Machine learning–based proteomic phenogrouping identifies an inflammatory high-risk subtype in HFpEF : results from the PACIFIC-PRESERVED study

Abstract Background Heart failure with preserved ejection fraction (HFpEF) represents more than half of heart failure cases and remains a major unmet clinical challenge due to marked biological and clinical heterogeneity. Conventional classifications fail to identify mechanistically distinct subgroups that may explain differential prognosis and therapeutic response. We aimed to define biologically relevant HFpEF phenotypes using deep multimodal profiling combined with machine learning, and to evaluate their prognostic implications. Methods PACIFIC-PRESERVED is a prospective, deeply phenotyped cohort enrolling HFpEF, HFrEF, and control participants in a 3:2:1 ratio. Among 155 individuals with complete multimodal data (clinical, imaging, biological, and high-throughput proteomics [~5,400 proteins]), multiview spectral clustering within a meta-consensus framework was applied to identify phenogroups. Cluster robustness was assessed by stability metrics and recursive feature elimination. Proteomic signatures were characterized using differential expression and pathway enrichment analyses. A supervised XGBoost classifier was trained and externally validated in the independent MEDIA-DHF cohort (n=456 HFpEF). Results Among the tested models, a four-cluster solution emerged as the most stable and biologically informative configuration, distinguishing two HFpEF subgroups (PEF1: 33%; PEF2: 55%), in addition to HFrEF and controls. While PEF1 and PEF2 showed broadly overlapping clinical characteristics and were not separable by conventional diagnostics, high-throughput proteomic profiling revealed marked molecular divergence. PEF1 was enriched in inflammatory pathways, particularly TNF receptor signaling, and in markers of cardio-skeletal muscle injury, and was associated with a higher prevalence of renal dysfunction. In contrast, PEF2 displayed signatures of endothelial stress and a relatively preserved clinical profile. Sensitivity clustering analyses excluding the proteomic dataset resulted in a marked decline in clustering performance, notably impairing discrimination between PEF1 and PEF2 clusters, underscoring the central contribution of proteomic data to subgroup identification. The multimodal classifier integrating clinical, imaging, and proteomic variables achieved high predictive performance in the validation cohort MEDIA-DHF (F1-score 0.87) and successfully replicated the phenotypic structure. Importantly, PEF1 was independently associated with adverse outcomes (death or HF hospitalization), with an adjusted hazard ratio of 2.27 (95% CI 1.04–4.94; p=0.039). Conclusions Integrative proteomics combined with machine learning revealed two biologically and prognostically distinct subtypes within HFpEF, a syndrome that appears clinically unified yet is molecularly heterogeneous. These findings establish a classifier-driven, biologically informed stratification of HFpEF, and provide a foundation for precision medicine in heart failure.