Targeted proteomics for the prediction of sudden cardiac death in individuals with heart failure across the spectrum of left ventricular ejection fraction

Background: Sudden cardiac death (SCD) substantially contributes to heart failure (HF)-related mortality. While current prediction models have relevant shortcomings, specifically risk stratification by left ventricular ejection fraction (LVEF), the role of proteomics for SCD risk stratification in individuals with HF is unclear.

Methods: Data from the MyoVasc study (NCT04064450) on HF of ischemic and nonischemic etiology, classified according to the 2021 Universal Definition of HF, were analyzed. Participants underwent deep cardiovascular phenotyping, including echocardiography, resting and Holter electrocardiography (ECG). Plasma concentrations of 357 proteins were measured using a targeted immuno-qPCR-based proteomics assay (Olink Proteomics, Sweden). Elastic net-regularized Cox regression was used to identify a molecular signature for SCD. Metascape was employed for pathway enrichment analysis. Based on the individual importance of the proteins selected, a weighted proteomic score was computed and analyzed against various cardiovascular traits by linear and robust Poisson regression. The prognostic relevance of the proteomic score was investigated in Cox regression analyses under consideration of competing risk. Events were independently adjudicated by evaluation of death certificates and medical records.                                              

Results: The analysis sample comprised 2630 individuals with HF stage B to D (mean age 66.5 years, 34% female). Among 1678 individuals with symptomatic HF (stage C/D), 43% displayed reduced EF. Over a median follow up of 7.3 years, 52 fatal SCD events occurred. A proteomic signature of 50 proteins (top 5: COL18A1, FGF-23, SPON1, ADM, PTX3) was found to predict SCD with an area under the curve of 0.71 after 10-fold cross-validation and adjustment for age and sex. Pathway analysis revealed protein clusters involving extracellular matrix remodeling and fibrosis, cytokine signaling, and complement activation. The proteomic score for SCD was associated with LVEF (β-estimate per standard deviation [SD] increase of SCD score -0.28 [-0.32; -0.24], p<0.001) and LV mass/height^2.7 (β _SD 0.22 [0.18, 0.26], p<0.001), which was also evident adjusting for the clinical profile. In addition to a positive relationship with QRS duration (β _SD 0.19 [0.15; 0.23], p<0.001) and QT interval (β _SD 0.33 [0.29; 0.37,] p<0.001),  higher SCD scores translated into a greater likelihood for the presence of ventricular tachycardia on Holter ECG (Prevalence ratio [PR] _SD 1.44 [1.16; 1.79], p=0.003) and history of survived cardiac arrest (PR _SD 1.42 [1.18; 1.7], p=0.002). The proteomic score was the most powerful predictor of SCD in Cox regression analyses corrected for multiple testing among clinical variables including LVEF, risk factors, comorbidities, NT-proBNP,  and high sensitivity Troponin I (Hazard ratio _SD 3.35 [2.64; 4.26], p<0.001). Adding the score to a multivariable clinical model provided incremental prognostic value for SCD (C-index 0.87 with vs 0.75 without SCD score). 

Conclusions:  Among individuals with HF, a supervised machine learning approach using targeted proteomics identified a molecular signature of SCD, including potentially druggable targets. The weighted proteomic score derived from the SCD signature was related to cardiovascular traits and might substantially improve SCD risk prediction over established clinical risk factors. Currently, a replication analysis is conducted in the UK Biobank for external validation.