Background: Sex specific differences in the proteomic profile of patients with heart failure with preserved ejection fraction (HFpEF) have been described. Recent data showed that the complex interplay of proteins is often insufficiently pictured by conventional statistical approaches. Therefore, we aimed to investigate sex specific differences in circulating proteins in HFpEF patients compared to patients without heart failure (No-HF) through a machine learning (ML) approach.

Methods: Patients of the LIFE-Heart study cohort with HFpEF and No-HF were included in the present study. All patients underwent an extensive proteomics analysis using two high-throughput panels (OLINK Cardiovascular II and Immuno-Oncology panel) comprehending 169 proteins. In a first step an eXtreme gradient boosting (XGboost) model containing all available proteins was used to identify proteins most closely associated to HFpEF in men and women, thereby establishing a female and male HFpEF prediction model. In a second step the two models were consecutively reduced to include only the minimum number of proteins required to maintain an AUC as comparable as possible (i.e. <3% deviation) to the initial model containing all proteins.

Results: Overall, 1810 patients were included, of those 996 had HFpEF (66.3±9 years, ♀ 41%, ♂ 59%) and 814 (63.8±8 years, ♀ 41%, ♂ 59%) No-HF. After the stepwise reduction, 15 proteins were retained in the male HFpEF prediction model (initial AUC 0.82, AUC after reduction 0.79) and 18 proteins in the female HFpEF prediction model (initial AUC 0.77, AUC after reduction 0.76) (Figure 1A). Of those proteins 5 (Adrenomedullin, Angiopoetin-2, Renin, Sortilin-1 and Stem cell factor) were included in both the female and male HFpEF prediction model. Those 5 proteins accounted for 49% and 36% of the model’s importance among male and female patients, respectively. In males, Angiopoetin-2 was the most influential protein (21%) and Adrenomedullin (13%) the most important one in females. On linear regression analysis, those two proteins showed the highest association with HFpEF typic features (Figure 1B) like a higher HFA-PEFF score, increased filling pressures, inflammation and obesity. Stem cell factor on the other hand was the only protein of the 5 shared proteins with effects showing inverse association with HFpEF features, which was more pronounced among female patients. Among sex specific proteins associations with either pro-inflammatory or hemodynamic variables were observed for respective sexes, but association with HFpEF features was more pronounced among the shared proteins.

Conclusion: ML identifies sex specific features of HFpEF patients on a circulating proteome level. While only five out of 15 and 18 were shared among sex, those proteins more holistically reflected HFpEF typic features than sex specific proteins. However, although important, those five proteins explained less than 50% of the model, highlighting the importance of sex specific proteins. While those proteins might only show partial associations to HFpEF specific features, they might be important to understand pathophysiological differences between males and females. Accounting for these sex specific differences might possibly pave the way for novel scientific approaches and foster our understanding of the complex HFpEF syndrome.