Using SHAP and LIME to Explain Machine Learning Models Predicting Comorbid Depression and Stroke From Daily Dietary Nutrient Intake in a US Population-Based Study

Abstract

While comorbid depression and stroke are a major concern for public health, the effect of dietary nutrient patterns on their concurrent occurrence is still largely unexplored. From NHANES, a survey of the U.S. civilian, non-institutionalized population, we included 814 participants with complete data on diet, depression, and stroke. Of these, 140 were identified with comorbid depression and stroke. Baseline characteristics were compared between groups, and Weighted Quantile Sum (WQS) regression was used to evaluate the collective effects of nutrient mixtures. Machine learning models aimed at predicting comorbid conditions were developed, incorporating Synthetic Minority Oversampling Technique (SMOTE) for oversampling and Boruta for selecting features. The interpretability of these models was analyzed using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Participants with comorbidities were younger and had lower socioeconomic status, along with reduced intake of thiamin, vitamin B6, total folate, added vitamin B12, and vitamin C. Although neither WQS-negative nor WQS-positive indices showed statistically significant associations with comorbidity risk, specific nutrients such as alcohol, alpha-carotene, added vitamin B12, theobromine, and vitamin E emerged as predominant contributors within the mixture models. The Random Forest classifier achieved the highest area under the receiver operating characteristic curve (AUC = 0.945) when adjusted for covariates and maintained consistently high performance in the unadjusted setting. SHAP and LIME analyses consistently identified vitamin B1, vitamin B12, zinc, vitamin C, and caffeine as influential predictors, with SHAP plots revealing mirrored feature contribution patterns depending on comorbidity status. Covariate adjustment improved directional stability and interpretability, particularly in SHAP dependence plots and waterfall visualizations. LIME explanations at the individual level corroborated these findings, showing consistent yet class-dependent feature effects. Although the overall mixture effect was not significant, machine learning identified nutrient-specific signals associated with comorbid depression and stroke. These results indicate that integrating dietary indicators with explainable artificial intelligence may improve transparency in risk prediction and guide future longitudinal and interventional research.