Scholarlink

Electroencephalogram (EEG) signal classification plays a very important role in the detection of neurological disorders, particularly epilepsy. However, traditional deep learning (DL) models face significant challenges as they are non-stationary, high-dimensional and non-linear in character. In this research, we present a comprehensive evaluation of four advanced hybrid DL models like 1 Dimensional(1D) Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM), 1D CNN - Gated Recurrent Unit (CNN-GRU), 1D CNN - Long Short-Term Memory (CNN-LSTM), and LSTM–Kolmogorov–Arnold Network (LSTM-KAN) for binary seizure classification on the widely used UCI Epileptic Seizure Recognition dataset. K-Means-SMOTE is employed to address data imbalance, and the features are standardized using Z-score normalization. Stratified five-fold cross-validation is used for training each model. These models are assessed using five standard performance metrics. Empirical results reveal that while the 1D CNN–BiLSTM secure the highest average accuracy of 98.53%, LSTM-KAN model offers a balance between interpretability and efficiency. The models were also evaluated using the SEED benchmark dataset, and the performance of 1D CNN–BiLSTM reached an exceptional level of 99.64% accuracy.

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

Advancing EEG Signal Classification Using Hybrid Deep Learning Architectures and Kolmogorov–Arnold Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Madhurima Paul,
  • Sanjay Chakraborty,
  • Saikat Basu,
  • Koushik Majumder

摘要

Electroencephalogram (EEG) signal classification plays a very important role in the detection of neurological disorders, particularly epilepsy. However, traditional deep learning (DL) models face significant challenges as they are non-stationary, high-dimensional and non-linear in character. In this research, we present a comprehensive evaluation of four advanced hybrid DL models like 1 Dimensional(1D) Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM), 1D CNN - Gated Recurrent Unit (CNN-GRU), 1D CNN - Long Short-Term Memory (CNN-LSTM), and LSTM–Kolmogorov–Arnold Network (LSTM-KAN) for binary seizure classification on the widely used UCI Epileptic Seizure Recognition dataset. K-Means-SMOTE is employed to address data imbalance, and the features are standardized using Z-score normalization. Stratified five-fold cross-validation is used for training each model. These models are assessed using five standard performance metrics. Empirical results reveal that while the 1D CNN–BiLSTM secure the highest average accuracy of 98.53%, LSTM-KAN model offers a balance between interpretability and efficiency. The models were also evaluated using the SEED benchmark dataset, and the performance of 1D CNN–BiLSTM reached an exceptional level of 99.64% accuracy.