Nutrient‐, Food‐Group‐, and Food‐Item‐Level Dietary Signatures of Psoriasis Comorbidity Among Adults With Anxiety Disorder: An Explainable Machine Learning Study

ABSTRACT

Psoriasis and anxiety disorders frequently co‐occur, yet the dietary features associated with this comorbidity remain poorly characterized. This cross‐sectional study compared nutrient‐, food‐group‐, and food‐item‐level dietary representations among adults with medication‐defined anxiety disorder using linked National Health and Nutrition Examination Survey (NHANES), Food Patterns Equivalents Database/MyPyramid Equivalents Database (FPED/MPED), and What We Eat in America (WWEIA) data. Three analytic datasets were constructed at the nutrient level ( n  = 1016; 37 psoriasis‐comorbidity cases), food‐group level ( n  = 1215; 46 cases), and food‐item level ( n  = 1257; 46 cases). Each dataset was divided using a fixed stratified 7:3 training‐validation split. Correlation filtering, shadow‐feature selection, scaling, class‐imbalance handling, and hyperparameter tuning were confined to the training workflow, with five‐fold stratified cross‐validation and an untouched validation set. Seven models, including conventional multivariable logistic regression, were compared. In locked validation, the nutrient‐level logistic regression model showed the strongest validation signal (area under the receiver operating characteristic curve [AUC‐ROC] 0.645, 95% confidence interval [CI] 0.464–0.811; area under the precision‐recall curve [AUC‐PR] 0.063; Brier score 0.176). The food‐group logistic regression model (AUC‐ROC 0.396, 95% CI 0.256–0.539) and the food‐item extreme gradient boosting (XGBoost) model (AUC‐ROC 0.401, 95% CI 0.273–0.537) provided more limited validation signals in the present sample. SHapley Additive exPlanations (SHAP) analyses of the cross‐validation‐selected models identified niacin, carbohydrate intake, β‐cryptoxanthin, fat‐related variables, hypertension, body mass index, and selected food categories as influential model features. These contributions are best interpreted as model‐derived signals that prioritize candidate dietary and clinical features for further study. Together, the findings highlight nutrient‐level dietary information as the most promising representation for subsequent development of nutrition‐informed risk assessment in adults with anxiety. Larger prospective cohorts and external validation may further strengthen the clinical applicability of these signals.