Related Experiment Video
Updated: Apr 28, 2026
07:52
Multiscale Investigations of Cortical Processing by Integrating Laminar Polytrodes and Optogenetics with Micro Electrocorticography in Rodents
Published on: May 23, 2025
995
Multiscale Cosine Convolution Neural Network for Robust and Interpretable Epileptic EEG Detection
Jiale Chen1, Weidong Zhou1,2, Guoyang Liu1,2
1School of Integrated Circuits, Shandong University, Jinan 250101, China.
Biosensors
|April 27, 2026
Summary
This study introduces a novel deep learning network for accurate epileptic seizure detection from electroencephalogram (EEG) signals. The proposed model enhances performance and interpretability, offering potential for improved patient monitoring and diagnostics.
Area of Science:
- Artificial Intelligence
- Biomedical Engineering
- Neurology
Background:
- Accurate epileptic seizure detection via electroencephalogram (EEG) is crucial for clinical diagnosis but faces challenges in performance and interpretability.
- Existing methods often struggle with low detection rates and lack transparency in their decision-making processes.
Purpose of the Study:
- To develop an advanced deep learning network, the Multiscale Cosine Convolutional Heterogeneous Two-Stream Cosine Convolution Network (MCC-HTSCC), to address limitations in EEG-based epilepsy detection.
- To improve both the accuracy and interpretability of seizure detection models.
Main Methods:
- The proposed MCC-HTSCC network utilizes a Multiscale Cosine Convolution (MCC) module for extracting multiscale temporal features from raw EEG signals.
- Spatial convolutional layers process temporal features to create comprehensive spatiotemporal representations.
- A Heterogeneous Two-Stream Cosine Convolution (HTSCC) module with deep and shallow subnetworks performs hierarchical feature extraction and classification.
Main Results:
- The MCC-HTSCC model achieved high performance on the CHB-MIT dataset (98.52% accuracy, 97.98% sensitivity, 98.50% specificity) and the SH-SDU dataset (94.56% accuracy, 88.09% sensitivity, 95.89% specificity).
- Cosine convolution operators reduced model parameters by ~18.12% compared to traditional methods, enhancing suitability for embedded systems.
- Gradient-Weighted Class Activation Mapping (Grad-CAM) provided model interpretability.
Conclusions:
- The MCC-HTSCC network demonstrates significant potential for accurate and interpretable epileptic seizure detection.
- The model's efficiency and transparency make it a promising tool for patient-specific epilepsy monitoring and clinical diagnostics.
