Abstract
BACKGROUND: The efficacy of dupilumab in atopic dermatitis (AD) has been widely validated; however, systematic investigations into treatment adherence are lacking. OBJECTIVE: To analyze clinical factors influencing dupilumab adherence in patients with moderate-to-severe AD and develop a multidimensional adherence prediction model to support precision management of biologic therapies. METHODS: Using a single-center prospective cohort, a three-stage modeling approach was applied: (1) univariable Cox proportional hazards regression to identify potential predictors; (2) XGBoost modeling with SHAP method for feature importance ranking and dimensionality reduction; (3) multivariable Cox proportional hazards model for final prediction. RESULTS: Univariable analysis indicated that treatment discontinuation was significantly associated with age, sex, combination therapy, baseline disease activity, and treatment response. Machine learning identified EASI/NRS and EASI-75/SLS-75 as key predictors of baseline disease activity and treatment response, respectively. The multivariable model confirmed independent predictive value for age, baseline EASI/NRS scores, and achievement of EASI-75/SLS-75. CONCLUSION: This study identified key determinants of dupilumab adherence and developed a predictive adherence model that offers personalized risk visualization via nomograms, providing an evidence-based tool for the precision management of AD biologic therapies.