Data Sheet 1: Objective Rosacea Erythema Severity Assessment via Multimodal AI

A retrospective cohort of 1,001 patients with rosacea was analyzed to develop an objective severity assessment tool. The dataset, created by Jun Zhang and last updated in May 2026, integrates handcrafted image-derived tabular features with deep learning representations from standardized VISIA® imaging. It serves as a proof-of-concept for transforming subjective clinical inspection into an objective digital measurement.

Use Cases

• Training multimodal AI models for objective Clinician’s Erythema Assessment (CEA) grading based on VISIA® imaging and derived features.
• Benchmarking model performance using metrics like Quadratic Weighted Kappa and Intraclass Correlation Coefficient against expert consensus.
• Conducting error analysis for clinical safety, focusing on misclassifications restricted to adjacent severity categories.
• Performing interpretability analysis via layer-wise relevance propagation to validate model focus on erythema-prone regions.

Strengths

• Dataset is based on a cohort of 1,001 patients, providing a substantial sample size.
• Reference standard established by unanimous consensus from three expert dermatologists.
• The described multimodal fusion model achieved a Macro-AUC of 0.902 and a QWK of 0.800, indicating strong performance.
• Error analysis shows over 95% of misclassifications for intermediate grades were restricted to adjacent categories.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect temporal or source bias inherent to the retrospective, single-study cohort.

Provenance

Source

Jun Zhang via figshare

Collection Method

Retrospective analysis of patient imaging data with expert consensus labeling.

Freshness

Last updated 2026-05-11 05:26:08; freshness should be verified.