Assessing diabetes risk factors using explainable AI and SMOTEENN: a nutrition-based analysis in South Korea

Abstract

The coronavirus disease (COVID-19) pandemic has influenced diabetes risk and management at the population level, underscoring the need for interpretable and data-driven risk assessment approaches. We developed an explainable machine learning framework integrating the KNHANES and COVID-19 datasets using a condition-based feature alignment (CBFA) approach. Data preprocessing involved multicollinearity screening, factor analysis, Mahalanobis distance–based outlier detection, and hybrid resampling with SMOTEENN to address class imbalance. Nutritional intake patterns were summarized through weighted clustering, and model interpretability was evaluated using SHAP and LIME. The MASN_XGB model achieved an accuracy of 93.12% (95% CI: 92.93–93.39), an AUC of 95.72%, and an F1-score of 93.32%. Glycaemic biomarkers were the primary contributors to prediction, while nutritional and COVID-19–related factors provided complementary information for population-level risk stratification. Integrating structured preprocessing with explainable machine learning provides a transparent and scalable approach to diabetes risk assessment. This framework can support population-level screening and preventive strategies in evolving public health settings.