Plasma proteomic signatures achieved AUCs of 0.88 for large artery atherosclerosis and 0.89 for cardioembolism to predict etiologies in cryptogenic acute ischemic stroke.
Do plasma proteomic signatures accurately classify acute ischemic stroke etiologies and predict the etiology of cryptogenic strokes?
A plasma proteomic signature of three proteins combined with clinical factors accurately classified acute ischemic stroke etiologies and predicted etiologies in cryptogenic strokes.
BACKGROUND: Identifying acute ischemic stroke (AIS) etiology guides targeted therapy implementatoin to prevent recurrent stroke. We derive plasma protein signatures of non-cryptogenic AIS etiologies and apply them to cryptogenic strokes to predict etiologies. METHODS: We studied adults at Yale-New Haven Hospital with an AIS from 2015-2020 and stored plasma samples. Proteins were measured with a SomaScan 11K Assay. Etiology was adjudicated by > 2 board-certified vascular neurologists. ANOVA tests identified proteins significantly different among the 4 non-cryptogenic etiologies (large artery atherosclerosis (LAA), cardioembolism (CE), small vessel disease (SVD), and other rare, determined etiologies (ODE)). Proteins with fold change > 1.2 and p-value < 0.05 were selected without multiple testing adjustment in this exploration. We built logistic regression models to classify 4-level and binary etiologies versus not with: A) age, B) age, sex, C) age, sex, hypertension, D) proteins, and E) age, sex, hypertension, proteins selected by stepwise selection for each outcome. We computed 95% confidence intervals for binary models with 2,000 bootstrap replicates. The PheWeb 2019 database was used to link predictive proteins with phenotypes. We applied the 4-level model to cryptogenic strokes to predict non-cryptogenic etiologies. RESULTS: We included 64 patients (median age 69 years IQR 58-76, 42.2% female, last known well to sample collection time: median 27 hours IQR 22-68, LAA n=15; CE n=23; SVD n=6; ODE n=7; cryptogenic n=13). Of 11,083 proteins, there were 40 differentially expressed proteins among 4 etiologies ( Figure 1 ). Three proteins (lithostathine-1-beta, transcription factor SOX-21, creatine kinase M- and B-types) classified 4-level etiologies: areas under the curve (AUC) Model A: 0.69, B: 0.70, C: 0.76, D: 0.84, E: 0.88; Table ). AUCs for each etiology were: 0.88 LAA (95% CI 0.78-0.97), 0.89 CE (0.80-0.98), 0.98 SVD (0.94-1.0), and 0.97 ODE (0.94-1.00). Identified proteins are linked with malignancy, varicella zoster, inflammation, hematologic conditions, and atrial fibrillation ( Figure 2) . In the cryptogenic stroke cohort, 7 patients had highest predicted probabilities for LAA (range: 0.55-0.83) and 6 for CE (0.47-0.85). CONCLUSION: We derived plasma proteomic signatures of non-cryptogenic etiologies with biologic plausibility and applied them to predict etiologies in cryptogenic AIS. Further studies are needed to evaluate their generalizability.
Lee et al. (2026) studied this question. Plasma proteomic signatures achieved AUCs of 0.88 for large artery atherosclerosis and 0.89 for cardioembolism to predict etiologies in cryptogenic acute ischemic stroke.