Scholarlink

<p>This study presents a graph-based learning framework employing a Graph Isomorphism Network (GIN) for binary dementia classification using the Open Access Series of Imaging Studies (OASIS)-2 dataset. Clinical and volumetric features were transformed into a subject-similarity graph via cosine-based k-nearest neighbor construction, enabling the model to capture inter-subject relationships often missed by conventional approaches. The proposed two-layer GIN achieved 99.40% accuracy, 98.67% precision, 100% recall, an Area Under the Curve (AUC) of 0.9998, and Cohen’s Kappa of 0.9879 under five-fold stratified cross-validation. Model interpretability was enhanced through Shapley Additive exPlanations (SHAP) analysis, which identified Clinical Dementia Rating (CDR), Mini-Mental State Examination (MMSE), and education level as key decision drivers, aligning with established clinical understanding. Comparative experiments demonstrated superior performance over baseline Graph Convolutional Network (GCN) and GraphSAGE architectures. These results indicate that integrating graph-based relational modeling with explainable Artificial Intelligence (AI) provides an effective and interpretable solution for dementia classification.</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

Graph isomorphism network with explainable learning for dementia screening using neurocognitive assessments

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Chand Rani,
  • Pooja Dogra,
  • Helen Babu,
  • Varun P. Gopi,
  • Nalina Gupta

摘要

This study presents a graph-based learning framework employing a Graph Isomorphism Network (GIN) for binary dementia classification using the Open Access Series of Imaging Studies (OASIS)-2 dataset. Clinical and volumetric features were transformed into a subject-similarity graph via cosine-based k-nearest neighbor construction, enabling the model to capture inter-subject relationships often missed by conventional approaches. The proposed two-layer GIN achieved 99.40% accuracy, 98.67% precision, 100% recall, an Area Under the Curve (AUC) of 0.9998, and Cohen’s Kappa of 0.9879 under five-fold stratified cross-validation. Model interpretability was enhanced through Shapley Additive exPlanations (SHAP) analysis, which identified Clinical Dementia Rating (CDR), Mini-Mental State Examination (MMSE), and education level as key decision drivers, aligning with established clinical understanding. Comparative experiments demonstrated superior performance over baseline Graph Convolutional Network (GCN) and GraphSAGE architectures. These results indicate that integrating graph-based relational modeling with explainable Artificial Intelligence (AI) provides an effective and interpretable solution for dementia classification.