Objective assessment of rosacea erythema severity: a multimodal artificial intelligence framework integrating VISIA®...

Rosacea is a chronic inflammatory skin disorder characterized by persistent facial erythema. Its clinical assessment relies on the Clinician's Erythema Assessment (CEA), a subjective scale prone to inter-observer variability. To address the need for diagnostic consistency, this study developed a multimodal artificial intelligence framework for objective CEA grading using standardized VISIA® imaging. We analyzed a retrospective cohort of 1,001 patients. To establish a robust reference standard, three expert dermatologists conducted a multi-step collective audit to reach a unanimous consensus for each case. The framework integrated handcrafted image-derived tabular features with deep learning representations. During training, spatial data augmentations and Focal Loss were implemented to address dataset imbalance and mitigate overfitting. Our results demonstrated that the multimodal fusion model achieved statistically significant improvements over the strong image-only baseline (McNemar's $p=0.031$ ; DeLong's $p=0.024$ ), yielding a Macro-AUC of 0.902 (95% CI: 0.862-0.937). Furthermore, to address the ordinal nature of the disease severity, the fusion model achieved a Quadratic Weighted Kappa (QWK) of 0.800 and an Intraclass Correlation Coefficient (ICC) of 0.801 (95% CI: 0.720-0.860), indicating excellent alignment with expert consensus. Error analysis revealed that over 95% of misclassifications in intermediate grades (CEA3) were restricted to adjacent categories, reflecting strong clinical safety. Interpretability analysis via layer-wise relevance propagation confirmed the model's focus on clinically recognized erythema-prone regions. This study establishes a robust proof-of-concept tool that transforms rosacea assessment from subjective inspection into an objective digital measurement, offering significant translational potential for clinical trials and teledermatology.