Inflammatory-neurological deficit versus metabolic dysregulation: unsupervised clustering and SHAP analysis.

OBJECTIVE

This retrospective cohort study enrolled 20,538 middle-aged and older patients with acute ischemic stroke (AIS) to identify immune-metabolic clinical subtypes by unsupervised clustering, to examine the differential association between subtype characteristics and post-stroke epilepsy (PSE) susceptibility, and to clarify immune-related threshold biomarkers for individualized PSE risk stratification.

METHODS

Seven standardized variables were used for K-means clustering: age; National Institutes of Health Stroke Scale (NIHSS) score; lipid profile components (triglycerides, low-density lipoprotein cholesterol [LDL-c], high-density lipoprotein cholesterol [HDL-c]); glycated hemoglobin (HbA1c); and the immune-inflammatory marker C-reactive protein (CRP). Two immune-metabolic subtypes were determined. Multivariable logistic regression was performed to quantify the association between Cluster and PSE. An Extra Trees-based SHapley Additive exPlanations (SHAP) framework was used to interpret the hierarchical contribution of immune and metabolic indicators to PSE risk within each subtype.

RESULTS

Clustering produced two clinically distinct subtypes. Cluster 1, the inflammatory-neurological deficit cluster (n = 7,790), had older age, higher NIHSS scores, higher CRP, and higher HDL-c. Cluster 2, the metabolic dysregulation cluster (n = 12,748), had higher HbA1c, LDL-c, and triglycerides (TG). The incidence of PSE was substantially higher in the inflammatory-neurological deficit cluster than in the metabolic dysregulation cluster (7.2% vs. 2.4%, P < 0.001). After full adjustment, membership in the inflammatory-neurological deficit cluster was associated with 4.31-fold higher odds of PSE. Extra Trees-based SHAP analysis identified distinct drivers in each cluster. In the inflammatory-neurological deficit cluster, CRP emerged as the leading predictor with a risk-acceleration threshold above 24 mg/L. In the metabolic dysregulation cluster, NIHSS was the top predictor, and HbA1c values below 7% were associated with a protective effect.

CONCLUSION

Unsupervised clustering separated AIS patients into two subtypes. The inflammatory-neurological deficit cluster had higher PSE risk. Cluster-specific SHAP results suggested candidate thresholds for individualized risk assessment.