Machine-Learning Models of Cognitive Test Performance Incorporating Exposomic Pesticide Biomarkers in Older U.S. Adults.

Background/Objectives: Environmental exposures may contribute to heterogeneity in cognitive aging, yet population-scale datasets integrating exposure biomarkers with cognitive testing remain underused. We evaluated associations between cognitive test performance and demographic, lifestyle, psychosocial, and pesticide-exposure variables in older U.S. adults. Methods: Using NHANES 2011-2014 data, we analyzed adults aged ≥60 years with complete cognitive assessments, demographic/lifestyle covariates, Patient Health Questionnaire-9 scores, binge drinking status, and urinary concentrations of eight pesticide biomarkers. Least absolute shrinkage and selection operator (LASSO) and ridge regression models were trained to predict CERAD immediate learning composite scores, Animal Fluency, and Digit Symbol Substitution Test (DSST) performance. Models were evaluated using mean absolute error, mean squared error, root mean squared error, and R-squared. Results: The analytic sample included 429 participants. Sex, race/ethnicity, and educational attainment were among the strongest predictors across outcomes. LASSO showed the best overall performance in the full cohort, particularly for DSST (R-squared = 0.4680). DEET and desethyl hydroxy-DEET retained non-zero negative coefficients for DSST, and a reduced model including demographics and these two biomarkers achieved comparable performance (DSST R-squared = 0.475). Age-stratified analyses suggested stronger predictor importance in adults aged 60-69 years. Conclusions: These findings support incorporating exposomic biomarkers alongside sociodemographic factors when modeling cognitive test performance in older adults, while emphasizing that cross-sectional NHANES data cannot establish temporality or causality.