review provides a practical roadmap for clinicians, researchers, and policymakers to advance intelligent, equitable, and precision dermatological care.

The AI development and the core algorithmic foundation

ML and DL of AI in dermatology

Dermatological AI relies on two key components: data engineering and data structuring, with ML and its deep learning (DL) subset serving as the foundation of diagnostic frameworks (Fig. 1). ML enables autonomous learning and adaptive optimization by iteratively improving algorithms using training data to enhance predictive accuracy and decision-making[38,39]. In dermatology, ML assists in pattern recognition, image classification, lesion segmentation, and personalized treatment, while DL excels in complex image analysis and automated feature extraction.

ML comprises four primary paradigms: supervised, unsupervised, semi-supervised, and reinforcement learning (RL)[40]. In supervised learning, lesions are classified using labeled images, such as distinguishing between benign and malignant skin lesions with pathologically annotated data[41]. Unsupervised learning involves handling unlabeled data through clustering and pattern recognition techniques, aiding tasks such as segmentation and anomaly detection[38]. Semi-supervised learning seeks to minimize annotation costs by utilizing a blend of limited labeled data and extensive unlabeled datasets, thereby improving diagnostic accuracy[42]. Reinforcement learning optimizes diagnostic decisions through a trial-and-error iterative process guided by feedback from clinicians. Its effective application can increase the sensitivity of melanoma detection from 61.4% to 79.5%[43]. Despite its effectiveness, RL encounters challenges due to its substantial computational demands and requires further optimization for personalized clinical applications[43–45].

DL is a subset of ML (Table 1) that utilizes artificial neural networks (ANNs), particularly convolutional neural networks (CNNs), to analyze medical images. ANNs enhance regression and classification tasks by capturing nonlinear relationships within data, while the multilayer convolutional architecture of CNNs enables automatic extraction of dermatological lesion features for classification, detection, and segmentation[42,46]. Each neural network layer captures distinct features (Fig. 2), and model performance is closely related to network depth, architecture, and data quality[47]. Training CNNs for dermatological diagnosis typically begins with acquiring high-resolution images of skin lesions, which are often sourced from datasets provided by the International Skin Imaging Consortium (ISIC), one of the most representative open-access dermatology databases, which has accumulated over 540,000 publicly available images[48]. Image preprocessing generally includes resizing images to a standardized resolution (e.g., 256 × 256 pixels), adjusting illumination, and applying data augmentation techniques such as flipping and rotation to improve generalization performance. A typical CNN architecture consists of six layers: an input layer, convolutional layers, activation functions, pooling layers, fully connected layers, and an output layer (Fig. 2). Preprocessed dermoscopic images are fed into the input layer and are typically represented as a three-dimensional array: height × width × number of channels. For instance, a dermoscopic color image with a resolution of 256 × 256 pixels is represented as a 256 × 256 × 3 array. The convolutional layer is the core component of CNNs and is responsible for automatically learning features from the input image, such as edges, textures, and shapes of skin lesions. A convolutional kernel, often with a small spatial size (e.g., 3 × 3 × 3), contains learnable parameters–weights corresponding to the input channels. The dot product between the kernel and the local receptive fields of the input image produces a two-dimensional matrix known as a feature map. A convolutional layer typically contains multiple kernels, each generating a distinct feature map, which are then stacked to form the output feature representation. Following convolution, a nonlinear activation function is applied to enable the network to learn complex nonlinear patterns. Pooling layers perform downsampling to reduce the spatial dimensions of feature maps, thereby decreasing computational cost and improving translation invariance, meaning that minor positional variations do not significantly affect recognition performance. Through successive convolutional layers, the network extracts hierarchical features, progressing from basic elements (e.g., colors and edges) to increasingly complex textures and patterns. These features are subsequently used to classify skin lesions and estimate the likelihood of skin disease (e.g., 85% melanoma vs. 15% nevus). Despite their high diagnostic accuracy, CNN models often face challenges in interpretability, commonly referred to as the “black-box” problem. To address this limitation, explainable artificial intelligence (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM), are used to generate heatmaps highlighting critical regions (e.g., lesion boundaries), thereby enhancing clinicians’ trust and clarifying the decision-making process.

Several prominent CNN architectures widely applied in dermatology—including GoogleNet, Inception-V3/V4, ResNet, Inception-ResNet V2, and DenseNet—have demonstrated outstanding performance in large-scale image classification tasks by overcoming challenges such as vanishing gradients and improving the recognition of complex features[39,42]. These models are typically trained on diverse and reliable datasets, such as the ISIC archive, DermNet, synthetic dermatological datasets, and collections that include samples from multiple skin tones[48–53].

Evaluation in AI diagnosis for dermatology

The assessment of AI-driven diagnostic and therapeutic systems frequently utilizes statistical parameters well-known to dermatologists, such as sensitivity, specificity, and accuracy[54,55]. Nevertheless, the advent of AI has brought forth supplementary statistical methods commonly employed in evaluating AI algorithms, such as the area under the receiver operating characteristic curve (AUROC), cross-validation, and the F1 score.

The AUROC assesses a model’s capacity to differentiate between normal and diseased states across various thresholds, with a value of 1 indicating flawless classification and 0.5 representing random chance[56]. The F1 score, which merges precision and recall, proves particularly beneficial in imbalanced datasets, offering a balanced metric ranging from 0 (complete error) to 1 (perfect classification) when precision and recall exhibit notable discrepancies[57,58]. Cross-validation, a pivotal technique for evaluating the generalizability of models, entails the iterative training and testing on distinct data subsets[59]. In dermatology, characterized by limited data availability, cross-validation enhances efficiency and diminishes bias compared to single-split evaluations[60]. For instance, in 5-fold cross-validation with 100 dermatological images, the dataset is divided into five subsets, each used once as the test set, providing a mean F1 score (e.g., 0.77) to indicate model stability across diverse samples.

The transferable application of AI technology from other diseases to dermatology

Beyond skin disease identification and assistance, AI possesses cross-disease expertise in the fields of retinal pathology, tumor detection, and image-assisted diagnosis. Gulshan et al.’s diabetic retinopathy system achieved an F1 score of 0.95 in diabetic retinopathy detection using DL (compared to 0.91 for ophthalmologists), earning U.S. Food and Drug Administration (FDA) approval[61]. Trained on hundreds of thousands of annotated images, the system learned complex pathological features, enhancing its generalization ability to enable automated initial screening of fundus images. This reduces physician workload while offering scalability for deployment across different regions and devices. AI models consistently achieve area under the curve (AUC) values exceeding 90% in applications such as lung cancer screening[62], breast cancer detection[63], prostate segmentation[64], pediatric disease diagnosis[65], and liver cancer detection[66]. Notably, in real-time multimodal fusion for liver cancer diagnosis and antimicrobial susceptibility testing, Sturm et al. validated ML-based nanomotion technology achieving an astonishing 98.9% accuracy in antibiotic sensitivity detection. This technology is transferable to pathogen identification and treatment monitoring in dermatology[67]. The key points for transferring these AI systems are: (1) massive standardized datasets, highlighting the importance of strengthening multi-center data sharing and standardization; (2) comprehensive integration of diagnosis, treatment, and management through advanced modeling techniques (e.g., reinforcement learning/multi-task learning); (3) seamless integration of AI with EHRs to advance intelligent diagnostic consultations and risk/prognosis assessment tools; (4) establishment of robust multi-center validation frameworks to ensure clinical relevance, interpretability, and data consistency in dermatological AI applications.

Research, development and application of AI algorithm in dermatological diagnosis

Feature focused image recognition model

Image recognition models are computer vision systems that utilize ML and DL algorithms to analyze image content, classify, and locate specific features within images. In dermatology, initial approaches relied on conventional ML techniques incorporating manually engineered features (e.g., color, texture, and shape). The emergence of DL, as demonstrated by AlexNet, facilitated the automatic identification of conditions such as melanoma. Subsequently, models pretrained on extensive datasets (e.g., ImageNet), such as ResNet and VGG, have been widely employed for the classification of skin diseases[68,69].

Current dermatological image recognition models (Fig. 2) focus on several key directions: (1) Mainstream architectures such as CNN, ResNet, DenseNet, and VGG enable automated extraction of texture and shape features through convolutional layers[70]. (2) Lightweight models such as MobileNet and EfficientNet prioritize compact design, computational efficiency, and low power consumption, supporting mobile and embedded device deployment[69,70]. (3) Medical-specific models, including DermaNet, U-shaped convolutional network (UNet), you only look once (YOLO), and generative adversarial network (GAN), are optimized for medical imaging tasks such as lesion segmentation and structural recognition, enhancing dermatological image analysis[71–73]. (4) Techniques such as transfer learning and few-shot learning fine-tune pre-trained models to mitigate data scarcity. Advanced architectures (e.g., Swin Transformer, two-branch networks) incorporate attention mechanisms to capture both local and global features, improving performance on complex and small-sample datasets[74,75]. (5) Interpretability-focused and multitask models integrate lesion localization, attribute recognition, and classification to align AI decision-making with dermatologists’ diagnostic logic, advancing clinical trust and adoption[76,77].

Image recognition models are extensively utilized in the early detection of skin cancer, particularly in identifying melanoma and basal cell carcinoma. These models also aid in the automated categorization of common skin conditions such as acne, eczema, and psoriasis, thereby enhancing the efficiency of outpatient workflows. Furthermore, they facilitate the quantification of disease progression for remote monitoring and follow-up purposes. Mobile applications (e.g., Aysa, Ruifu) have been integrated into outpatient services, enabling patients to conduct preliminary self-assessment via image uploads, thus streamlining pre-hospital diagnosis[78,79]. Ongoing challenges include the need to improve model performance for darker skin tones and rare diseases through the diversification of training datasets[80].The current mainstream dermatological public datasets, such as ISIC, HAM10000, and SCIN, still show significant imbalances in terms of skin color representation. The widely used ISIC database and its derivative dataset HAM10000 mainly consist of dermoscopy images from Europe and Australia, thus overall featuring a majority of individuals with light skin tones[48]. HAM10000 contains a total of 10,015 dermoscopy images, but this dataset does not provide clear Fitzpatrick skin type labels, making it difficult to directly conduct performance evaluations between different skin tone subgroups[48,81]. In the ISIC 2018 dataset, the individual typology angles of most images are mainly concentrated in the 34.5°–48° range, which mainly corresponds to lighter skin tones, while the proportion of dark skin samples is significantly insufficient[82]. In contrast, the Skin Condition Image Network (SCIN) dataset, designed at the outset with a clear emphasis on skin color diversity, includes over 5,000 volunteers who contributed more than 10,000 images[83]. It also provides self-reported and expert estimated skin color labels in the Fitzpatrick III–VI types, offering a wider coverage. Moreover, SCIN not only includes neoplastic lesions but also incorporates inflammatory, infectious, and allergic skin diseases, which are typically underrepresented in traditional dermoscopy datasets focused mainly on skin cancer[84].

