Prognosis and treatment response in cardiometabolic HFpEF: deep learning-based phenomapping with a pragmatic clinical classifier

Background and purpose Heart failure with preserved ejection fraction (HFpEF) is a highly heterogeneous syndrome. Although a cardiometabolic phenogroup is widely recognized, its clinical and prognostic implications remain unclear. We aimed to identify sub-phenotypes within cardiometabolic HFpEF, develop a simplified decision tool, compare their clinical characteristics, and evaluate their prognosis and treatment responses. Methods We enrolled 1,158 hospitalized patients diagnosed with incident cardiometabolic HFpEF between April 2014 and December 2024. An unsupervised DeepCluster with a fully connected neural network was used to identify cardiometabolic HFpEF phenogroups. We assessed feature importance using random forest Gini impurity score and SHapley Additive exPlanations (SHAP), then derived a classification tree (CART) based on the top 15 non-collinear features. Clinical characteristics were compared across identified phenogroups. All-cause mortality, heart failure rehospitalization, and treatment responses were evaluated using multivariable Cox proportional hazards models. Results Two distinct phenogroups among cardiometabolic HFpEF patients (Figure 1A) were identified based on 87 demographic and echocardiographic variables, clinical biomarkers, medical and medication history. Phenogroup 1 (n = 356) showed more severe symptoms and diastolic dysfunction, higher disease burden of diabetes and chronic kidney disease, and elevated levels of NT-proBNP, US-CRP, cTnT, uric acid, Urea, creatinine, glucose, and HbA1c (Figure 1B-D). Phenogroup 2 (n = 802) had higher prevalence of hyperlipidemia and more frequent usage of SGLT2 inhibitor, statin, cholesterol-absorption inhibitors, and metformin. For prognostic outcomes, phenogroup 1 had significantly higher all-cause mortality than phenogroup 2 (HR=5.19, 95%CI: 3.04-8.86, Figure 1E). Statin use (HR = 0.55, 95% CI 0.35-0.86) and metformin (HR = 0.19, 95% CI 0.05-0.78) showed significant protective effects overall. Antiplatelet medication use had a significant treatment-by-phenogroup interaction. It was significantly associated with lower all-cause mortality in phenogroup 2 (HR=0.26, 95%CI: 0.10-0.67, p-interaction=0.004). We derived a simplified CART classifier based on NT-proBNP, E/e’, and left atrial diameter that well discriminated the phenogroups (macro F1-score=0.844, Figure 1F). Conclusion Our DeepCluster model identified two distinct cardiometabolic HFpEF phenogroups. The high-risk group was characterized by more severe diastolic dysfunctions and a greater cardiovascular-kidney-metabolic burden. The derived phenomapping classifier may support a pragmatic risk stratification and inform individualized treatment in cardiometabolic HFpEF.