Introduction
Atopic dermatitis (AD) is a chronic inflammatory skin disorder characterized by eczematous lesions, intense pruritus, and a relapsing course. Facial involvement in AD is particularly distressing for patients due to cosmetic and psychological impact, posing a significant disease burden. Dupilumab, a monoclonal antibody targeting IL-4 and IL-13 pathways, has demonstrated efficacy in moderate-to-severe AD including facial lesions. However, therapeutic response is heterogeneous, and reliable early predictors of treatment outcomes are lacking, hindering personalized management.
Study Background
Accurate and non-invasive methods to anticipate dupilumab response are urgently needed to optimize treatment strategies, avoid unnecessary exposure, and manage patient expectations. Reflectance confocal microscopy (RCM) allows in vivo visualization of skin morphology at near histological resolution without tissue damage, offering unique insights into microstructural changes under therapy. Incorporating deep learning techniques to analyze complex imaging data has the potential to augment predictive accuracy.
Study Design and Methods
This prospective cohort study enrolled 52 patients with facial AD treated with dupilumab between May 2023 and June 2025. Patients were classified as responders or non-responders based on achievement of at least 75% improvement in the Eczema Area and Severity Index (EASI-75) at week 16. RCM imaging was performed on facial lesions at baseline and week 4 post-treatment to capture early morphological features.
A deep learning model based on the ResNet101 architecture was trained on week 4 RCM images to discriminate responders from non-responders. Model performance was assessed via the area under the receiver operating characteristic curve (AUC). Key RCM features were also analyzed individually for their association with treatment response.
Key Findings
The study identified dermal papillary dilation and tortuous dilation of papillary capillaries on RCM at week 4 as significant predictors of poor therapeutic response, with AUC values of 0.83 (P<0.001) and 0.81 (P<0.001), respectively. These findings suggest that persistent microvascular alterations early after treatment initiation portend resistance to dupilumab.
The ResNet101 deep learning model, trained exclusively on week 4 post-treatment RCM images, achieved a robust predictive performance with a test AUC of 0.796. Heatmap visualizations of the model's decision-making highlighted microvascular and dermal structural patterns as critical features driving classification.
The integration of RCM imaging with advanced machine learning thus enables an effective, non-invasive assessment tool for early identification of patients unlikely to benefit from dupilumab therapy on facial AD lesions.
Limitations
The relatively small sample size (52 patients) limits the generalizability of findings. Larger and multi-center studies are needed to validate and refine the predictive model. Additionally, the study focused on facial lesions, and applicability to other body sites warrants further investigation. The short interval (4 weeks) for early prediction is a strength but requires confirmation regarding long-term outcomes.
Expert Commentary
This study innovatively combines reflectance confocal microscopy with deep learning to address a critical unmet need in AD management—early prediction of biologic treatment response. The use of non-invasive imaging markers correlated with microvascular remodeling offers mechanistic plausibility, as dupilumab’s modulation of type 2 inflammation could influence dermal vascular changes. Although promising, the approach necessitates sophisticated imaging resources and computational expertise, which may limit immediate clinical adoption.
The distinction between responders and non-responders at week 4 is clinically valuable, allowing potential early treatment modification and optimization of healthcare resources. Future integration with serum biomarkers and clinical phenotyping could enhance predictive accuracy further.
Conclusion
Early post-treatment RCM features, specifically dermal papillary and tortuous papillary capillary dilation, combined with ResNet101-based deep learning analysis, provide a promising non-invasive approach to predict dupilumab response in facial atopic dermatitis. This methodology supports personalized treatment pathways by identifying patients who may require alternative therapies early in their treatment course. Larger validation studies are needed to confirm these encouraging results and facilitate translation into routine dermatologic practice.
Funding and Clinical Trial Registration
The study was supported by institutional grants from participating dermatology departments. No clinical trial registration number was indicated.
References
1. Simpson EL et al. Two phase 3 trials of dupilumab versus placebo in atopic dermatitis. N Engl J Med. 2016;375(24):2335-2348.
2. Wollenberg A et al. Early improvement and predictors of response to dupilumab in patients with atopic dermatitis. J Eur Acad Dermatol Venereol. 2021;35(4):e219-e221.
3. Rajadhyaksha M et al. In vivo confocal scanning laser microscopy of human skin I: Morphologic features of normal skin. J Invest Dermatol. 1995;104(6):946-952.
4. He K et al. Deep Residual Learning for Image Recognition. 2016 IEEE CVPR.
5. Beck LA et al. Dupilumab treatment in adults with moderate-to-severe atopic dermatitis. N Engl J Med. 2014;371(2):130-139.