To overcome the inherent deficiency of current mainstream dermatology datasets in terms of skin color representation and prevent AI from exacerbating health inequalities in clinical deployment, future research should focus on data source expansion and targeted collection for minority ethnic groups, and conduct multi-center collaborations across countries or regions. For instance, a prospective image acquisition cohort specifically targeting Fitzpatrick IV–VI types could be established, or active collection of images of common skin diseases (such as inflammation and infection) among non-white ethnic groups could be carried out in grassroots community clinics, thereby filling the representational gap in the absolute quantity of data distribution[85,86]. However, merely expanding the data may not solve the problem; fairness considerations should be introduced in the algorithm design through loss functions to correct model bias. In conventional training, the standard cross-entropy loss is prone to driving the model into the trap of “majority class optimality”, masking the misdiagnosis rate of dark skin. To address this issue, dynamic intervention mechanisms can be introduced during the training phase. For example, sample reweighting or focal loss can be used to assign higher error penalty weights to the less frequent dark skin samples in the training set; or adversarial debiasing networks can be utilized to force the extraction of lesion features independent of skin color features, thereby mathematically compelling the model to achieve performance balance across different skin color subgroups[87]. Before clinical deployment, the AI model must undergo external multi-ethnic validation tests. The AI model cannot merely be internally evaluated on a homogeneously divided test set but must undergo rigorous double-blind tests in an independent external cohort that is completely separate from the training environment and has a different demographic distribution (for example, an independent validation set collected in front-line medical institutions in Asia or Africa). Only when its generalization ability and diagnostic fairness are proven in diverse populations can the AI system be regarded as a safe and reliable universal medical tool[88].

LLMs based on inquiry-analysis-feedback

ML and DL models have facilitated the automated classification and diagnosis of various skin conditions such as skin cancer, pigmented lesions, eczema, acne, and psoriasis[13,48,89]. Nevertheless, current models are constrained to specific classification tasks and necessitate validation by dermatologists. They do not possess the ability to generate natural language reports or engage with patients autonomously, thereby limiting symptom description and feedback. The incorporation of LLMs into dermatological workflows presents a novel opportunity to improve doctor-patient communication, provide diagnostic support, and enhance health education.

LLMs, a type of DL-based NLP technology, are trained on extensive datasets (e.g., books, articles, online content) to perform various tasks including text generation, translation, question answering, and code composition[90]. Built on the Transformer architecture, LLMs utilize multi-layer encoder-decoder structures to capture contextual dependencies and produce coherent natural language output. These models tokenize input text into words or sub-words, convert them into vector representations, and leverage attention mechanisms to predict and construct coherent sentences or paragraphs. In the field of dermatology, LLMs analyze patient histories and symptoms to propose diagnoses and referrals, demonstrating proficiency in natural language reasoning and context-aware recommendations as opposed to traditional algorithms focused on image analysis (Fig. 3). SkinGPT-4 combines a visual Transformer with an LLM to generate diagnostic reports with accuracy comparable to that of dermatologists[33]. LLMs are also employed to automate pathology reports, aid in complex case reviews, and provide guidance in image analysis and immunohistochemistry[91]. In dermatology certification examinations involving image-text cases, GPT-4 achieved a 90% accuracy, meeting clinical standards[92]. LLMs designed for patient interaction translate symptom descriptions, such as “red and itchy skin”, into easily understandable health information, offering personalized advice.

LLMs are instrumental in advancing equitable healthcare. GPT-4 Vision, mitigates skin color bias by accurately identifying melanoma across diverse skin tones, enhancing diagnostic fairness[93]. The SkinGEN framework combines an LLM with an image-generation model and GAN to explain diagnoses, boosting patient trust in AI[94].

Multi-omics models based on molecular typing

The integration of multi-omics technologies with AI has advanced the field of dermatology by transitioning from single-modality to comprehensive data analysis. Through the amalgamation of genomic, epigenomic, transcriptomic, and proteomic data, these models facilitate precise subtyping. A multi-omics model based on gene co-expression networks (GCN) that combines RNA expression and DNA methylation has demonstrated a substantial enhancement in the accuracy of tumor classification[95]. Comparable methodologies are being utilized in the study of skin conditions such as psoriasis and melanoma to achieve more nuanced subtype distinctions.

Recent DL advances, including sparse matrix factorization and self-supervised learning, have improved dimensionality reduction and feature extraction for high-dimensional multi-omics data, retaining key molecular signatures while reducing computational demands. These methods facilitate biomarker discovery, precise diagnosis, and personalized treatment of skin disease subtypes. The PathME model, leveraging pathway-based features, enhances molecular subtyping accuracy for inflammatory skin diseases such as psoriasis[96]. Moreover, to enhance clinician confidence, multi-omics approaches are increasingly incorporating XAI by integrating protein–protein interaction and ceRNA networks with causal inference models[97]. This integration enables the generation of diagnostically relevant outputs that are both biologically meaningful and interpretable.

Multi-omics models offer significant advantages in the clinical diagnosis and basic research of skin diseases: (1) integrating diverse omics layers to uncover disease pathways and therapeutic targets[98]; (2) leveraging single-cell and spatial transcriptomics for precise inflammatory cell profiling in conditions such as psoriasis and atopic dermatitis; (3) analyzing skin microbiota dynamics via combined metagenomic and proteomic data, offering new perspectives on host-microbe interactions[99]; (4) predicting personalized drug responses by evaluating skin barrier integrity, metabolic pathways, and immune status, advancing precision dermatology and cosmetology[100]. These models also refine immunotherapy strategies by dissecting gene expression patterns across tumor and stromal compartments in cancers such as melanoma, enabling targeted modulation of immune cells[98]. The interdisciplinary nature of multi-omics fosters collaboration across dermatology, microbiology, and immunology, accelerating comprehensive exploration of skin health and disease[101].

AI-driven multi-omics models have made significant strides in accurately classifying skin diseases, yet encounter challenges including data heterogeneity, noise interference, limited biological interpretability, and high computational demands[97,102]. To enhance interpretability, integrating causal inference with biological networks such as protein–protein interactions and pathways is crucial. Additionally, the analysis of high-dimensional omics data necessitates substantial computing resources and cost of sequencing, thereby restricting accessibility in resource-limited settings. Data scarcity, particularly concerning rare subtypes of skin diseases, further compromises the robustness of models. Federated learning presents a promising solution by facilitating the sharing of data across multiple institutions to enhance the generalizability of models. Future endeavors should prioritize the development of interpretable, scalable, and globally inclusive AI systems that integrate multi-omics data for dermatological applications.

Multimodal model for multidimensional information integration

Integrating cross-modal data, such as medical images, clinical records, omics, and dynamic monitoring, into multimodal models enhances diagnostic precision and treatment decisions, beyond the molecular emphasis of multi-omics and the linguistic capabilities of LLMs (Table 2)[103]. Skin conditions such as SLE, psoriasis, autoimmune bullous disorders (e.g., bullous pemphigoid, linear IgA disease), and melanoma are characterized by intricate interactions among immune responses, genetic predispositions, and environmental influences. Therefore, a comprehensive integration of data is essential to improve clinical outcomes in managing these conditions[104–107].

In dermatology, multimodal data fusion spans: (1) images (dermoscopic, clinical, pathological); (2) text (medical histories, diagnostics, lab reports); (3) multi-omics data; (4) dynamic monitoring (wearables); and (5) other formats (videos, audio). Traditionally analyzed in isolation, these diverse data sources are now being integrated using AI techniques to enhance both diagnostic processes and treatment outcomes. Hao et al. proposed a multimodal framework combining RNA and chromatin data to refine cell identity mapping and elucidate immune responses to vaccination[108]. Similarly, multimodal image fusion (CT, MRI, and PET) has improved diagnostic accuracy[109], while Medical Contrastive Language-Image Pre-training (MedCLIP) and Medical Generative Pre-trained Transformer (Med-GPT) integrate images, text, time-series, and omics data, advancing medical knowledge retrieval and clinical decision-making[103,110,111].

The clinical application of multimodal AI models typically follows five steps: (1) data acquisition and preprocessing (cleaning, standardization, noise reduction); (2) feature extraction to refine modality-specific data; (3) modality alignment and algorithmic fusion; (4) model training and optimization; and (5) deployment prioritizing practicality and interpretability[112]. Each modality requires tailored preprocessing: for example, U-Net segments dermatological lesions[113]; omics data undergo noise reduction and key feature selection to identify biomarkers[114–118]; support vector machine (SVM) detects hyphal features in fungal infections[119]; and NLP extracts disease-relevant information from EHRs[120].

Data fusion methods are categorized into early, mid, and late fusion. Early fusion combines features at the input level, which is beneficial for highly correlated modalities. For instance, merging dermoscopic images with medical history can enhance the assessment of severity[117]. Mid fusion involves extracting features independently and integrating them through joint learning, making it suitable for intricate analyses such as combining RNA sequencing with imaging data using co-attention mechanisms[121]. Late fusion merges final predictions through weighted voting or averaging, particularly effective when modalities are weakly correlated, as seen in integrating DNA methylation data with clinical scoring models in SLE[114]. Training models that incorporate multiple modalities poses challenges related to data heterogeneity, noise, and redundancy[118]. Dermatological images, for example, may be affected by lighting conditions, skin oils, or hair, while text data could suffer from semantic mismatches. A fundamental challenge lies in establishing connections between visual features (e.g., lesion texture) and symptoms (e.g., itching severity)[118]. Recent strategies to address these challenges include optimizing loss functions, such as cross-entropy, to enhance classification robustness[122], integrating ResNet with attention mechanisms for precise lesion detection[119], and utilizing Transformers to dynamically assign weights to modalities for improved alignment[123].

The potential of multimodal AI in dermatological diagnosis is significant, yet it encounters challenges such as data inconsistency, high computational costs, and limited interpretability[122,123]. Variability among institutions impedes model generalization, while the complexity of high-dimensional data increases training requirements.

AI promotes systematic innovation in skin disease drug development

AI accelerates drug development by enhancing target identification, lead optimization, molecule design, and trial simulation. This transformation, traditionally a decade or more, is now expedited through AI technologies such as omics analysis, DL, and NLP. Insilico Medicine’s PandaOmics identified TRAF2- and NCK-interacting kinase (TNIK) as a fibrosis target and completed preclinical candidate screening within 18 months using Pharma.AI. GENTRL algorithm designed six discoidin domain receptor 1 (DDR1) inhibitors in 21 days, resulting in four active compounds and one with promising in vivo outcomes[124–126]. In China, DeepBlock improved ligand design, increasing synthesizability by 6% and achieving a 0.54 drug-like score[127]. Stanford’s TrialPathfinder simulated lung cancer trials, suggesting that relaxed eligibility criteria could double enrollment and reduce mortality risk, benefiting more patients[128]. In dermatology, AI advances precision therapy by integrating omics, imaging, and clinical expertise for novel drug discovery, regimen optimization, and drug repurposing (Table 3).

