Graph-Based Minimally Supervised Risk Stratification for Prognosis Prediction in Heart Failure With Preserved Ejection Fraction.
Abstract
Heart failure with preserved ejection fraction (HFpEF) is associated with poor outcomes, with 5-year survival after hospitalisation as low as 35-40%. Despite its growing prevalence, effective treatments remain limited. A major barrier to therapeutic success is the substantial biological and clinical heterogeneity within HFpEF, suggesting that distinct patient profiles may require different management strategies. This study investigates whether graph-based machine learning methods can identify distinct patient clusters within the HFpEF population and assess their relationships with clinical characteristics and outcomes. We analysed a multimodal dataset of 133 symptomatic HFpEF patients, comprising more than 240 K attributes derived from comprehensive cardiovascular assessment. These included cardiac MRI, echocardiography, circulating biomarkers of fibro-inflammation, and high-resolution proteomics. Five conventional unsupervised clustering methods were first applied but failed to produce clinically coherent groups, largely due to the complexity and multimodality of the data. To address this, we implemented a minimally supervised Graph Neural Network (GNN) that incorporated limited label information while modelling relationships between patients. This approach produced five clusters with potential clinical relevance. Importantly, cluster membership remained significantly associated with survival after adjustment for established clinical risk factors, including age, sex, BNP levels, symptom burden, and kidney function. Internal validation yielded an accuracy of 0.706 and an $F1$-score of 0.613, indicating reasonably consistent performance and suggesting that graph structure may aid generalisation and class-wise prognostic prediction. Overall, the findings show that minimally supervised graph-based learning can more effectively capture the underlying structure of multimodal HFpEF data than traditional methods. Visualisation and cluster quality metrics further support the effectiveness of the GNN approach.
Abstract as supplied by the indexing source. © the authors / publisher.