Scholarlink

<p>Autism spectrum disorder is a neurological condition that significantly impacts communication and social interactions. Psychological tests and neuroimaging are frequently used in traditional autism spectrum disorder detection techniques. However, electroencephalogram is an effective tool for identifying neurological disorders like autism spectrum disorder because it has the ability to capture real-time brain activity. Therefore, this paper introduces a novel framework called a hybrid convolutional-dynamic multi-level attention transformer model. The proposed framework integrates convolutional neural networks with a dynamic multi-level attention mechanism, which replaces the traditional self-attention layer in transformers, significantly reducing computational complexity while maintaining high detection accuracy. The electroencephalogram data recordings are initially preprocessed through re-referencing, filtering, and normalization procedures. The pre-processed data signals are segmented and transformed into spectrograms using the short-time Fourier transform technique. These spectrograms are used as model input for the training approach. The proposed model used VGG-16 and ResNet-50 to extract features, while attention-based transformer captures complex temporal patterns in the electroencephalogram spectrograms. Following feature extraction, the model refines the representation before classification by applying a dense layer to refine the combined feature set. The feature maps obtained from triple-stream modules are fused to form hybrid features and are classified into two distinct classes: non-autistic and autistic cases using the softmax classifier. The evaluation of the HC-DMAformer model attained a higher accuracy of 99.23%, across various Electroencephalogram datasets. Furthermore, the model is computationally efficient, with a training duration of 1&#xa0;h 29&#xa0;min, and a computation time of 13.18&#xa0;s. The result demonstrates that the proposed method outperforms than other methods.</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

HC-DMAformer: hybrid convolutional-dynamic multi-level attention transformer for efficient and accurate EEG-based autism detection

  • 期刊论文

  • 发表日期：

    2025年8月5日

查看全文
  • V. Kavitha,
  • R. Vidhya,
  • G. L. Swathi Mirthika,
  • K. Suresh,
  • S Hemavathi

摘要

Autism spectrum disorder is a neurological condition that significantly impacts communication and social interactions. Psychological tests and neuroimaging are frequently used in traditional autism spectrum disorder detection techniques. However, electroencephalogram is an effective tool for identifying neurological disorders like autism spectrum disorder because it has the ability to capture real-time brain activity. Therefore, this paper introduces a novel framework called a hybrid convolutional-dynamic multi-level attention transformer model. The proposed framework integrates convolutional neural networks with a dynamic multi-level attention mechanism, which replaces the traditional self-attention layer in transformers, significantly reducing computational complexity while maintaining high detection accuracy. The electroencephalogram data recordings are initially preprocessed through re-referencing, filtering, and normalization procedures. The pre-processed data signals are segmented and transformed into spectrograms using the short-time Fourier transform technique. These spectrograms are used as model input for the training approach. The proposed model used VGG-16 and ResNet-50 to extract features, while attention-based transformer captures complex temporal patterns in the electroencephalogram spectrograms. Following feature extraction, the model refines the representation before classification by applying a dense layer to refine the combined feature set. The feature maps obtained from triple-stream modules are fused to form hybrid features and are classified into two distinct classes: non-autistic and autistic cases using the softmax classifier. The evaluation of the HC-DMAformer model attained a higher accuracy of 99.23%, across various Electroencephalogram datasets. Furthermore, the model is computationally efficient, with a training duration of 1 h 29 min, and a computation time of 13.18 s. The result demonstrates that the proposed method outperforms than other methods.