AI-assisted new drug discovery: cross-disease, high-throughput, targeted precision

AI accelerates target discovery and candidate compound generation by advancing drug development from “single-target” to “multi-target network” paradigms. Through the integration of transcriptomics, protein interactions, and structural data, AI constructs intricate drug–target–pathway–disease maps. Graph neural networks (GNNs) uncover critical regulatory modules, such as the IL-23/Th17 and JAK-STAT pathways’ interaction in psoriasis[129]. Yu et al. introduced DrugMAN, a multi-graph attention model that merges protein and chemical networks to pinpoint targets for psoriasis, surpassing graph attention network (GAT) and GCN benchmarks[130]. Fang et al. devised the DGMM model, which merges D-VAE encoding with a genetic algorithm to generate drug-like, synthesizable small molecules[131]. By simulating the evolution of drugs, the model achieves human-level performance in molecular optimization. DGMM establishes a comprehensive AI framework from mechanism discovery to compound generation, particularly suitable for intricate, multi-mechanism skin diseases.

Personalized treatment: prediction of efficacy, and population typing

In the realm of limited clinical drug trial data, the utilization of AI to tailor treatment approaches through the prediction of drug responses and the provision of interpretable guidance is a prominent area of research. Maulana et al. introduced a framework that integrates extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and Random Forest, incorporating various features such as skin inflammation scores, serum markers, and family history to categorize the severity of atopic dermatitis and optimize dosage[132]. The DRExplainer model merges dynamic graph convolutional networks with interpretable subgraph mining to forecast psoriasis outcomes and identify JAK/ILs-related gene subnetworks[133]. Additionally, ASGARD employs single-cell expression data to delineate drug response landscapes, initially focusing on tumors and subsequently adapting for inflammatory skin conditions such as vitiligo[134]. STEDR, on the other hand, simulates treatment trajectories based on EHRs, offering personalized classification and risk assessment strategies[135].

Drug repurposing

AI facilitates identifying drugs capable of reversing disease-specific gene expression patterns, by comparing differentially expressed genes (DEGs) in lesional tissues with drug-induced transcriptional profiles, leveraging the inverse expression principle. Databases such as the Library of Integrated Network-Based Cellular Signatures (LINCS) L1000 enable large-scale matching between disease perturbation signatures and pharmacologic effects, supporting candidate prioritization through network-based and pathway co-modulation analyses[136,137].

Wen et al. integrated clinical feature vectors with CMap data to reverse-match drugs for off-label indications, improving screening accuracy[138]. Liu et al. applied counterfactual simulations to reconstruct hypothetical patient responses, enabling personalized efficacy comparisons[139]. DrugAgent combined multi-agent models (e.g., DeepPurpose drug-target interaction [DTI]) with knowledge graphs (DGIdb, DrugBank, CTD, STITCH) to map drug–pathway–literature networks and validate candidates for atopic dermatitis[140]. Greugny et al. employed an ordinary differential equation (ODE)-based kinetic model to simulate skin flora recovery, identifying a therapeutic window for microecological interventions in acne and eczema[141].

Drug development to clinical implementation: model compression, system integration, and treatment closed-loop control

AI integration is accelerating innovation in dermatological drug development, with early efforts centered on image recognition. Current advancements emphasize model compression, system integration, and closed-loop treatment control. Velasco et al. showed that MobileNet achieved 94.1% accuracy for psoriasis, acne, and ringworm, making it suitable for mobile deployment[142]. A distilled melanoma model was reduced to 0.26 million (M) parameters, enabling wearable applications for efficacy monitoring[143]. Han et al. further combined AI risk scoring with microneedle delivery to create a closed-loop drug release system for psoriasis treatment[144]. These developments highlight AI’s transition from a diagnostic tool to an active participant in therapeutic modulation.

Development of dermatological AI clinical deployment and related standards and consensus

Effective implementation of AI in dermatology requires seamless integration of ML, DL, LLMs, and multimodal models with clinical tools such as dermatoscopes and EHRs to ensure data interoperability. AI should augment, not replace, clinician decision-making. Compliance with regulations such as General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA) is essential to protect patient privacy and ensure secure clinical deployment. In clinical practice, AI systems must adapt to real-time feedback and maintain strong interpretability and user-centered design to gain clinician and patient trust (Fig. 4). Promoting data sharing via standardized formats such as Fast Healthcare Interoperability Resources (FHIR) and Health Level Seven (HL7) is essential for cross-institutional collaboration and the large-scale implementation of dermatology AI.

Trends in AI model deployment for dermatology: from traditional algorithms to hybrid and specialized models

AI applications in dermatology have evolved from classical ML to multimodal models. Early algorithms such as traditional ML models such as least absolute shrinkage and selection operator (LASSO), SVM, and naive Bayes (NB) have supported gene marker identification[146], personalized treatment planning[147], and infectious disease screening[148,149]. Yet, their dependence on manual feature extraction limits performance in high-dimensional tasks. DL overcomes this via end-to-end processing, with CNN models achieving 96%–97% accuracy on HAM10000 and surpassing conventional methods in tumor classification (e.g., VGG16, ResNet50: 88% vs. 72%)[150]. Transformer-based LLMs such as GPT-4 and BERT also enhance text-driven diagnostic support[151]. As diagnostic needs grow, specialized and hybrid models are gaining traction. Tailored DL networks for skin cancer improve efficiency and suit mobile deployment through parameter simplification[152]. Models adapted for diverse skin types enhance diagnostic fairness[153]. Hybrid approaches that combine CNN feature extraction with SVM classification improve accuracy, especially in small datasets[154,155]. These models are advancing personalized care and decision support in dermatology.

Research standards and application specifications

A robust standardization framework is essential for clinical deployment, including: (1) foundational standards: defining AI concepts, model structures, and evaluation metrics; (2) data standards: such as HL7 and FHIR for cross-platform data exchange; (3) technical standards: detailing model training and validation protocols; (4) application standards: outlining clinical decision-making criteria; and (5) safety and ethical standards: ensuring patient data protection under GDPR and HIPAA regulations[156–158].

Effective deployment of AI in dermatology relies on the complementary roles of FHIR and the Common Data Model (e.g., Observational Medical Outcomes Partnership [OMOP] ). FHIR functions as a communication protocol facilitating standardized exchange of clinical, imaging, and omics data through modular APIs and structured formats such as JSON/XML[159–161]. It enables genomic-clinical integration in conditions such as psoriasis and SLE[162], enhancing data interoperability across hospital networks[163]. OMOP offers a unified schema for multi-center AI research, aiding in integrating EHRs, imaging, and omics data for disease modeling and personalized treatment[164–166]. However, OMOP encounters challenges in handling unstructured inputs such as dermoscopic images and physician notes. Improving OMOP’s flexibility and enhancing FHIR’s data granularity are crucial for developing robust AI models in dermatology.

To advance AI-driven dermatology, existing frameworks should be extended to accommodate dermoscopic imaging, genomics, and clinical history. Integration tools such as OHDSI’s ATLAS can support data mapping and quality control. Establishing dermatology-specific databases and fostering multi-center, cross-national data exchange will strengthen diagnostic precision and individualized care through standardized, interoperable data infrastructures.

Real-world translational applications of AI in dermatology

Telemedicine and mobile health

Telemedicine and mobile health technologies have revolutionized the provision of care for skin diseases, effectively tackling the worldwide impact of dermatological conditions. In Europe, an estimated annual affliction of over 185 million adults underscores the significant burden[167]. Scarce dermatological resources pose challenges to prompt care delivery, especially in marginalized regions. The incorporation of AI in diagnostic processes, monitoring strategies, and teleconsultation services enhances accessibility and operational effectiveness. Research indicates that the implementation of smartphone-enabled remote management leads to a substantial 70% decrease in physical consultations, thereby optimizing results and resource distribution[168,169].

AI-powered telemedicine enables patients to capture high-resolution skin images via mobile devices, allowing physicians to remotely assess lesion characteristics, such as morphology, color, and texture, for faster, more accurate diagnosis[170,171]. This is especially beneficial for managing chronic conditions such as eczema and psoriasis. Wearables and health apps transmit data on skin moisture, inflammation, and erythema, supporting real-time monitoring and treatment adjustment[172,173]. Integrated platforms provide clinicians with medical history, symptom evolution, and family history, enhancing diagnostic precision for complex autoimmune skin diseases[174,175]. EHR systems further assist in tracking treatment responses and refining personalized care. Additionally, telehealth tools empower patients with educational content and self-monitoring functions. Surveys show over 90% acceptance of online consultations and e-prescriptions[176].

The success of AI-driven telemedicine relies on the integration of diverse data modalities. Although the use of omics data in telehealth is nascent, home genetic testing enables early identification of hereditary skin disorders including keratosis and congenital pigmentation abnormalities[162,177]. Wearable devices capture dynamic parameters such as skin temperature and humidity, which can indicate inflammatory activity in real time[172,173]. The fusion of these inputs requires advanced AI processing capabilities and underscores the importance of training clinicians in intelligent diagnostic workflows.

Diagnostic accuracy of telemedicine applications depends heavily on reliable, timely data transmission, and any delays or data loss can impair outcomes. Inconsistent data formats, such as variations in EHR systems across institutions, further hinder integration efforts[178]. Ensuring patient privacy remains critical, requiring strict adherence to GDPR, HIPAA, and best practices for data encryption and anonymization. Current AI tools are limited by data quality and user proficiency among clinicians and patients. Strengthening multimodal modeling is essential to improve adaptability and performance[179]. With continued technological progress, telemedicine is expected to play a growing role in delivering dermatological care and expanding global access.

Typical tool applications of dermatological AI

We summarized major AI tools currently applied in this field, covering diagnosis, image classification, subtype identification, inflammation quantification, clinical integration, and drug development. Table 4 presents an overview of each system’s functions, target conditions, implementation status, and supporting studies, offering a reference framework for advancing intelligent dermatological diagnostics and therapies.

Practical AI considerations for dermatologists: data, algorithms, and computational resources

As shown in Figure 5, the clinical utilization of AI in dermatology is underpinned by three fundamental components: data, algorithms, and computational infrastructure. Among these, the quality of the training data, which encompasses skin lesion images, pathology slides, EHRs and clinical documentation, plays a pivotal role in determining model efficacy[206]. Nonetheless, raw data frequently exhibit imperfections such as noise, redundancy, and gaps in annotation, necessitating systematic preprocessing procedures such as error rectification, standardization, and outlier elimination to enhance training resilience[207]. In supervised learning, precise expert annotation of diagnostic labels and lesion delineations is indispensable for enhancing model training and predictive accuracy.

