Large-Scale Plasma Proteomics Profiles for Predicting Ischemic Stroke Risk in the General Population.

Abstract

Background: We aimed to develop and validate a protein risk score for ischemic stroke (IS) risk prediction and to compare its predictive capability with IS clinical risk factors and IS polygenic risk score.

Methods: The prospective cohort study included 53 029 participants from UKB-PPP (UK Biobank Pharmaceutical Proteomics Project). IS protein risk score was calculated as the weighted sum of proteins selected by the least absolute shrinkage and selection operator regression. The discrimination ability of models was assessed by C statistic. IS risk factors included age, sex, smoking, waist-to-hip ratio, antihypertensive medication use, systolic and diastolic blood pressure, coronary heart disease, diabetes, total cholesterol/high-density lipoprotein cholesterol ratio, and estimated glomerular filtration rate. Polygenic risk score was computed using identified susceptibility variants.

Results: After exclusions, 38 060 participants from England were randomly divided into the training set and the internal validation set in a 7:3 ratio, and 4970 participants from Scotland/Wales were assigned as the external validation set. Of 43 030 participants included (mean age, 59.0 years; 54.0% female), 989 incident IS occurred during a median follow-up of 13.6 years. In the training set, IS protein risk score was constructed using 17 out of 2911 proteins. In the internal validation set, compared with the basic model (age and sex: C statistic,0.720 [95% CI, 0.691-0.749]), IS protein risk score had the highest predictive performance for IS risk (C statistic, 0.765 [95% CI, 0.736-0.793]), followed by clinical risk factors of IS (C statistic, 0.753 [95% CI, 0.725-0.781]), and IS polygenic risk score (C statistic, 0.730 [95% CI, 0.701-0.759]). The top 5 proteins with the largest absolute coefficients in the IS protein risk score, including GDF15 (growth/differentiation factor 15), PLAUR (urokinase plasminogen activator surface receptor), NT-proBNP (N-terminal pro-B-type natriuretic peptide), IGFBP4 (insulin-like growth factor-binding protein 4), and BCAN (brevican core protein), contributed most of the predictive ability of the IS protein risk score, with a cumulative C statistic of 0.761 (95% CI, 0.733-0.790). These results were verified in the external validation cohort.

Conclusions: A simple model, including age, sex, and the IS protein risk score (or only the top 5 proteins) had a good predictive performance for IS risk.