From Nutritional Epidemiology to Clinical Prediction: Interpretable Machine Learning Reveals Subtle Associations Between Dietary Antioxidant Profiles and Non-Melanoma Skin Cancer in Older Adults.

The growing burden of non-melanoma skin cancer (NMSC) among older adults underscores the need for early risk prediction based on modifiable dietary factors. In this cross-sectional study using data from the 2007-2010 and 2017-2018 cycles of the National Health and Nutrition Examination Survey (NHANES), we evaluated dietary antioxidant profiles and developed an interpretable machine learning model for NMSC risk prediction. Baseline feature analysis revealed that participants with NMSC generally had higher socioeconomic status and higher oxidative balance scores (OBS), with significant differences also observed in age and race. Through feature importance ranking using the Random Forest (RF) algorithm, caffeine and luteolin were identified as the most predictive dietary variables. After comparing 12 machine learning algorithms, linear discriminant analysis (LDA) was selected as the final screening-oriented prediction model (AUC = 0.742). Although its overall accuracy was lower than that of several ensemble models, LDA demonstrated higher sensitivity (0.730) and favorable imbalance-aware performance, thereby mitigating the "accuracy paradox" and supporting its use for preliminary screening to identify participants with higher predicted NMSC risk. SHapley Additive exPlanations (SHAP) global explanations further confirmed that, after controlling for age as a primary risk factor, niacin and quercetin act as protective factors in the NMSC risk prediction model. In summary, this study developed an interpretable and robust machine learning model for NMSC risk prediction in older adults and provides hypothesis-generating evidence regarding the potential role of dietary antioxidant profiles in prevention-oriented strategies.