Early dermatological AI heavily relied on traditional models for their interpretability, but they faced challenges with high-dimensional, heterogeneous data and lacked the scalability of DL models[208–210]. DL addresses these limitations by enabling end-to-end learning to automatically extract intricate features from raw data. CNNs excel in skin image classification and lesion detection, while Long Short-Term Memory networks (LSTMs) handle time-series data for dynamic disease monitoring. GANs enhance and synthesize medical images[39,211–213]. Despite their effectiveness, DL models require significant computational resources, such as GPUs and TPUs, and are often opaque, complicating result interpretation and potentially undermining clinical trust. In the realm of medical AI research, there is a growing interest in specialized and hybrid models for their task-specific adaptability. Lightweight CNNs, designed for dermatological image analysis on mobile devices, optimize computational efficiency in resource-constrained environments[214]. Hybrid models frequently merge CNN-based feature extraction with classifiers such as SVM or random forests to enhance diagnostic versatility[215]. An innovative hybrid quantum computing model proposed by Vijayasri Iyer et al., leverages quantum systems to encode dermatological image features and achieves high classification accuracy on the HAM10000 dataset, showcasing the promising integration of quantum computation in dermatological diagnostics[216]. However, these models may impose substantial computational requirements, particularly in real-time and multimodal scenarios, necessitating distributed computing or multi-GPU systems for practical implementation.

For clinicians, the clinical deployment of AI modules in dermatology requires comprehensive planning across multiple dimensions: (1) data preparation: Identify relevant data types (e.g., clinical images, omics data, medical records, pathology slides, teleconsultation transcripts, and follow-up records); determine whether preprocessing (e.g., noise reduction, standardization) is necessary; consider patient stratification based on clinical features; and evaluate the need for optimization strategies to improve data quality. (2) Algorithm design: decide whether to adopt traditional interpretable models or generalizable models suited for large-scale data; assess the influence of clinician expertise on supervised/semi-supervised models; and clarify whether the algorithm prioritizes biological plausibility or predictive performance. (3) Computing and integration: ensure adequate computational resources (e.g., GPU clusters) to support model training and deployment; confirm compatibility across AI modules to achieve system-level coherence. (4) Clinical applicability: evaluate the real-world feasibility of the AI model, including ease of integration into clinical workflows, user-friendliness for dermatologists, and equitable access for patients in diverse healthcare settings.

Challenges in real-world deployment: regulation, governance, and economics

The transformation of AI models for dermatology from experimental models to clinical tools inevitably requires strict regulatory review. Currently, AI systems are mainly regulated as “Software as a Medical Device (SaMD)”. The U.S. FDA has recognized the iterative nature of ML and introduced the Pre-Certification for SaMD program, allowing algorithms to self-update within a specific range after deployment without the need for reapplication for pre-market approval[217]. In Europe, the regulatory framework is dominated by the Medical Device Regulation (MDR) and the newly introduced EU AI Act. Under these frameworks, AI systems for dermatological diagnosis are typically classified as Class IIa/IIb medical devices or “high-risk” AI systems[218]. These classifications mandate: (1) Data governance: the training, validation, and testing datasets of AI systems must be of high quality. Developers must document the source (provenance) of the data, the original purpose of data collection, and conduct rigorous bias assessment. If the dataset is severely imbalanced in terms of skin color (Fitzpatrick classification), gender, or region, which may harm patient health or lead to discrimination, the system will be denied market entry. (2) Algorithm transparency: AI systems must be designed with sufficient explainability to enable medical professionals to understand the system’s output and take appropriate actions. Additionally, manufacturers must provide detailed technical documentation and user instructions, clearly informing doctors of the system’s capability boundaries, expected diagnostic accuracy, and in which specific populations it may fail. (3) Cybersecurity and robustness: AI must have a high level of defense against cyber attacks, hacking, and data tampering. It must also seamlessly integrate with data regulations such as the EU’s General Data Protection Regulation (GDPR). (4) Pre-market assessment: high-risk medical AI cannot be marketed through the manufacturer’s “self-declaration”. They must undergo strict audits by independent third-party notified bodies (such as BSI, GMED). The review includes quality management systems (ISO 13485) and medical device risk management (ISO 14971), and incorporates the special requirements of the AI Act. Only after meeting all standards can they obtain the CE marking and enter the European market. In China, AI medical independent software with auxiliary diagnostic functions (such as directly indicating the risk of melanoma) is uniformly classified as the highest-risk Class III medical device and requires extremely strict clinical evaluation and technical review by the National Medical Products Administration (NMPA) authority to be approved.

Post-market supervision and model performance drift monitoring for dermatology AI One key challenge in clinical deployment is the phenomenon of performance degradation, often referred to as “dataset shift”[219]. An AI model primarily trained on specific demographic characteristics (e.g., light-skinned populations) or specific imaging hardware may experience a sharp decline in diagnostic accuracy when exposed to a diverse real-world population or new dermoscopy devices. To mitigate this issue, a robust post-market supervision mechanism must be enforced to detect performance drift in real time. This includes regular audits using diverse, cross-institutional clinical data, and the implementation of an automatic warning system when the model’s confidence level drops below a preset threshold[220]. If the performance of the maintained AI model falls below the threshold, dermatologists should re-examine the misdiagnosed skin lesions and feed them back to the model for re-training to maintain its long-term clinical utility. At the same time, it is necessary to consider: when an AI system misses or misdiagnoses a malignant melanoma, how to define the fault: should it be attributed to the attending dermatologist, the medical institution, or the software developer[221]. Additionally, clinicians may become victims of “automation bias”, that is, over-reliance on algorithm outputs and abandonment of their own clinical judgment. To address these governance challenges, dermatology AI must be strictly positioned as an auxiliary tool rather than an autonomous decision-maker, ensuring that “AI is practiced by humans in human medicine”[222]. Professional dermatology societies must establish clear clinical governance frameworks and practice guidelines, clearly stipulating that although AI provides probability-based risk assessments, the ultimate responsibility for diagnosis and treatment lies with the qualified doctor.

As an atypical medical device, the wide clinical application of dermatology AI fundamentally depends on the establishment of a sustainable reimbursement model. Currently, healthcare payers lack standardized coding systems (such as specific Current Procedural Terminology [CPT] codes) to provide reasonable compensation for AI-assisted image analysis or remote dermatology triage. To clear the reimbursement channel, developers and researchers must generate strong health technology assessment and health economics evidence[223]. This requires conducting prospective trials that not only prove the non-inferiority of the model in diagnosis but also demonstrate its cost-effectiveness. For example, an AI triage system must prove its financial rationality by providing evidence such as significantly reducing unnecessary excision biopsies for benign skin lesions (such as common moles), optimizing the workflow efficiency of dermatology clinics, or moving the detection of skin cancer forward, thereby avoiding extremely expensive late-stage tumor treatment costs[224]. Without convincing cost-saving and value-based medical evidence, the large-scale promotion of AI in routine dermatology practice will face significant financial obstacles.

Conclusion and perspectives

In dermatology, AI has approached clinical utility in disease classification, severity scoring, therapeutic response prediction, and remote monitoring, marking its transition from an “experimental tool” to a “clinical assistant”. These advances are driving dermatologic care toward greater intelligence, personalization, and precision.

The diverse and complex nature of dermatologic data has surpassed the capabilities of single-modality models. This challenge has led to the development of multimodal and multi-omics approaches that integrate visual, textual, and molecular data inputs. Current AI research is increasingly focusing on integrated frameworks. For instance, early multimodal systems such as SkinGPT-4 and DeepDR-LLM combine image recognition with natural language processing to aid in diagnostics[33,225]. The DeepGEM model integrates imaging with genomic data to improve spatial mutation prediction[226], while Evo, a generative model for DNA sequences, expands AI applications to target identification and synthetic biology[227].

The clinical integration of AI in dermatology encounters challenges including data heterogeneity, limited interpretability, clinician skepticism, and inadequate regulatory and ethical frameworks[228]. These obstacles are especially pronounced in intricate immune-mediated skin conditions, where modeling and assessment methods lack standardization. Future developments in dermatologic AI will emphasize: (1) Model architecture: image recognition model, multi-omics models, LLMs, and multimodal frameworks will enable a shift from simple data fusion to mechanism-driven modeling. Omics-based algorithms will identify key molecular targets and integrate clinical context for precision stratification and treatment planning. LLMs will power intelligent pre-consultation systems, enhancing diagnostic workflows. (2) Clinical deployment: next-generation platforms will fuse imaging, text, and lab data; knowledge graphs will inform intelligent decision-making; dynamic physiological monitoring will enable “digital twin” simulations; and protein structure prediction will support personalized drug design. (3) Training and infrastructure: improving model interpretability and training efficiency will be critical, alongside implementing standardized frameworks such as FHIR and OMOP to enable data integration, governance, and equitable deployment. By aligning technical innovation with clinical needs and embedding AI into ethical and regulatory structures, dermatology can advance toward intelligent, inclusive, and patient-centered care.

With ongoing progress in algorithmic design, clinical adaptation, and treatment feedback, AI is anticipated to transition from a decision-support tool to an intelligent collaborator throughout the dermatologic care process. This evolution is poised to facilitate the advancement of more equitable, efficient, and reliable precision medicine in the foreseeable future.

References

[1] Miller DD, Brown EW. Artificial intelligence in medical practice: the question to the answer? Am J Med. 2018;131(2):129-133.

[2] Turing AM. Computing machinery and intelligence. Mind. 1950;49:433-460.

[3] Sevilla J, Heim L, Ho A, Besiroglu T, Hobbhahn M, Villalobos P. Compute trends across three eras of machine learning. In: Proceedings of the 2022 International Joint Conference on Neural Networks (IJCNN), 18-23 July 2022. 2022:1-8.

[4] Ahmadi R. Emerging innovations in AI-driven medical imaging: elevating diagnostic precision and therapeutic decision-making. JAPMI. 2024;1(1):1-28.

[5] Khalifa M, Albadawy M. AI in diagnostic imaging: revolutionising accuracy and efficiency. Comput Methods Programs Biomed Update. 2024;5:100146.

[6] Ramírez JGC. AI in healthcare: revolutionizing patient care with predictive analytics and decision support systems. JAIGS. 2024;1(1):1-37.

[7] Szeto MD, Alhanshali L, Rundle CW, et al. Dermatologic data from the Global Burden of Disease Study 2019 and the PatientsLikeMe online support community: comparative analysis. JMIR Dermatol. 2024;7:e50449.

[8] Peng D, Sun J, Wang J, Qi X, Li G. Burden of skin disease-China, 1990-2019. China CDC Weekly. 2021;3(22):472-475.

[9] Yardman-Frank JM, Nambudiri V. Ending the pandemic of inaccessibility in dermatology. Clin Dermatol. 2023;41(2):280-283.

