Scholarlink

<p>Electronic health records (EHRs) encode complex, temporally evolving interactions among patients, providers, and clinical events in intensive care units (ICUs). While graph neural networks (GNNs) have shown strong performance for clinical outcome prediction, existing approaches rarely integrate multi-scale topological structure from persistent homology (PH) in a principled and learnable way within graph representation learning. We present a topology-aware graph learning framework for healthcare networks that incorporates both graph and hypergraph persistent homology into a unified GNN architecture through a persistence-aware pooling mechanism, LAI-Pool. We first construct patient-specific healthcare networks from MIMIC-IV EHR data, where nodes represent patients, admissions, diagnoses, procedures, providers, ICU stays, and ICU units, and temporal edges and hyperedges encode structured care interactions. We then compute PH over time-based filtrations of graph clique complexes and weighted hypergraph nerves to capture complementary multi-scale structural patterns. We transform topological features into learnable persistence representations and integrate with node embeddings via topology-guided pooling and cross-attention mechanisms. Experiments on 65,366 ICU patients demonstrate that integrating graph and hypergraph PH with GNN embeddings consistently improves mortality prediction compared to graph-only baselines. Joint fusion of both topological signals achieves the strongest performance, indicating that persistent topological structure provides complementary predictive information beyond standard healthcare network representations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Topological graph learning on healthcare networks for critical care prediction

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Charles Fanning,
  • Ahmed Guler,
  • Corey Verkouteren,
  • Djalil Sawadogo,
  • Mehmet Emin Aktas

摘要

Electronic health records (EHRs) encode complex, temporally evolving interactions among patients, providers, and clinical events in intensive care units (ICUs). While graph neural networks (GNNs) have shown strong performance for clinical outcome prediction, existing approaches rarely integrate multi-scale topological structure from persistent homology (PH) in a principled and learnable way within graph representation learning. We present a topology-aware graph learning framework for healthcare networks that incorporates both graph and hypergraph persistent homology into a unified GNN architecture through a persistence-aware pooling mechanism, LAI-Pool. We first construct patient-specific healthcare networks from MIMIC-IV EHR data, where nodes represent patients, admissions, diagnoses, procedures, providers, ICU stays, and ICU units, and temporal edges and hyperedges encode structured care interactions. We then compute PH over time-based filtrations of graph clique complexes and weighted hypergraph nerves to capture complementary multi-scale structural patterns. We transform topological features into learnable persistence representations and integrate with node embeddings via topology-guided pooling and cross-attention mechanisms. Experiments on 65,366 ICU patients demonstrate that integrating graph and hypergraph PH with GNN embeddings consistently improves mortality prediction compared to graph-only baselines. Joint fusion of both topological signals achieves the strongest performance, indicating that persistent topological structure provides complementary predictive information beyond standard healthcare network representations.