Scholarlink

<p>The Autism Spectrum Disorder (ASD) is neurodevelopment disorder with unusual social communication, and repetitive behaviours. Early and accurate diagnosis remains a clinical challenge, particularly in resource-limited settings where neuroimaging-based facilities are not available. This study proposed a graph neural network (GNN) models for ASD classification using behavioural features extracted from the AVSRT dataset. The nineteen features were extracted and used to construct patient similarity graphs through k-nearest-neighbour adjacency. The three GNN architectures includes Graph Convolutional Network (GCN), Gated Graph Neural Network (GGNN), and Message Passing Neural Network (MPNN) were trained and evaluated under a fully inductive method with the multi-seed protocol and compared against Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR) baselines. The MPNN achieved the highest balanced accuracy of 0.597, followed by GGNN at 0.575, outperforming all conventional classifiers. The feature importance analysis revealed that the inter-stimulus intervals and stimulus count features as primary discriminative biomarkers. The Ablation studies demonstrated MPNN’s feature dependence with removal of standard inter-stimulus intervals at step 9 causing a 19.97% balanced accuracy drop by 19.97%, The statistical significance testing further confirmed that GCN was significantly outperformed by both GGNN and MPNN, with p-values of 0.00084 and 0.0046 on accuracy, respectively. The computational analysis established that all models are deployable within routine clinical settings, with MPNN requiring only 0.043&#xa0;MB and achieving inference in approximately 495 microseconds per sample. These findings demonstrated that the graph-based machine learning models applied to behavioural reaction time features can provide clinically interpretable ASD 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

Investigating the efficacy of graph convolutional, gated graph, and message passing neural networks for autism spectrum disorder diagnosis

  • 期刊论文

  • 发表日期：

    2026年7月31日

查看全文
  • J. Revathi

摘要

The Autism Spectrum Disorder (ASD) is neurodevelopment disorder with unusual social communication, and repetitive behaviours. Early and accurate diagnosis remains a clinical challenge, particularly in resource-limited settings where neuroimaging-based facilities are not available. This study proposed a graph neural network (GNN) models for ASD classification using behavioural features extracted from the AVSRT dataset. The nineteen features were extracted and used to construct patient similarity graphs through k-nearest-neighbour adjacency. The three GNN architectures includes Graph Convolutional Network (GCN), Gated Graph Neural Network (GGNN), and Message Passing Neural Network (MPNN) were trained and evaluated under a fully inductive method with the multi-seed protocol and compared against Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR) baselines. The MPNN achieved the highest balanced accuracy of 0.597, followed by GGNN at 0.575, outperforming all conventional classifiers. The feature importance analysis revealed that the inter-stimulus intervals and stimulus count features as primary discriminative biomarkers. The Ablation studies demonstrated MPNN’s feature dependence with removal of standard inter-stimulus intervals at step 9 causing a 19.97% balanced accuracy drop by 19.97%, The statistical significance testing further confirmed that GCN was significantly outperformed by both GGNN and MPNN, with p-values of 0.00084 and 0.0046 on accuracy, respectively. The computational analysis established that all models are deployable within routine clinical settings, with MPNN requiring only 0.043 MB and achieving inference in approximately 495 microseconds per sample. These findings demonstrated that the graph-based machine learning models applied to behavioural reaction time features can provide clinically interpretable ASD classification.