[10] Freeman E, Lester J, Padovese V, et al. Introducing the global health and equity section in the British Journal of Dermatology. Br J Dermatol. 2022;186(2):201-202.

[11] Bajaj S, Marchetti MA, Navarrete-Dechent C, Dusza SW, Kose K, Marghoob AA. The role of color and morphologic characteristics in dermoscopic diagnosis. JAMA Dermatol. 2016;152:676-682.

[12] Lowenstein EJ, Sidlow R. Diagnostic heuristics in dermatology, part 2: metacognition and other fixes. Br J Dermatol. 2018;179(6):1270-1276.

[13] Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115-118.

[14] Tschandl P, Rinner C, Apalla Z, et al. Human-computer collaboration for skin cancer recognition. Nat Med. 2020;26(8):1229-1234.

[15] Mollersen K, Kirchesch H, Zortea M, Schopf TR, Hindberg K, Godtliebsen F. Computer-aided decision support for melanoma detection applied on melanocytic and nonmelanocytic skin lesions: a comparison of two systems based on automatic analysis of dermoscopic images. Biomed Res Int. 2015;2015:579282.

[16] Zhou J, He XZ, Sun L, et al. SkinGPT-4: an interactive dermatology diagnostic system with visual large language model. medRxiv. 2023.

[17] Daneshjou R, Barata C, Betz-Stablein B, et al. Checklist for evaluation of image-based artificial intelligence reports in dermatology: CLEAR Derm consensus guidelines from the International Skin Imaging Collaboration Artificial Intelligence Working Group. JAMA Dermatol. 2022;158(1):90-96.

[18] Perednia DA. What dermatologists should know about digital imaging. J Am Acad Dermatol. 1991;25(1):89-108.

[19] Liang J, Han Z, Feng J, et al. Targeted metabolomics combined with machine learning to identify and validate new biomarkers for early SLE diagnosis and disease activity. Clin Immunol. 2024;264:110235.

[20] Liu L, Ju M, Hu Y, Luan C, Zhang J, Chen K. Genome-wide DNA methylation and transcription analysis in psoriatic epidermis. Epigenomics. 2023;15(4):209-226.

[21] Kan H, Nagar S, Patel J, Wallace DJ, Molta C, Chang DJ. Longitudinal treatment patterns and associated outcomes in patients with newly diagnosed systemic lupus erythematosus. Clin Ther. 2016;38(3):610-624.

[22] Yu H, Lin J, Yuan J, Sun X, Wang C, Bai B. Screening mitochondria-related biomarkers in skin and plasma of atopic dermatitis patients by bioinformatics analysis and machine learning. Front Immunol. 2024;15:1367602.

[23] Khozeimeh F, Alizadehsani R, Roshanzamir M, Khosravi A, Layegh P, Nahavandi S. An expert system for selecting wart treatment method. Comput Biol Med. 2017;81:167-175.

[24] Cazzaniga S, Sassi F, Mercuri SR, Naldi L. Prediction of clinical response to excimer laser treatment in vitiligo by using neural network models. Dermatology. 2009;219(2):133-137.

[25] Muddassir M, Limbert G, Navarro-Alarcon D. Development of a numerical multi-layer model of skin subjected to pulsed laser irradiation to optimise thermal stimulation in photorejuvenation procedure. Comput Methods Programs Biomed. 2022;216:106653.

[26] Tsutsui K, Ogata D, Tsukamoto S, et al. Post-immunotherapy robotic-assisted resection for primary anorectal melanoma: a case report. Case Rep Oncol. 2022;15(1):56-61.

[27] Goodman RS, Patrinely JR, Osterman T, Wheless L, Johnson DB. On the cusp: considering the impact of artificial intelligence language models in healthcare. Med. 2023;4(3):139-140.

[28] Nambudiri VE. Decoding the ABCs of natural language processing in dermatology. J Eur Acad Dermatol Venereol. 2024;38(12):2201-2202.

[29] Luger T, Dirschka T, Eyerich K, et al. Developments and challenges in dermatology: an update from the Interactive Derma Academy (IDeA) 2019. J Eur Acad Dermatol Venereol. 2020;34 Suppl 7:3-18.

[30] Bronsnick T, Murzaku EC, Rao BK. Diet in dermatology: part I. atopic dermatitis, acne, and nonmelanoma skin cancer. J Am Acad Dermatol. 2014;71(6):1039.e1-1039.e12.

[31] Adly M, Woo TE, Traboulsi D, Klassen D, Hardin J. Understanding dermatologic concerns among persons experiencing homelessness: a scoping review and discussion for improved delivery of care. J Cutan Med Surg. 2021;25(6):616-626.

[32] Goodman RS, Patrinely JR, Stone CA, et al. Accuracy and reliability of chatbot responses to physician questions. JAMA Netw Open. 2023;6(10):e2336483.

[33] Zhou J, He X, Sun L, et al. Pre-trained multimodal large language model enhances dermatological diagnosis using SkinGPT-4. Nat Commun. 2024;15(1):5649.

[34] Goessinger EV, Gottfrois P, Mueller AM, Cerminara SE, Navarini AA. Image-based artificial intelligence in psoriasis assessment: the beginning of a new diagnostic era? Am J Clin Dermatol. 2024;25(6):861-872.

[35] Deo RC. Machine learning in medicine. Circulation. 2015;132(20):1920-1930.

[36] Wu WT, Li YJ, Feng AZ, et al. Data mining in clinical big data: the frequently used databases, steps, and methodological models. Mil Med Res. 2021;8(1):44.

[37] Talavera-Martinez L, Tschandl P. Automated dermatoscopic pattern discovery by clustering neural network output for human-computer interaction. J Eur Acad Dermatol Venereol. 2023;37(11):e1268-e1270.

[38] Chartrand G, Cheng PM, Vorontsov E, et al. Deep learning: a primer for radiologists. Radiographics. 2017;37(7):2113-2131.

[39] Alzubaidi L, Zhang J, Humaidi AJ, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data. 2021;8(1):53.

[40] Du AX, Emam S, Gniadecki R. Review of machine learning in predicting dermatological outcomes. Front Med (Lausanne). 2020;7:266.

[41] Jones OT, Matin RN, van der Schaar M, et al. Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review. Lancet Digit Health. 2022;4(6):e466-e476.

[42] Xie Z, Tu E, Zheng H, Gu Y, Yang J. Semi-supervised skin lesion segmentation with learning model confidence. In: 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 2021:1135-1139.

[43] Barata C, Rotemberg V, Codella NCF, et al. A reinforcement learning model for AI-based decision support in skin cancer. Nat Med. 2023;29(8):1941-1946.

[44] Thomsen K, Iversen L, Titlestad TL, Winther O. Systematic review of machine learning for diagnosis and prognosis in dermatology. J Dermatolog Treat. 2020;31(5):496-510.

[45] Chan S, Reddy V, Myers B, Thibodeaux Q, Brownstone N, Liao W. Machine learning in dermatology: current applications, opportunities, and limitations. Dermatol Ther (Heidelb). 2020;10(3):365-386.

[46] Puri P, Comfere N, Drage LA, et al. Deep learning for dermatologists: part II. current applications. J Am Acad Dermatol. 2022;87(6):1352-1360.

[47] Erickson BJ, Korfiatis P, Kline TL, Akkus Z, Philbrick K, Weston AD. Deep learning in radiology: does one size fit all? J Am Coll Radiol. 2018;15(3 Pt B):521-526.

[48] Tschandl P, Rosendahl C, Kittler H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci Data. 2018;5:180161.

[49] Zhu CY, Wang YK, Chen HP, et al. A deep learning based framework for diagnosing multiple skin diseases in a clinical environment. Front Med (Lausanne). 2021;8:626369.

[50] Rebouças Filho PP, Peixoto SA, Medeiros da Nóbrega RV, et al. Automatic histologically-closer classification of skin lesions. Comput Med Imaging Graph. 2018;68:40-54.

[51] Bajwa MN, Muta K, Malik MI, et al. Computer-aided diagnosis of skin diseases using deep neural networks. Applied Sciences. 2020;10(1):2488.

[52] Kim A, Saharkhiz N, Sizikova E, et al. S-SYNTH: knowledge-based, synthetic generation of skin images. Springer Nature Switzerland. 2024;LNCS 15003:734-744.

[53] Kamulegeya L, Bwanika J, Okello M, et al. Using artificial intelligence on dermatology conditions in Uganda: a case for diversity in training data sets for machine learning. Afr Health Sci. 2023;23(2):753-763.

[54] Zaar O, Larson A, Polesie S, et al. Evaluation of the diagnostic accuracy of an online artificial intelligence application for skin disease diagnosis. Acta Derm Venereol. 2020;100(16):adv00260.

[55] Nelson CA, Pérez-Chada LM, Creadore A, et al. Patient perspectives on the use of artificial intelligence for skin cancer screening: a qualitative study. JAMA Dermatol. 2020;156(5):501-512.

[56] Hajian-Tilaki K. Receiver operating characteristic (ROC) curve analysis for medical diagnostic test evaluation. Caspian J Intern Med. 2013;4(2):627-635.

[57] Li CX, Fei WM, Shen CB, et al. Diagnostic capacity of skin tumor artificial intelligence-assisted decision-making software in real-world clinical settings. Chin Med J (Engl). 2020;133(17):2020-2026.

[58] Stutz D, Cemgil A, Roy A, et al. Evaluating AI systems under uncertain ground truth: a case study in dermatology. Med Image Anal. 2023;103:1361-8415.

[59] Pillai A, Joseph S, Kreutz J, Trabousi D, Gandhi M, Hardin J. Evaluating the diagnostic and treatment recommendation capabilities of GPT-4 vision in dermatology. 2024.

[60] Rikhye R, Loh A, Hong G, et al. Closing the AI generalization gap by adjusting for dermatology condition distribution differences across clinical settings. arXiv:2402.15566. 2024.

[61] Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402-2410.

[62] Ardila D, Kiraly AP, Bharadwaj S, et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med. 2019;25(6):954-961.

[63] Rodriguez-Ruiz A, Lång K, Gubern-Merida A, et al. Stand-alone artificial intelligence for breast cancer detection in mammography: comparison with 101 radiologists. J Natl Cancer Inst. 2019;111(9):916-922.

[64] Milletarì F, Navab N, Ahmadi S-A. V-Net: fully convolutional neural networks for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV). 2016:565-571.

[65] Liang H, Tsui BY, Ni H, et al. Evaluation and accurate diagnoses of pediatric diseases using artificial intelligence. Nat Med. 2019;25(3):433-438.

[66] Huang L, Sun H, Sun L, et al. Rapid, label-free histopathological diagnosis of liver cancer based on Raman spectroscopy and deep learning. Nat Commun. 2023;14(1):48.

[67] Sturm A, Jóźwiak G, Verge MP, et al. Accurate and rapid antibiotic susceptibility testing using a machine learning-assisted nanomotion technology platform. Nat Commun. 2024;15(1):2037.

