ABSTRACT
Background: Coronary microvascular dysfunction (CMD) is a major cause of ischemia with no obstructed coronary arteries. Objectives: The authors sought to assess protein biomarker signature for CMD. Methods: We quantified 184 unique cardiovascular proteins with proximity extension assay in 1,471 women with angina and no obstructive coronary artery disease characterized for CMD by coronary flow velocity reserve (CFVR) by transthoracic echo Doppler. We performed Pearson's correlations of CFVR and each of the 184 biomarkers, and principal component analyses and weighted correlation network analysis to identify clusters linked to CMD. For prediction of CMD (CFVR < 2.25), we applied logistic regression and machine learning algorithms (least absolute shrinkage and selection operator, random forest, extreme gradient boosting, and adaptive boosting) in discovery and validation cohorts. Results: Sixty-one biomarkers were correlated with CFVR with strongest correlations for renin (REN), growth differentiation factor 15, brain natriuretic protein (BNP), N-terminal-proBNP (NT-proBNP), and adrenomedullin (ADM) (all P < 1e-06). Two principal components with highest loading on BNP/NTproBNP and interleukin 6, respectively, were strongly associated with low CFVR. Weighted correlation network analysis identified 2 clusters associated with low CFVR reflecting involvement of hypertension/vascular function and immune modulation. The best prediction model for CFVR <2.25 using clinical data had area under the receiver operating characteristic curve (ROC-AUC) of 0.61 (95% CI: 0.56-0.66). ROC-AUC was 0.66 (95% CI: 0.62-0.71) with addition of biomarkers (P for model improvement = 0.01). Stringent two-layer cross-validated machine learning models had ROC-AUC ranging from 0.58 to 0.66; the most predictive biomarkers were REN, BNP, NT-proBNP, growth differentiation factor 15, and ADM. Conclusions: CMD was associated with pathways particularly involving inflammation (interleukin 6), blood pressure (REN, ADM), and ventricular remodeling (BNP/NT-proBNP) independently of clinical risk factors. Model prediction improved with biomarkers, but prediction remained moderate.
