Abstract DP045: Machine Learning Analysis of Serum Immune Markers Identifies a Panel Associated with Incident Stroke in the Northern Manhattan Study

Background: Inflammation contributes to cerebrovascular disease and is linked to stroke risk. Serum immune markers may capture these pathways and improve risk stratification beyond traditional factors. We used machine learning to identify an immune panel, and tested its predictive value in the multiethnic, population-based Northern Manhattan Study (NOMAS). Methods: Stroke-free NOMAS participants aged ≥55 years with serum levels of 60 immune markers measured by a customized Luminex assay were studied. We applied the least absolute shrinkage and selection operator (LASSO) Cox model, a penalized regression form of machine learning, with 5-fold cross-validation to log-transformed, standardized markers and all immune marker–marker interactions, using time-to-incident stroke as the outcome. Stability was assessed with event-stratified bootstrap resampling (B=300), retaining terms in ≥80% of resamples. Nested Cox models evaluated incremental predictive value of: (a) any single LASSO-selected marker, and (b) selected marker(s) plus stable interactions over traditional stroke predictors (Models 1–3: age only; +sociodemographics; +vascular risk factors: Table 2), using likelihood ratio tests (LRT) and, Harrell’s C-index. We assessed improvement in risk categorization via continuous net reclassification improvement (NRI). Results: Among 1,176 participants (mean age 70±9 years; 60% female; 68% Hispanic), 131 strokes occurred over 14 (IQR 6.5) years. LASSO selected one marker, CXCL9, and five statistical interaction terms, i.e., marker pairs whose effects varied with each other’s levels: IL13 with βNGF; CXCL10 with CCL2; IL13 with TRAIL; CXCL10 with IL1α; and IL1α with IL21. Adding CXCL9 alone improved model discrimination (ΔC-index≈+0.011, LRT p