[68] Oliveira R, Pereira A, Tavares J. Pattern recognition in macroscopic and dermoscopic images for skin lesion diagnosis. In: Tavares J., Natal Jorge R. (eds). VipIMAGE 2017. ECCOMAS 2017. Lecture Notes in Computational Vision and Biomechanics, vol 27. Cham: Springer. 2018:504-514.

[69] Pathak P, Punetha Y, Kratika . Identification of skin diseases using convolutional neural network. In: Sharma TK, Ahn CW, Verma OP, Panigrahi BK (eds) Soft Computing: Theories and Applications. Advances in Intelligent Systems and Computing, vol 1381. Singapore: Springer. 2021.

[70] Saha A, Suddapalli R, Sahoo AK, Singh J. Leveraging advanced deep learning techniques for automated dermatological image classification. In: 2024 IEEE 4th International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC). Bhubaneswar, India. 2024:1-6.

[71] Sorour S, Hany AA, Elredeny M, Sedik A, Hussien R. An automatic dermatology detection system based on deep learning and computer vision. IEEE Access. 2023;11:137769-137778.

[72] Yang J, Sun X, Liang J, Rosin P. Clinical skin lesion diagnosis using representations inspired by dermatologist criteria. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2018:1258-1266.

[73] Groger F, Gottfrois P, Amruthalingam L, et al. Towards scalable foundation models for digital dermatology. arXiv:2411.05514[cs.CV]. 2024.

[74] Ren Y, Xu W, Mao Y, Wu Y, Fu B, Thanh D. Few‐shot learning for dermatological conditions with lesion area aware Swin Transformer. Int J Imaging Syst Technol. 2023;33(5):1549-1560.

[75] Tan L, Wu H, Xia J, Liang Y, Zhu J. Skin lesion recognition via global-local attention and dual-branch input network. Eng Appl Artif Intell. 2024;127:107385.

[76] Zhou Y, Liu W, Gao Y, et al. A novel multi-task model imitating dermatologists for accurate differential diagnosis of skin diseases in clinical images. arXiv:2307.08308[cs.CV]. 2023:202-212.

[77] Talavera-Martínez L, Bibiloni P, González-Hidalgo M. A multitasking learning framework for dermoscopic image analysis. Cham: Springer. 2021;12702:34-44.

[78] Marri SS, Albadri W, Hyder MS, Janagond AB, Inamadar AC. Efficacy of an artificial intelligence app (Aysa) in dermatological diagnosis: cross-sectional analysis. JMIR Dermatol. 2024;7:e48811.

[79] Wu H, Yin H, Chen H, et al. A deep learning-based smartphone platform for cutaneous lupus erythematosus classification assistance: simplifying the diagnosis of complicated diseases. J Am Acad Dermatol. 2021;85(3):792-793.

[80] Daneshjou R, Vodrahalli K, Novoa RA, et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Sci Adv. 2022;8(32):eabq6147.

[81] Kabir S. SCGAN-generated skin-tone-balanced extension of the HAM10000 dataset for fairer dermatological AI. PLoS One. 2026;21(4):e0331081.

[82] Kinyanjui NM, Odonga T, Cintas C, et al. Estimating skin tone and effects on classification performance in dermatology datasets. In: Proceedings of the Fairness of Classifiers Across Skin Tones in Dermatology (MICCAI Workshop/CVPR Workshop). 2020.

[83] Daneshjou R, et al. SCIN: a new resource for representative dermatology images. Google Research. Updated March 19, 2024.

[84] OpenMedLab. SCIN dataset official release. 2024.

[85] Daneshjou R, Smith MP, Sun MD, Rotemberg V, Zou J. Lack of transparency and potential bias in artificial intelligence data sets and algorithms: a scoping review. JAMA Dermatol. 2021;157(11):1362-1369.

[86] Groh M, Harris C, Soenksen L, et al. Evaluating deep neural networks trained on clinical images in dermatology with the Fitzpatrick 17k dataset. In: Proceedings of the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021:1820-1828.

[87] Bevan PJ, Atapour-Abarghouei A. Detecting melanoma fairly: skin tone detection and debiasing for skin lesion classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2022.

[88] Wen D, Khan SM, Ji Xu A, et al. Characteristics of publicly available skin cancer image datasets: a systematic review. Lancet Digit Health. 2022;4(1):e64-e74.

[89] Li L-F, Wang X, Hu W-J, Xiong NN, Du Y-X, Li B-S. Deep learning in skin disease image recognition: a review. IEEE Access. 2020;8:208264-208280.

[90] Kung TH, Cheatham M, Medenilla A, et al. Performance of ChatGPT on USMLE: potential for AI-assisted medical education using large language models. PLOS Digit Health. 2023;2(2):e0000198.

[91] Shah A, Wahood S, Guermazi D, Brem CE, Saliba E. Skin and syntax: large language models in dermatopathology. Dermatopathology (Basel). 2024;11(1):101-111.

[92] Fan KS, Fan KH. Dermatological knowledge and image analysis performance of large language models based on specialty certificate examination in dermatology. Dermato. 2024;4(4):124-135.

[93] Cirone K, Akrout M, Abid L, Oakley A. Assessing the utility of multimodal large language models (GPT-4 vision and large language and vision assistant) in identifying melanoma across different skin tones. JMIR Dermatol. 2024;7:e55508.

[94] Lin B, Xu Y, Bao X, et al. SkinGEN: an explainable dermatology diagnosis-to-generation framework with interactive vision-language models. arXiv:2404.14755v2[cs.MM]. 2024.

[95] Yin C, Cao Y, Sun P, et al. Molecular subtyping of cancer based on robust graph neural network and multi-omics data integration. Front Genet. 2022;13:884028.

[96] Lemsara A, Ouadfel S, Fröhlich H. PathME: pathway based multi-modal sparse autoencoders for clustering of patient-level multi-omics data. BMC Bioinformatics. 2020;21(1):146.

[97] Chierici M, Bussola N, Marcolini A, et al. Integrative network fusion: a multi-omics approach in molecular profiling. Front Oncol. 2020;10:1065.

[98] Liu Y, Wang H, Taylor M, et al. Classification of human chronic inflammatory skin disease based on single-cell immune profiling. Sci Immunol. 2022;7(70):eabl9165.

[99] Yu T, Xu X, Liu Y, et al. Multi-omics signatures reveal genomic and functional heterogeneity of Cutibacterium acnes in normal and diseased skin. Cell Host Microbe. 2024;32(7):1129-1146.e8.

[100] Dessì A, Pintus R, Fanos V, Bosco A. Integrative multiomics approach to skin: the sinergy between individualised medicine and futuristic precision skin care? Metabolites. 2024;14(3):157.

[101] Chen M, Wang R, Wang T. Gut microbiota and skin pathologies: mechanism of the gut-skin axis in atopic dermatitis and psoriasis. Int Immunopharmacol. 2024;141:112658.

[102] Wu D, Wang D, Zhang MQ, Gu J. Fast dimension reduction and integrative clustering of multi-omics data using low-rank approximation: application to cancer molecular classification. BMC Genomics. 2015;16:1022.

[103] Domingues I, Muller H, Ortiz A, Dasarathy BV, Abreu PH, Calhoun VD. Guest editorial: information fusion for medical data: early, late, and deep fusion methods for multimodal data. IEEE J Biomed Health Inform. 2020;24(1):14-16.

[104] Dreyer L, Faurschou M, Mogensen M, Jacobsen S. High incidence of potentially virus-induced malignancies in systemic lupus erythematosus: a long-term followup study in a Danish cohort. Arthritis Rheum. 2011;63(10):3032-3037.

[105] Petit A, Dadzie OE. Multisystemic diseases and ethnicity: a focus on lupus erythematosus, systemic sclerosis, sarcoidosis and Behçet disease. Br J Dermatol. 2013;169 Suppl 3:1-10.

[106] Bulger DA, Minhas S, Asbeutah AA, et al. Chronic systemic inflammatory skin disease as a risk factor for cardiovascular disease. Curr Probl Cardiol. 2021;46(5):100799.

[107] Vassileva S, Drenovska K, Manuelyan K. Autoimmune blistering dermatoses as systemic diseases. Clin Dermatol. 2014;32(3):364-375.

[108] Hao Y, Hao S, Andersen-Nissen E, et al. Integrated analysis of multimodal single-cell data. Cell. 2021;184(13):3573-3587.e29.

[109] Tan W, Tiwari P, Pandey HM, Moreira C, Jaiswal AK. Multimodal medical image fusion algorithm in the era of big data. Neural Comput Appl. 2025;37:22995-23015.

[110] AlSaad R, Abd-Alrazaq A, Boughorbel S, et al. Multimodal large language models in health care: applications, challenges, and future outlook. J Med Internet Res. 2024;26:e59505.

[111] Fiaidhi J, Mohammed S. Multimodal learning for supporting evidence base medicine. In: Proceedings of the 2024 Prognostics and System Health Management Conference (PHM), 28-31 May 2024. 2024:14-19.

[112] Eliseev M, Zhelyabina O, Vladimirov S, et al. SAT0646 Comparison of US, CT, X-RAY and MRI efficacy for sequential assessment of chronic gout manifestations. Ann Rheum Dis. 2017;76(Supplement 2):1019.

[113] Gound RS, Gadre PS, Gaikwad JB, Wagh PK. Skin disease diagnosis system using image processing and data mining. Int J Comput Appl. 2018;179(16):38-40.

[114] Gan L, Wu X, Song J. Comprehensive analysis of crucial m⁶A-Related differentially expressed genes in psoriasis. Front Biosci (Landmark Ed). 2024;29(9):311.

[115] Castillo RL, Sidhu I, Dolgalev I, et al. Spatial transcriptomics stratifies psoriatic disease severity by emergent cellular ecosystems. Sci Immunol. 2023;8(84):eabq7991.

[116] Fan F, Huang Z, Chen Y. Integrated analysis of immune-related long noncoding RNAs as diagnostic biomarkers in psoriasis. PeerJ. 2021;9:e11018.

[117] Yu Z, Gong Y, Cui L, et al. High-throughput transcriptome and pathogenesis analysis of clinical psoriasis. J Dermatol Sci. 2020;98(2):109-118.

[118] Guo Y, Luo L, Zhu J, Li C. Multi-omics research strategies for psoriasis and atopic dermatitis. Int J Mol Sci. 2023;24(9):8018.

[119] Mäder U, Quiskamp N, Wildenhain S, et al. Image-processing scheme to detect superficial fungal infections of the skin. Comput Math Methods Med. 2015;2015:851014.

[120] Wanat KA, Dominguez AR, Carter Z, Legua P, Bustamante B, Micheletti RG. Bedside diagnostics in dermatology: viral, bacterial, and fungal infections. J Am Acad Dermatol. 2017;77(2):197-218.

