Abstract Body: Introduction. Confident cardioembolic (CE) etiology prediction in cryptogenic stroke may enable accurate patient selection for advanced diagnostic workup in these patients. Thrombus analysis by flow cytometry may serve as a complementary source of powerful biomarkers for this purpose.
Methods. A retrospective analysis was performed to study the association of a combination of biomarkers with high-risk CE etiology in cryptogenic stroke. High-risk CE etiology was determined upon extensive diagnostic workup. A total of 51 descriptors from serologic analysis, stroke severity, diagnostic cardiac imaging and thrombus flow cytometry were collected for each patient per-protocol. A random forest model with extreme gradient boosting (XGBRF) was implemented for high-risk CE prediction. Recursive feature elimination (RFE) analysis based on feature importance and validation performance of the XGBRF model was conducted to identify the most important predictors of CE cause. Model results were extracted using 50-fold Monte Carlo cross-validation, reserving 20% of cases for validation on each fold. All experiments, were conducted in two branches: 1) including all descriptors, and 2) excluding flow cytometry biomarkers. The area under the receiver operating characteristic (AUROC) curve and F1-score were used for predictive performance assessment.
Results. A sample of 89 cases (63, 70.8% cases in the high-risk CE group) with cryptogenic stroke treated with thrombectomy were included. RFE analysis identified three predictors for optimal CE etiology prediction when excluding cytometric markers (left atrial area, baseline NIHSS and NTproBNP), and five predictors when including all predictors (same three + monocyte/lymphocyte ratio and lymphocyte %). Validation AUROC showed superior discrimination ability when including cytometric biomarkers (AUROC: 0.84, 95%CI 0.81-0.87; F1: 0.83, 95%CI 0.80-0.85) that when excluded (AUROC 0.80, 95%CI 0.77-0.83; F1: 0.73, 95%CI 0.71-0.76).
Conclusions. Thrombus biomarkers by flow cytometry may add predictive value in multi-variable models for CE etiology prediction in cryptogenic stroke. Early prediction of CE etiology in cryptogenic stroke may guide diagnostic studies after acute stroke.