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

Several biochemical features exhibited statistically significant heterogeneity of variance that can distort standard ANOVA inference.

• 49 biochemical features were analyzed in a cohort of 120 patients.
• Features including Fe, Transf, RDW%, and LDL showed statistically significant heterogeneity of variance (p < 0.05).
• The study applied variance-aware statistical testing to detect this heterogeneity.
• The authors note this variance heterogeneity 'is known to distort standard ANOVA inference.'

A gradient boosting model with restricted tree depth achieved high discriminative accuracy across all five clinical classes.

• The model used a maximum tree depth of 3.
• F1-scores ranged between 0.87 and 0.96 across all five clinical classes.
• The five clinical classes included AIS, CCCI, type 2 diabetes mellitus, SIVD, and a control/fifth group.
• Standard machine-learning classifiers demonstrated variable performance across clinical groups, in contrast to the gradient boosting model.

SHAP analysis identified iron (Fe), transferrin, and glucose as key stable biomarkers with synergistic interactions in model predictions.

• Shapley Additive Explanations (SHAP) were used to interpret the gradient boosting model.
• These three biomarkers were described as having 'synergistic interactions' in model predictions.
• The biomarkers were characterized as 'key stable biomarkers' based on their SHAP values.
• SHAP analysis was applied to a model trained on 49 biochemical features from 120 patients.

Comparative analysis indicated consistency between statistical significance and SHAP-based feature importance ranks.

• Spearman correlation coefficients between statistical significance ranks and SHAP values reached 0.53 for groups 1–2.
• Spearman correlation coefficients reached 0.59 for groups 1–5.
• The analysis compared feature importance ranks from statistical testing with those derived from SHAP values.

Unsupervised KMeans clustering revealed poor correspondence with clinical labels.

• KMeans clustering was performed with k = 5 clusters.
• The Adjusted Rand Index (ARI) was 0.198, indicating poor alignment with clinical labels.
• The Normalized Mutual Information (NMI) was 0.286.
• These results indicate that 'statistical structures in biochemical data do not always map to meaningful clinical categories.'

The study cohort comprised patients with four age-related neurological and metabolic conditions.

• The four conditions studied were acute ischemic stroke (AIS), chronic cerebral circulation insufficiency (CCCI), type 2 diabetes mellitus (DM), and subcortical ischemic vascular dementia (SIVD).
• The total cohort consisted of 120 patients.
• 49 biochemical features were measured per patient.
• Five clinical classes total were used in the classification framework.