[121] He H, Bissonnette R, Wu J, et al. Tape strips detect distinct immune and barrier profiles in atopic dermatitis and psoriasis. J Allergy Clin Immunol. 2021;147(1):199-212.

[122] Kawahara J, Daneshvar S, Argenziano G, Hamarneh G. 7-point checklist and skin lesion classification using multi-task multi-modal neural nets. IEEE J Biomed Health Inform. 2018.

[123] Zhang Y, Hu Y, Li K, Pan X, Mo X, Zhang H. Exploring the influence of transformer-based multimodal modeling on clinicians’ diagnosis of skin diseases: a quantitative analysis. Digit Health. 2024;10:20552076241257087.

[124] Ivanenkov YA, Polykovskiy D, Bezrukov D, et al. Chemistry42: an AI-driven platform for molecular design and optimization. J Chem Inf Model. 2023;63(3):695-701.

[125] Kamya P, Ozerov IV, Pun FW, et al. PandaOmics: an AI-driven platform for therapeutic target and biomarker discovery. J Chem Inf Model. 2024;64(10):3961-3969.

[126] Ren F, Aliper A, Chen J, et al. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nat Biotechnol. 2025;43(1):63-75.

[127] Li P, Zhang K, Liu T, et al. A deep learning approach for rational ligand generation with toxicity control via reactive building blocks. Nat Comput Sci. 2024;4(11):851-864.

[128] Orcutt X, Chen K, Mamtani R, Long Q, Parikh RB. Evaluating generalizability of oncology trial results to real-world patients using machine learning-based trial emulations. Nat Med. 2025;31(2):457-465.

[129] Zhao L, Chen J, Bai B, et al. Topical drug delivery strategies for enhancing drug effectiveness by skin barriers, drug delivery systems and individualized dosing. Front Pharmacol. 2023;14:1333986.

[130] Zhang Y, Wang Y, Wu C, et al. Drug-target interaction prediction by integrating heterogeneous information with mutual attention network. BMC Bioinformatics. 2024;25(1):361.

[131] Yan S, Yu Z, Primiero C, et al. A general-purpose multimodal foundation model for dermatology. arXiv:2410.15038v1[cs.CV]. 2024.

[132] Maulana A, Noviandy TR, Suhendra R, et al. Enhanced prediction of atopic dermatitis severity using advanced machine learning techniques. In: Proceedings of the 2024 International Conference on Electrical Engineering and Informatics (ICELTICs), 12-13 Sept. 2024. 2024:164-169.

[133] Shi H, Xu T, Li X, et al. DRExplainer: quantifiable interpretability in drug response prediction with directed graph convolutional network. Artif Intell Med. 2025;163:103101.

[134] He B, Xiao Y, Liang H, et al. ASGARD is a single-cell guided pipeline to aid repurposing of drugs. Nat Commun. 2023;14(1):993.

[135] Lee S, Liu R, Cheng F, Zhang P. A deep subgrouping framework for precision drug repurposing via emulating clinical trials on real-world patient data. KDD. 2025;2025(v1):2347-2358.

[136] Bang D, Koo B, Kim S. Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction. Bioinformatics. 2024;40(Suppl 1):i130-i139.

[137] Subramanian A, Narayan R, Corsello SM, et al. A next generation connectivity map: L1000 platform and the first 1,000,000 profiles. Cell. 2017;171(6):1437-1452.e17.

[138] Wen Q, Liu R, Zhang P. Clinical connectivity map for drug repurposing: using laboratory results to bridge drugs and diseases. BMC Med Inform Decis Mak. 2021;21(Suppl 8):263.

[139] Liu R, Wei L, Zhang P. When deep learning meets causal inference: a computational framework for drug repurposing from real-world data. arXiv:2007.10152 [stat.AP]. 2020.

[140] Inoue Y, Song T, Fu T. DrugAgent: explainable drug repurposing agent with large language model-based reasoning. arXiv:2408.13378v3[cs.AI]. 2024.

[141] Greugny E, Fages F, Radulescu O, Szmolyan P, Stamatas G. A skin microbiome model with AMP interactions and analysis of quasi-stability vs stability in population dynamics. Theor Comput Sci. 2023;983(1):114294.

[142] Velasco JS, Catipon JV, Monilar EG, Amon VM, Virrey GC, Tolentino LKS. Classification of skin disease using transfer learning in convolutional neural networks. Int J Emerg Technol Adv Eng. 2023;13(4):1-7.

[143] Khan MS, Alam KN, Dhruba AR, Zunair H, Mohammed N. Knowledge distillation approach towards melanoma detection. Comput Biol Med. 2022;146:105581.

[144] Han Y, Qin X, Lin W, et al. Microneedle-based approaches for skin disease treatment. Nanomicro Lett. 2025;17(1):132.

[145] Zhavoronkov A, Ivanenkov YA, Aliper A, et al. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nat Biotechnol. 2019;37(9):1038-1040.

[146] Deutelmoser H, Scherer D, Brenner H, et al. Robust Huber-LASSO for improved prediction of protein, metabolite and gene expression levels relying on individual genotype data. Brief Bioinform. 2021;22(4):bbaa230.

[147] Melbin K, Raj Y. Integration of modified ABCD features and support vector machine for skin lesion types classification. Multimed Tools Appl. 2020;80:8909-8929.

[148] Vera-Salmerón E, Domínguez-Nogueira C, Romero-Béjar JL, Sáez JA, Mota-Romero E. Decision-tree-based approach for pressure ulcer risk assessment in immobilized patients. Int J Environ Res Public Health. 2022;19(18):11161.

[149] Balaji V, Suganthi S, Rajadevi R, Kumar V, Saravana Balaji B, Pandiyan S. Skin disease detection and segmentation using dynamic graph cut algorithm and classification through Naive Bayes classifier. Measurement. 2020;163:107922.

[150] Mayank SMK. Enhancing skin cancer detection through deep learning techniques. Int Sci J Eng Manag. 2024.

[151] Abhishek K, Kawahara J, Hamarneh G. Predicting the clinical management of skin lesions using deep learning. Sci Rep. 2021;11(1):7769.

[152] Nguyen TH, H T, Mai T, et al. A lightweight model to skin disease recognition. Res Dev Inf Commun Technol. 2024.

[153] Gakire M. The use of machine learning in predicting disease progression. Newport Int J Res Med Sci. 2024.

[154] Picard M, Scott-Boyer MP, Bodein A, Périn O, Droit A. Integration strategies of multi-omics data for machine learning analysis. Comput Struct Biotechnol J. 2021;19:3735-3746.

[155] Ravi V. Deep learning-based skin cancer diagnosis using dermoscopy images. In: 2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC-ROBINS). 2024:8-13.

[156] Ghassemi M, Naumann T, Schulam P, Beam AL, Chen IY, Ranganath R. Practical guidance on artificial intelligence for health-care data. Lancet Digit Health. 2019;1(4):e157-e159.

[157] Zhang ZF, Yang ZX, Huang YY, Zhan JF. Big medical data and medical AI standards: status quo, opportunities and challenges. Med J Peking Un Med Coll Hosp. 2021;12(5):614-620..

[158] Hallinan CM, Ward R, Hart GK, et al. Seamless EMR data access: integrated governance, digital health and the OMOP-CDM. BMJ Health Care Inform. 2024;31(1):e100953.

[159] Borisov V, Minin A, Basko V, Syskov A. FHIR data model for intelligent multimodal interface. In: 2018 26th Telecommunications Forum (TELFOR). 2018:420-425.

[160] Brehmer A, Sauer CM, Salazar Rodríguez J, et al. Establishing medical intelligence-leveraging Fast Healthcare Interoperability Resources to improve clinical management: retrospective cohort and clinical implementation study. J Med Internet Res. 2024;26:e55148.

[161] Gruendner J, Schwachhofer T, Sippl P, et al. KETOS: clinical decision support and machine learning as a service-a training and deployment platform based on Docker, OMOP-CDM, and FHIR web services. PLoS One. 2019;14(10):e0223010.

[162] Alterovitz G, Warner J, Zhang P, et al. SMART on FHIR Genomics: facilitating standardized clinico-genomic apps. J Am Med Inform Assoc. 2015;22(6):1173-1178.

[163] Gruendner J, Deppenwiese N, Folz M, et al. The architecture of a feasibility query portal for distributed COVID-19 Fast Healthcare Interoperability Resources (FHIR) patient data repositories: design and implementation study. JMIR Med Inform. 2022;10(5):e36709.

[164] Klann JG, Joss MAH, Embree K, Murphy SN. Data model harmonization for the All of Us Research Program: transforming i2b2 data into the OMOP Common Data Model. PLoS One. 2019;14(2):e0212463.

[165] Jeon K, Park WY, Kahn CE, Nagy P, You SC, Yoon SH. Advancing medical imaging research through standardization: the path to rapid development, rigorous validation, and robust reproducibility. Invest Radiol. 2025;60(1):1-10.

[166] Kalokyri V, Kondylakis H, Sfakianakis S, et al. MI-Common data model: extending Observational Medical Outcomes Partnership-Common Data Model (OMOP-CDM) for registering medical imaging metadata and subsequent curation processes. JCO Clin Cancer Inform. 2023;7:e2300101.

[167] Trakatelli M, Richard MA, Rouillard A, Paul C, Röcken M, Stratigos A. The burden of skin disease in Europe. J Eur Acad Dermatol Venereol. 2023;37 Suppl 7:3-5.

[168] Lee JJ, English JC. Teledermatology: a review and update. Am J Clin Dermatol. 2018;19(2):253-260.

[169] Rizvi SMH, Schopf T, Sangha A, Ulvin K, Gjersvik P. Teledermatology in Norway using a mobile phone app. PLoS One. 2020;15(4):e0232131.

[170] Chiou AS, Omiye JA, Gui H, et al. Multimodal image dataset for AI-based skin cancer (MIDAS) benchmarking. NEJM AI. 2024.

[171] Khan MA, Muhammad K, Sharif M, Akram T, Albuquerque VHC. Multi-class skin lesion detection and classification via teledermatology. IEEE J Biomed Health Inform. 2021;25(12):4267-4275.

[172] Gao X, He P, Zhou Y, Qin X. A smart healthcare system for remote areas based on the edge-cloud continuum. Electronics. 2024;13(21):4152.

[173] Kalasin S, Surareungchai W. Challenges of emerging wearable sensors for remote monitoring toward telemedicine healthcare. Anal Chem. 2023.

[174] Plazzotta F, Sommer JA, Marquez Fosser SN, Luna DR. Asynchronous dermatology teleconsultations using a personal health record. Stud Health Technol Inform. 2018;247:690-694.

[175] Sud E, Anjankar A. Applications of telemedicine in dermatology. Cureus. 2022;14(8):e27740.

