Scholarlink

<p>The growing complexity of modern healthcare requires advanced analytical frameworks capable of integrating heterogeneous clinical data. We present a novel hybrid graph-based framework that combines Graph Neural Networks, Bidirectional Long Short-Term Memory networks, and ClinicalBERT embeddings, enhanced with a Variational Graph Autoencoder for medical link prediction. This unified design jointly models temporal dynamics, contextual semantics, and structural relationships between patients, symptoms, and treatments, enabling a comprehensive representation of clinical interactions. Extensive experiments conducted on real-world datasets—MIMIC-III, DrugBank, and SIDER—demonstrate the superior performance of our approach, achieving an Accuracy of&#xa0;0.85, F1-score of&#xa0;0.82, and AUC-ROC of&#xa0;0.84, outperforming all baseline models. By fusing temporal, textual, and graph-based reasoning, the proposed framework delivers interpretable, scalable, and clinically relevant insights for decision support. These findings highlight its potential to enhance diagnostic precision, enable personalized treatment strategies, and advance AI-driven medical research.</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

Hybrid graph-based models in medical decision support

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Riadh Bouslimi,
  • Fatma Zitouni,
  • Takwa Ben Smida,
  • Wahiba Ben abdsalem Karaa

摘要

The growing complexity of modern healthcare requires advanced analytical frameworks capable of integrating heterogeneous clinical data. We present a novel hybrid graph-based framework that combines Graph Neural Networks, Bidirectional Long Short-Term Memory networks, and ClinicalBERT embeddings, enhanced with a Variational Graph Autoencoder for medical link prediction. This unified design jointly models temporal dynamics, contextual semantics, and structural relationships between patients, symptoms, and treatments, enabling a comprehensive representation of clinical interactions. Extensive experiments conducted on real-world datasets—MIMIC-III, DrugBank, and SIDER—demonstrate the superior performance of our approach, achieving an Accuracy of 0.85, F1-score of 0.82, and AUC-ROC of 0.84, outperforming all baseline models. By fusing temporal, textual, and graph-based reasoning, the proposed framework delivers interpretable, scalable, and clinically relevant insights for decision support. These findings highlight its potential to enhance diagnostic precision, enable personalized treatment strategies, and advance AI-driven medical research.