Machine learning reveals metabolic and inflammatory predictors of exercise adaptation in HFpEF

Abstract Background Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous syndrome with exercise intolerance (low peak oxygen uptake; VO2 peak) as the cardinal symptom. Exercise training can improve VO2 peak in HFpEF, but individual responses vary widely. The multicenter Ex-DHF trial showed that combined endurance/resistance training improved VO2 peak over 3 to 12 months compared to usual care. Purpose Using machine learning, we aimed to identify individual characteristics and plasma biomarkers predictors of VO2 peak response in HFpEF patients. Methods We analysed Ex-DHF trial data (N=322 HFpEF patients) at baseline and after 3- and 12-months intervention. Baseline plasma proteomic profiling (Olink cardiovascular panel II and ELISA) measured 97 biomarkers. Predictive models for percent change in weight-normalized VO2 peak at 3 and 12 months were built using multiple algorithms (linear regression, Lasso, random forest, gradient boosting, XGBoost). SHAP values were computed for explainability. Missing values were handled via multiple imputation (Bayesian ridge regression), and feature importance was aggregated across imputations and models. Results Across models, predictive performance was modest and comparable (3-month RMSE mean = 0.34, SD = 0.03; 12-month RMSE mean = 0.27, SD = 0.02). Some features consistently predicted changes in VO2 peak/kg (Figs. 1 and 2): baseline vitality and physical limitation, exercise group assignment and adherence. Short-term biomarkers associated with VO2 peak responses were RAGE (receptor for advanced glycation end-products), adiponectin, gastric inhibitory polypeptide (GIP), Brother of CDO (BOC), Integrin Subunit Beta 1 Binding Protein 2 (ITGB1BP2), proline/arginine-rich end leucine-rich repeat protein (PRELP), and pentraxin-3 (PTX3). Long-term response predictors included fibroblast growth factor-21 (FGF-21), ADAMTS13 (a metalloprotease), renin, interleukin-18 (IL-18), B-type natriuretic peptide (BNP), and RAGE. Age was associated with short-term changes, while goal self-concordance (motivation) was a long-term predictor. These biomarkers reflect potential mechanisms: ADAMTS13 suggests thrombo-inflammation via neutrophil extracellular traps (NETs) and von Willebrand factor (vWF), both implicated in HF. RAGE and its ligands may promote NET formation and vascular inflammation, while PTX3 can bind NETs. BOC has been linked to venous thromboembolism, and PRELP is associated with heart fibrosis. Conclusions Our analysis highlights pathways linked to exercise adaptation in HFpEF patients. The identified biomarkers, many involved in inflammation, oxidative stress, and metabolism, suggest these processes influence heterogeneity in training benefit. Integrating proteomic profiling with clinical factors may help personalize HFpEF exercise programs. Future studies should validate these candidate predictors and assess whether targeting these pathways enhances exercise responsiveness.Top features (top: 3 m, bottom: 12 m)SHAP values (top: 3 m, bottom: 12 m)