[176] Greis C, Meier Zürcher C, Djamei V, Moser A, Lautenschlager S, Navarini AA. Unmet digital health service needs in dermatology patients. J Dermatolog Treat. 2018;29(7):643-647.

[177] Nam Y, Kim J, Jung SH, et al. Harnessing artificial intelligence in multimodal omics data integration: paving the path for the next frontier in precision medicine. Annu Rev Biomed Data Sci. 2024;7(1):225-250.

[178] Giansanti D. Advancing dermatological care: a comprehensive narrative review of tele-dermatology and mHealth for bridging gaps and expanding opportunities beyond the COVID-19 pandemic. Healthcare (Basel). 2023;11(13):1911.

[179] Yim W-W, Abacha AB, Fu Y, Sun Z, Yetisgen M, Xia F. Overview of the MEDIQA-MAGIC Task at ImageCLEF 2024: multimodal and generative telemedicine in dermatology. In CLEF 2024 Conference and Labs of the Evaluation Forum. 2024:1456-1462.

[180] Habgood E, McCormack C, Walter FM, Emery JD. Patients’ experiences of using skin self-monitoring apps with people at higher risk of melanoma: qualitative study. JMIR Dermatol. 2021;4(2):e22583.

[181] Smak Gregoor AM, Sangers TE, Eekhof JA, et al. Artificial intelligence in mobile health for skin cancer diagnostics at home (AIM HIGH): a pilot feasibility study. EClinicalMedicine. 2023;60:102019.

[182] Rebedew D. VisualDx. Fam Pract Manag. 2015;22(1):33.

[183] Hasan MK, Elahi MTE, Alam MA, Jawad MT, Martí R. DermoExpert: skin lesion classification using a hybrid convolutional neural network through segmentation, transfer learning, and augmentation. Informatics in Medicine Unlocked. 2022;28:100819.

[184] Huang K, Jiang Z, Li Y, et al. The classification of six common skin diseases based on Xiangya-derm: development of a Chinese database for artificial intelligence. J Med Internet Res. 2021;23(9):e26025.

[185] Ward A, Li J, Wang J, et al. Creating an empirical dermatology dataset through crowdsourcing with web search advertisements. JAMA Netw Open. 2024;7(11):e2446615.

[186] Liu Y, Jain A, Eng C, et al. A deep learning system for differential diagnosis of skin diseases. Nat Med. 2020;26(6):900-908.

[187] Dhiman P. Dermonet: lightweight diagnostic system for dermatological conditions using SqueezeNet framework. J Mech Contin Math Sci. 2024;(11):186-199.

[188] Trinh P, Yekrang K, Phung M, et al. Partnering with a senior living community to optimise teledermatology via full body skin screening during the COVID-19 pandemic: a pilot programme. Skin Health Dis. 2022;2(3):e141.

[189] Guido N, Hagstrom E, Ibler E, et al. A novel total body digital photography smartphone application designed to detect and monitor skin lesions: a pilot study. J Surg Dermatol. 2018;3(2):1-5.

[190] Kommoss KS, Winkler JK, Mueller-Christmann C, et al. Observational study investigating the level of support from a convolutional neural network in face and scalp lesions deemed diagnostically ‘unclear’ by dermatologists. Eur J Cancer. 2023;185:53-60.

[191] Sies K, Winkler JK, Fink C, et al. Past and present of computer-assisted dermoscopic diagnosis: performance of a conventional image analyser versus a convolutional neural network in a prospective data set of 1,981 skin lesions. Eur J Cancer. 2020;135:39-46.

[192] Haenssle HA, Fink C, Toberer F, et al. Man against machine reloaded: performance of a market-approved convolutional neural network in classifying a broad spectrum of skin lesions in comparison with 96 dermatologists working under less artificial conditions. Ann Oncol. 2020;31(1):137-143.

[193] Li H, Zhang P, Wei Z, et al. Deep skin diseases diagnostic system with dual-channel image and extracted text. Front Artif Intell. 2023;6:1213620.

[194] Morton C, Chin Gillian XM, Carse J, McKenna S, Matin R, Fleming C. Teledermatology & digital dermatology symposiumoral presentationsBT01 developing a national artificial intelligence-assisted skin cancer pathway. Br J Dermatol. 2024;191(Supplement_1):i189-i189.

[195] Marsden H, Morgan C, Austin S, et al. Effectiveness of an image analyzing AI-based digital health technology to identify non-melanoma skin cancer and other skin lesions: results of the DERM-003 study. Front Med (Lausanne). 2023;10:1288521.

[196] Goreski H, Ilic D, Flacher V, et al. NETSKINMODELS: A European network for skin engineering and modeling. J Invest Dermatol. 2025;145(1):15-17.

[197] Wang Y, Aldahdooh J, Hu Y, et al. DrugRepo: a novel approach to repurposing drugs based on chemical and genomic features. Sci Rep. 2022;12(1):21116.

[198] Pan Y, Chen Z, Qi F, Liu J. Identification of drug compounds for keloids and hypertrophic scars: drug discovery based on text mining and DeepPurpose. Ann Transl Med. 2021;9(4):347.

[199] Chen S, Zeng J, Saad M, Lineaweaver WC, Chen Z, Pan Y. Precision drug repurposing: a deep learning toolkit for identifying 34 hyperpigmentation-associated genes and optimizing treatment selection. Ann Plast Surg. 2024;93(2S Suppl 1):S30-S42.

[200] Pandiarajan R, Hariharan V, Abinesh MS, Harrish N. DeepSkinNet: ReLU-enhanced CNNs for AI-based skin disease detection. In: Proceedings of the 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT). 2024:1-9.

[201] Josphineleela R, Raja Rao PBV, Shaikh A, Sudhakar K. A multi-stage Faster RCNN-based iSPLInception for skin disease classification using novel optimization. J Digit Imaging. 2023;36(5):2210-2226.

[202] Zhang D, Li A, Wu W, Yu L, Kang X, Huo X. CR-Conformer: a fusion network for clinical skin lesion classification. Med Biol Eng Comput. 2024;62(1):85-94.

[203] Wu H, Yin H, Chen H, et al. A deep learning, image based approach for automated diagnosis for inflammatory skin diseases. Ann Transl Med. 2020;8(9):581.

[204] Li Q, Yang Z, Chen K, et al. Human-multimodal deep learning collaboration in ‘precise’ diagnosis of lupus erythematosus subtypes and similar skin diseases. J Eur Acad Dermatol Venereol. 2024;38(12):2268-2279.

[205] Yin H, Chen H, Zhang W, et al. Image-based remote evaluation of PASI scores with psoriasis by the YOLO-v4 algorithm. Exp Dermatol. 2024;33(4):e15082.

[206] Yu KH, Healey E, Leong TY, Kohane IS, Manrai AK. Medical artificial intelligence and human values. N Engl J Med. 2024;390(20):1895-1904.

[207] Meier LJ. Predicting patient preferences with artificial intelligence: the problem of the data source. Am J Bioeth. 2024;24(7):48-50.

[208] Sun HW, Zhang X, Shen CC. The shared circulating diagnostic biomarkers and molecular mechanisms of systemic lupus erythematosus and inflammatory bowel disease. Front Immunol. 2024;15:1354348.

[209] Steardo L, Carbone EA, de Filippis R, et al. Application of support vector machine on fMRI data as biomarkers in schizophrenia diagnosis: a systematic review. Front Psychiatry. 2020;11:588.

[210] Wang DC, Xu WD, Wang SN, et al. Lupus nephritis or not? A simple and clinically friendly machine learning pipeline to help diagnosis of lupus nephritis. Inflamm Res. 2023;72(6):1315-1324.

[211] Koppe G, Guloksuz S, Reininghaus U, Durstewitz D. Recurrent neural networks in mobile sampling and intervention. Schizophr Bull. 2019;45(2):272-276.

[212] Van Steenkiste T, Ruyssinck J, De Baets L, et al. Accurate prediction of blood culture outcome in the intensive care unit using long short-term memory neural networks. Artif Intell Med. 2019;97:38-43.

[213] Rajabi E, Sahebari M, Thomas T. Analyzing systemic lupus erythematosus publications using neural network-based multi-label classification algorithms. Lupus. 2022;31(7):820-827.

[214] Konno N, Kijima Y, Watano K, et al. Deep distributed computing to reconstruct extremely large lineage trees. Nat Biotechnol. 2022;40(4):566-575.

[215] Langmead B, Nellore A. Cloud computing for genomic data analysis and collaboration. Nat Rev Genet. 2018;19(4):208-219.

[216] Iyer V, Ganti B, Vyshnavi AH, Namboori K, Iyer S. Hybrid quantum computing based early detection of skin cancer. J Interdiscip Math. 2020;23(2):347-355.

[217] Gerke S, Babic B, Evgeniou T, Cohen IG. The need for a system view to regulate artificial intelligence/machine learning-based software as medical device. NPJ Digit Med. 2020;3:53.

[218] Aboy M, Minssen T, Vayena E. Navigating the EU AI Act: implications for regulated digital medical products. NPJ Digit Med. 2024;7(1):237.

[219] Finlayson SG, Subbaswamy A, Singh K, et al. The clinician and dataset shift in artificial intelligence. N Engl J Med. 2021;385(3):283-286.

[220] Davis SE, Lasko TA, Chen G, Siew ED, Matheny ME. Calibration drift in regression and machine learning models for acute kidney injury. J Am Med Inform Assoc. 2017;24(6):1052-1061.

[221] Price WN, Gerke S, Cohen IG. Potential liability for physicians using artificial intelligence. JAMA. 2019;322(18):1765-1766.

[222] Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. J Am Med Inform Assoc. 2012;19(1):121-127.

[223] Vasey B, Ursprung S, Beddoe B, et al. Association of clinician diagnostic performance with machine learning-based decision support systems: a systematic review. JAMA Netw Open. 2021;4(3):e211276.

[224] Gomez Rossi J, Rojas-Perilla N, Krois J, Schwendicke F. Cost-effectiveness of artificial intelligence as a decision-support system applied to the detection and grading of melanoma, dental caries, and diabetic retinopathy. JAMA Netw Open. 2022;5(3):e220269.

[225] Li J, Guan Z, Wang J, et al. Integrated image-based deep learning and language models for primary diabetes care. Nat Med. 2024;30(10):2886-2896.

[226] Zhao Y, Xiong S, Ren Q, et al. Deep learning using histological images for gene mutation prediction in lung cancer: a multicentre retrospective study. Lancet Oncol. 2025;26(1):136-146.

[227] Nguyen E, Poli M, Durrant MG, et al. Sequence modeling and design from molecular to genome scale with Evo. Science. 2024;386(6723):eado9336.

[228] Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health. 2021;3(11):e745-e750.

Rights & permissions

The Author(s) 2026. This article is published by Higher Education Press on behalf of People’s Medical Publishing House.