Interpretable machine learning reveals the association of metabolic, inflammatory, and nutritional indices with ischemic stroke complicated by metabolic abnormality clustering

Background This study aimed to explore correlations between metabolic, inflammatory and nutritional composite indices and metabolic abnormality clustering (MAC) in acute ischemic stroke (AIS), and build an interpretable Clinlabomics model to support early risk screening and stratified clinical management. Methods This study retrospectively recruited 2,183 AIS patients from Suining Central Hospital between January 2023 and July 2025, who were divided into MAC ( n = 621) and non-MAC ( n = 1,562) groups. We included a panel of glucose and lipid metabolic parameters as well as their composite metabolic indices. The cohort was randomly split into a 7:3 training set ( n = 1,529) and internal test set ( n = 654), with an independent temporal validation set of 496 patients. In the training cohort, feature selection was performed via least absolute shrinkage and selection operator (LASSO) regression. Eleven machine learning algorithms were used to develop Clinlabomics models. The optimal algorithm was selected based on sensitivity, specificity, accuracy, F1-score, and AUC. Interpretability analysis was performed on the optimal Clinlabomics model using feature importance ranking and SHapley Additive exPlanations (SHAP) to identify core features. Subsequently, univariate and multivariate logistic regression analyses were conducted to identify metabolic signatures significantly associated with the comorbidity of MAC and AIS. Results A total of 22 feature variables were identified via LASSO analysis. Eleven machine learning algorithms were adopted to develop Clinlabomics models in the training cohort, among which the recursive partitioning and regression trees (rpart) algorithm was identified as the optimal machine learning algorithm. After hyperparameter tuning, the optimal complexity parameter (cp.) was 0.00015. Following tuning, the AUC values in the training, test, and temporal validation cohorts were as follows: 0.979 (95% CI: 0.970–0.988), 0.968 (95%CI: 0.951–0.985), and temporal validation cohort 0.974 (95% CI: 0.958–0.990). In the temporal validation cohort, this optimal Clinlabomics model correctly identified 129 (84.3%) MAC patients and 337 (98.3%) non-MAC patients. Finally, through integrated analyses, triglyceride–glucose index (TyG), Castelli risk index II (CRI-II), fasting blood glucose (FBG) and atherogenic index of plasma (AIP) were recognized as key metabolic indicators for MAC. Conclusions Based on the rpart algorithm, this study developed the optimal clinlabomics model for MAC-AIS comorbidity with satisfactory discriminative ability. TyG, CRI-II, FBG, and AIP were recognized as key metabolic signatures of MAC-AIS comorbidity, providing a reliable tool for early clinical intervention and personalized management.