An Anomaly-Aware Analysis of Metabolic Heterogeneity in Diabetes Risk During and After COVID-19

Diabetes risk differs substantially among individuals, and some of this variation may be missed when atypical observations are removed during data pre-processing.We examined whether these observations provide useful information for characterising diabetes-related risks before, during, and after the COVID-19 pandemic.The analysis included data from the Korea National Health and Nutrition Examination Survey collected between 2015 and 2024.Extreme observations were retained and represented by continuous anomaly scores and binary anomaly flags.We evaluated these variables, together with conventional predictors, using machine learning models and block-wise ablation.Participants classified as anomalies differed from the remaining population in terms of several metabolic and behavioural characteristics.Glycaemic biomarkers accounted for most of the discrimination in diabetes classification, whereas demographic, behavioural, and anomaly related variables contributed more to risk prediction.Cross-temporal analyses also showed modest changes in feature importance after the COVID-19 period.The model performance declined when the anomaly related variables were removed.Thus, retaining information from atypical metabolic profiles may help characterise the heterogeneity in diabetes risk across changing population conditions.