ABSTRACT NUMBER: ESOC2026A838 DEVELOPMENT AND VALIDATION OF MACHINE-LEARNING ALGORITHMS INTEGRATING GLIAL FIBRILLARY ACIDIC PROTEIN AND CLINICAL PARAMETERS FOR PREHOSPITAL STROKE TRIAGE

Abstract Background and aims Early prehospital differentiation between intracerebral hemorrhage (ICH) and ischemic stroke is of major clinical importance. The DETECT study showed that prehospital GFAP point-of-care testing can identify ICH with positive predictive value (PPV) >90% using age-stratified cut-offs. However, sensitivity ranged from 56% to 72%, being lowest in elderly patients, who constitute most stroke cases. We aimed to develop unified machine-learning models integrating GFAP with clinical variables to achieve consistent ICH detection across all ages for prehospital decision support. Methods We analyzed 353 patients from the DETECT cohort presenting within 6 hours of symptom onset, stratified into 80% training and 20% test sets. A limited-data logistic regression model used five objective variables obtainable even in uncooperative or aphasic patients (age, systolic and diastolic blood pressure, GFAP, altered consciousness). A full-data XGBoost model incorporated additional examination findings and clinical history. Results For prehospital diagnosis of ICH, the limited-data model achieved ROC AUC 0.876, sensitivity 67%, specificity 98%, and PPV 91%. The full-data model achieved ROC AUC 0.865, sensitivity 73%, specificity 96%, and PPV 85%. The two models show complementary performance: the limited model allows confident ICH identification when a thorough examination is not feasible, while the full model maximizes case detection, matching the sensitivity of DETECT's best-performing age subgroup but consistent across all ages. Conclusions The two complementary algorithms provide reliable support for ICH detection across all ages, especially in elderly patients, who are most affected by stroke, enabling optimized triage and earlier prehospital initiation of blood pressure management. Conflict of interest Deepak Bos: nothing to disclose