Scholarlink

<p>Personalized medication recommendation must ensure both efficacy and safety, particularly by avoiding drug–drug interactions (DDIs), a major cause of preventable adverse events. We introduce a <i>heterogeneous graph learning framework</i> that unifies electronic health records (EHRs) using explicit patient–diagnosis, patient–drug, and diagnosis–drug edges together with curated DDI interaction edges. Patient- and diagnosis-specific embeddings are learned inductively via <i>GraphSAGE</i>, while pharmacological risk is modeled through a <i>graph convolutional network</i> (GCN) on the DDI subgraph. A <i>DDI-aware cross-attention mechanism</i> fuses these representations, aligning patient context with safety signals and penalizing unsafe co-prescriptions during training. Evaluation on the <i>MIMIC-III</i> dataset demonstrates that our model consistently outperforms strong baselines, achieving superior predictive accuracy (Jaccard 0.5281, F1 0.6701, PR-AUC 0.7754) while reducing unsafe recommendations (DDI Rate 0.0581). All improvements are statistically significant (p<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\ &lt; 0.05\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mspace width="4pt" /> <mo>&lt;</mo> <mn>0.05</mn> </mrow> </math></EquationSource> </InlineEquation>). By integrating inductive heterogeneous graph embeddings with safety-aware attention, our framework bridges predictive accuracy and pharmacological safety, generalizes to unseen patients, and avoids clinically hazardous combinations.</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

Safe and personalized drug recommendation via heterogeneous graph learning: integrating graphSAGE representations with DDI-aware cross-attention

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Navid Khaledian,
  • Sonya Falahati

摘要

Personalized medication recommendation must ensure both efficacy and safety, particularly by avoiding drug–drug interactions (DDIs), a major cause of preventable adverse events. We introduce a heterogeneous graph learning framework that unifies electronic health records (EHRs) using explicit patient–diagnosis, patient–drug, and diagnosis–drug edges together with curated DDI interaction edges. Patient- and diagnosis-specific embeddings are learned inductively via GraphSAGE, while pharmacological risk is modeled through a graph convolutional network (GCN) on the DDI subgraph. A DDI-aware cross-attention mechanism fuses these representations, aligning patient context with safety signals and penalizing unsafe co-prescriptions during training. Evaluation on the MIMIC-III dataset demonstrates that our model consistently outperforms strong baselines, achieving superior predictive accuracy (Jaccard 0.5281, F1 0.6701, PR-AUC 0.7754) while reducing unsafe recommendations (DDI Rate 0.0581). All improvements are statistically significant (p \(\ < 0.05\) < 0.05 ). By integrating inductive heterogeneous graph embeddings with safety-aware attention, our framework bridges predictive accuracy and pharmacological safety, generalizes to unseen patients, and avoids clinically hazardous combinations.