Blood RNA Signatures Enable Accurate Discrimination of Stroke Subtype and Onset Time at Hospital Admission.

Despite advances in thrombolytic therapy (IVT) and mechanical thrombectomy (MT) for acute ischemic stroke, their benefit is strongly time-dependent and contingent on rapid exclusion of intracranial hemorrhage (ICH), where thrombolysis can be fatal. Here we evaluated whether a peripheral blood transcriptome-based machine learning model could rapidly identify hemorrhagic stroke, distinguish ischemic stroke from mimics, and predict eligibility for thrombolysis. Whole-blood samples (n=314) were collected from acute stroke patients admitted at emergency department of Grady Memorial Hospital (Atlanta, GA). Two independent training (n=192) and validation cohorts (n=122) were sequenced, aligned to GRCh38, and quantified with StringTie2. Differentially expressed transcripts were used to train hierarchical machine-learning models (caret, hidden Markov models [HMMs]) to classify hemorrhagic stroke, then distinguish ischemic stroke from stroke mimics, and further predict thrombolysis eligibility (≤3.5 hours from onset) and stroke severity (NIHSS). HMM-based classifiers demonstrated robust performance: a three-transcript panel perfectly discriminated hemorrhagic from non-hemorrhagic stroke (100% accuracy), and a four-transcript panel achieved 97% accuracy with 100% sensitivity and 96% specificity in validation. Ischemic stroke panels accurately distinguished patients from stroke mimics, time-associated transcripts identified individuals within the thrombolysis window, and severity-associated RNA signatures strongly correlated with NIHSS scores. The findings indicate that RNA profiling at admission can rapidly identify stroke subtypes, time of stroke onset, and stroke severity, supporting point-of-care triage and timely thrombolytic therapy.