Routine Metabolic Laboratory Profile and Time-Stratified Machine-Learning Risk Stratification of Rosacea: A Combined Retrospective-Prospective Cohort Study

Rosacea has been linked with systemic metabolic abnormalities in case-control studies, but most such evidence compares rosacea patients with healthy controls rather than asking which patients within a rosacea population carry the greatest metabolic burden. We aimed to describe the metabolic profile of a real-world rosacea outpatient cohort and build a time-stratified, outcome-independent machine-learning model ranking which routine metabolic indicators predict metabolic syndrome status.

We conducted a combined retrospective-prospective study of 298 outpatients with rosacea, seen at a tertiary dermatology clinic from March 2021 to March 2026, contributing 5420 laboratory records (repeated measurements; 298 unique patients) across nine metabolic indicators. The outcome was metabolic syndrome status, defined with a modified NCEP-ATP III definition that includes fasting glucose, triglycerides, blood pressure and waist circumference (Laboratory Variables); fasting glucose and triglycerides were then excluded from the predictors; a random forest classifier was built from the remaining seven indicators, trained on patients enrolled in 2021-2022 (n=164) and evaluated on a time-stratified holdout enrolled in 2022-2023 (n=134). A subset of 83 patients with ≥2 follow-up visits was used to examine within-patient change over time.

Abnormal fasting glucose occurred in 23.7% of records and abnormal triglycerides in 31.4%. Patients with metabolic syndrome (n=64, 21.5%) had higher CRP (6.8±3.1 vs 2.4±1.6 mg/L, P<0.001) and higher severity scores (14.2±3.5 vs 10.1±3.0, P=0.003). On the holdout, the random forest model reached an AUC of 0.734 (95% CI 0.628-0.824), with CRP, total cholesterol, ALT, and HbA1c as leading contributors. Change in fasting glucose correlated with change in severity score on follow-up (ρ=0.31, P=0.004).

Outpatients with rosacea frequently show systemic metabolic disturbances aligned with metabolic syndrome status and clinical severity. A machine-learning model built from routinely available, non-diagnostic indicators retained discrimination on a time-stratified holdout, suggesting a low-cost approach to flagging metabolic risk in this population, which requires external validation in independent, multicenter cohorts before clinical use.