Interpretable Machine Learning with SHAP Identifies Key Biomarkers in a Multi-Factorial Spectrum of Age-Related Neurological and Metabolic Conditions

Vascular and metabolic disorders in the elderly—including acute ischemic stroke (AIS), chronic cerebral circulation insufficiency (CCCI), type 2 diabetes mellitus (DM), and subcortical ischemic vascular dementia (SIVD)—pose a major diagnostic challenge due to their reliance on multi-parameter blood chemistry. In this study, 49 biochemical features were analyzed within a cohort of 120 patients. The application of variance-aware statistical testing revealed that several features (e.g., Fe, Transf, RDW%, LDL) exhibited statistically significant heterogeneity of variance (p < 0.05), which is known to distort standard ANOVA inference. While standard machine-learning (ML) classifiers demonstrated variable performance across clinical groups, a gradient boosting model with restricted tree depth (max depth = 3) achieved high discriminative accuracy, yielding F1-scores between 0.87 and 0.96 across all five clinical classes. Through the use of Shapley Additive Explanations (SHAP), key stable biomarkers including iron (Fe), transferrin, and glucose were identified as having synergistic interactions in model predictions. A comparative analysis of feature importance ranks indicated consistency between statistical significance and SHAP values, with Spearman correlation coefficients reaching 0.53 for groups 1–2 and 0.59 for groups 1–5. Conversely, unsupervised KMeans clustering (k = 5) revealed a poor correspondence with clinical labels, yielding an Adjusted Rand Index (ARI) of 0.198 and Normalized Mutual Information (NMI) of 0.286. These results underscore that statistical structures in biochemical data do not always map to meaningful clinical categories and advocate for the adoption of variance-aware workflows and interpretable ML to enhance diagnostic reliability in aging populations.