A new protein panel for the diagnosis of HFpEF: combining machine-learning and liquid-chromatography mass-spectrometry proteomics.

INTRODUCTION: The diagnosis of heart failure with a preserved left ventricular ejection fraction (HFpEF) remains challenging. Plasma concentrations of NT-proBNP can rule out HFpEF with good sensitivity, but only moderate specificity. Therefore, there is a high unmet need for better diagnostic biomarkers with improved specificity. METHODS: LC-MS plasma proteomics was performed in 249 patients with symptomatic HFpEF (LVEF ≥50%) from BIOSTAT-CHF and 99 control subjects without heart failure from the PREVEND study. Machine learning was used to identify and rank proteins based on differential abundance. Biomarker candidates were technically validated using the SomaScan 7K assay. RESULTS: HFpEF patients had a median age of 77 years, a median LVEF of 59%, a median plasma NT-proBNP concentration of 1056 pg/mL and 44% were female. Non-HF controls had a median age of 75 years, a median plasma NT-proBNP concentration of 77 pg/mL and 44% were female. Principal component analysis showed a clear distinction between the protein profiles of patients with HFpEF versus controls. Overall, 211 differentially abundant proteins were observed. We identified two strongly downregulated proteins (tropomyosin-4 (TMP4) and 14-3-3ζ protein (YWHAZ)) as biomarker candidates. Compared to NT-proBNP, specificity increased from 73% to 91% (97% sensitivity) using a combined biomarker model, with a Net Reclassification Index of 1.57 (95% CI [1.26-1.83]). The AUC improved from 0.96 to 0.99 (95% CI [-0.00072; 0.065], p=0.036). CONCLUSION: We identified a protein combination with higher specificity than NT-proBNP for diagnosing HFpEF. Integrating these biomarkers into clinical practice could improve HFpEF diagnosis and management, although external validation is required before clinical implementation.