Machine Learning Applied to Proteomic, Metabolomic, and Multi-Omics Biomarkers for the Diagnosis and Risk Stratification of Heart Failure With Preserved Ejection Fraction (HFpEF): A Systematic Review

Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous and increasingly prevalent syndrome that remains challenging to diagnose and risk-stratify using conventional clinical and echocardiographic parameters. Advances in high-throughput proteomic and metabolomic technologies, combined with machine learning methods, have enabled the development of predictive models that capture complex molecular signatures associated with heart failure. This systematic review synthesizes current evidence on machine learning models derived from proteomic, metabolomic, and multi-omics datasets for the diagnosis, early detection, and prognostic assessment of heart failure, with particular focus on HFpEF. A structured search of PubMed, Scopus, and Web of Science identified eight eligible studies published between 2010 and 2026. Included studies applied machine learning techniques to high-dimensional molecular data to predict incident HF, classify HFpEF, identify molecular subtypes, or estimate mortality risk. Several models demonstrated strong discriminatory performance, with reported area under the curve (AUC) or C-index values generally ranging from approximately 0.78 to 0.98, and in some studies, demonstrated improved performance compared with established clinical tools such as natriuretic peptides (e.g., NT-proBNP) and conventional risk scores, including the Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC) score. Multi-omics integration showed particular promise in identifying individuals at risk of developing HFpEF years before symptom onset. However, substantial heterogeneity across molecular platforms, limited external validation in some studies, and vulnerability to overfitting in smaller datasets restrict generalizability. Methodological quality assessment using the Prediction model Risk Of Bias ASsessment tool (PROBAST) tool indicated variable risk of bias, with higher concerns observed in smaller, non-externally validated studies. No randomized trials have yet evaluated the clinical impact of ML-omics-guided risk stratification. Overall, machine learning-based molecular profiling represents a promising direction for refining HFpEF phenotyping and risk prediction, but standardization, cross-platform validation, and outcome-based testing are necessary before routine clinical implementation.