Scholarlink

Heart disease is a significant factor in global mortality rates, emphasizing the need for precise and dependable predictive models. Traditional models often struggle with datasets that are imbalanced. In real-world data, the number of patients suffering from heart disease is always significantly lower than those who are healthy. The Graph Convolutional Network (GCN) is one such model that can be highly effective for predictions but performs poorly when subjected to imbalanced data. There are several known techniques to handle data imbalance effectively. In this paper, we provide a comprehensive analysis of Boosting Graph Neural Networks, the Synthetic Minority Over-sampling Technique (SMOTE), a weighted loss function, and a curriculum learning approach to understand the relative effectiveness of these methods. Additionally, we aim to improve the clarity and understandability of the most promising model using Explainable AI technology. We utilize SHAP, a technique used to explain the individual contributions of features to the model’s predictions. Ultimately, our goal is to create more effective, reliable, and easily understood applications.

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

Leveraging Graph Convolutional Networks for Heart Disease Detection with Limited and Imbalanced Data: A Focus on Interpretability

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Bhumika S. Mandikel,
  • Chandana S,
  • Arundhati Sitharam,
  • Fiza Haneen,
  • Bhaskarjyoti Das

摘要

Heart disease is a significant factor in global mortality rates, emphasizing the need for precise and dependable predictive models. Traditional models often struggle with datasets that are imbalanced. In real-world data, the number of patients suffering from heart disease is always significantly lower than those who are healthy. The Graph Convolutional Network (GCN) is one such model that can be highly effective for predictions but performs poorly when subjected to imbalanced data. There are several known techniques to handle data imbalance effectively. In this paper, we provide a comprehensive analysis of Boosting Graph Neural Networks, the Synthetic Minority Over-sampling Technique (SMOTE), a weighted loss function, and a curriculum learning approach to understand the relative effectiveness of these methods. Additionally, we aim to improve the clarity and understandability of the most promising model using Explainable AI technology. We utilize SHAP, a technique used to explain the individual contributions of features to the model’s predictions. Ultimately, our goal is to create more effective, reliable, and easily understood applications.