EClinicalMedicine (Oct 2025)
Machine learning-based phenotyping and assessment of treatment responses in heart failure with preserved ejection fractionResearch in context
Abstract
Summary: Background: Heart failure with preserved ejection fraction (HFpEF) accounts for over half of heart failure cases, yet effective treatments remain limited due to its clinical heterogeneity. This study aimed to identify distinct HFpEF phenotypes using machine-learning based algorithm and to compare treatment responses across different phenogroups. Methods: Our training cohort included 2147 hospitalized patients with HF with left ventricular ejection fraction (LVEF) ≥50% at Peking University Third Hospital (2014–2023). A two-stage DeepCluster model with a fully connected neural network was used to identify HFpEF phenogroups based on 107 demographic and clinical variables from electronic medical record (EMR). Cox proportional hazard models were used to assess patients' prognosis and treatment responses. The phenotyping model was validated internally using leave-one-out cross-validation method and externally with data from the TOPCAT clinical trial (n = 1696) and a well-characterized HFpEF patient cohort from the University of Michigan Health System (UMHS, n = 128). Findings: Three distinct HFpEF phenogroups were identified. Phenogroup 1 (n = 815) had the highest burden of metabolic comorbidities, along with left ventricular hypertrophy, and both systolic and diastolic dysfunction. Phenogroup 2 (n = 608) comprised predominantly females with atrial fibrillation and structural abnormalities in the atria and right ventricle, with mainly diastolic dysfunction. Phenogroup 3 (n = 724) included younger men with unhealthy lifestyles, higher burdens of hyperlipidemia, liver dysfunction, and relatively normal cardiac morphology and function. Overall, phenogroup 1 had the highest risk of all-cause mortality. We didn't observe significant survival benefits from major HF therapies overall. However, in phenogroup 1, post-diagnostic use of sodium-glucose cotransporter 2 inhibitors (SGLT2i) was associated with a 55% reduced risk of HF rehospitalization (HR = 0.45, 95% CI 0.20–0.99), while angiotensin receptor-neprilysin inhibitors (ARNIs) were associated with a 66% lower risk of all-cause mortality (HR = 0.34, 95% CI 0.14–0.79). The effect of ARNI on all-cause mortality differed significantly across the three phenogroups (p for interaction = 0.048). In phenogroup 2, calcium channel blockers were associated with a lower risk of all-cause mortality (HR = 0.62, 95% CI 0.38–0.99) and HF rehospitalization (HR = 0.58, 95% CI 0.38–0.88). Furthermore, our DeepCluster model was internally validated with 96.0% consistency. In external validation in the TOPCAT and UMHS cohort, we observed consistent clinical and pathophysiologic characteristics across the three identified phenogroups. Interpretation: Machine learning-based algorithm identifies HFpEF phenogroups with distinct clinical features and treatment responses. This study suggests that SGLT2i and ARNI have a role in improving the outcomes in patients with metabolic comorbidities and both systolic and diastolic impairment, while calcium channel blockers were most likely to benefit HFpEF patients with atrial fibrillation. Future prospective studies are required to further validate these findings. Funding: This study was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0539000), National Natural Science Foundation of China (82300291), Beijing Nova Program (20230484275), University of Michigan Medical School (UMMS) and Peking University Health Science Center (PKUHSC) Joint Institute for Translational and Clinical Research (BMU2023JI004), Clinical Cohort Construction Program of Peking University Third Hospital (BYSYDL2023009), CAMS Innovation Fund for Medical Sciences (2021-I2M-5-003), Peking University Shi-Ji Jin-Yuan Medical Foundation (48014Y0243), and Beijing Excellent Clinical Research Program (BRWEP2024W014090203).
Keywords
