TL;DR
This paper introduces MSE-VGG, a novel deep learning model that uses electroencephalography (EEG) signals to rapidly detect ischemic stroke. By fusing novel connectivity features (cwPLI) with complexity measures (SaEn) and employing an enhanced VGG architecture with attention mechanisms, the model achieves high accuracy, paving the way for faster and more accessible stroke diagnosis.
What Problem Does This Paper Address?
Ischemic stroke, a leading cause of death and disability, requires immediate treatment to minimize damage. Current diagnostic methods like CT and MRI, while accurate, are expensive, time-consuming, and require specialized facilities, delaying critical interventions. Electroencephalography (EEG) offers a promising alternative due to its low cost, portability, and real-time capabilities. However, existing EEG-based stroke detection methods have struggled to achieve the high accuracy needed for clinical adoption. This paper aims to bridge this gap by developing a more robust and accurate EEG-based detection system.
Methods
The core of this research lies in a novel feature extraction and classification pipeline:
- Feature Extraction:
- Correlation-weighted Phase Lag Index (cwPLI): This is a novel feature designed to capture spatio-temporal information and functional connectivity between EEG channels. It builds upon the Phase Lag Index (PLI) by weighting the PLI values based on their correlation with stroke vs. non-stroke classifications, aiming to amplify discriminative patterns.
- Sample Entropy (SaEn): This standard nonlinear dynamic analysis technique is used to quantify the complexity and regularity of EEG signals, providing insights into brain state changes.
- Feature Fusion: The upper triangle of the cwPLI matrix is fused with the SaEn array to create a comprehensive feature representation. This fused feature is then converted into an RGB image for input into the classification model.
- Classification Model (MSE-VGG):
- The authors propose MSE-VGG, an enhanced version of the VGG neural network.
- Squeeze-and-Excitation (SE) Modules: Multiple SE modules are integrated into the VGG architecture. These modules act as attention mechanisms, allowing the network to dynamically re-weight channel-wise feature responses, focusing on more informative channels and improving feature representation.
- The model retains the VGG's deep convolutional structure (13 convolutional layers, 3 fully connected layers, 5 pooling layers) but with the added SE blocks after each pooling layer.
Datasets
The study utilized a private dataset named ZJU4H:
- Ischemic Stroke Patients: 20 individuals (11 males, 9 females; aged 47-87) diagnosed with ischemic stroke.
- Non-Stroke Controls: 19 healthy individuals (12 males, 7 females; aged 45-76) with no history of stroke.
- Data Acquisition: EEG data was collected using an XLTEK-EEG32U system with 32 channels, retaining 19 channels consistent with the international 10-20 electrode placement standard. The sampling rate was 256 Hz, and approximately 5 minutes of EEG data was recorded per patient.
- Labeling: Subjects were labeled based on MRI diagnosis.
Key Insights and Ideas
- cwPLI as a Superior Connectivity Measure: The research demonstrates that cwPLI, by incorporating class-specific weighting, provides a more pronounced distinction in functional connectivity between ischemic stroke patients and healthy controls compared to standard PLI. This suggests that tailoring connectivity features to the specific diagnostic task can significantly improve performance.
- Fusion is Key: Combining functional connectivity (cwPLI) with signal complexity (SaEn) yields a more powerful feature set. This fusion captures both the spatio-temporal relationships and the nonlinear dynamics of brain activity, which are crucial for accurate stroke detection.
- Attention Enhances CNNs for EEG: The integration of SE modules into the VGG architecture significantly boosts its ability to discern subtle EEG patterns. This highlights the effectiveness of attention mechanisms in deep learning for complex biological signal analysis.
- EEG's Practical Advantage: The study strongly emphasizes the feasibility of EEG for real-world stroke detection due to its portability, low cost, and speed, making it suitable for use outside of traditional hospital settings.
Key Results
- High Performance: The proposed MSE-VGG model with the cwPLI+SaEn fusion feature achieved an accuracy of 90.17%, sensitivity of 89.86%, and specificity of 90.41% on the ZJU4H dataset.
- Feature Superiority: The cwPLI+SaEn fusion feature outperformed individual features like PLI, cwPLI, PCC, MI, and PLV, achieving the highest AUC (0.9013) in ROC analysis.
- Model Superiority: MSE-VGG demonstrated better performance than other established classifiers (LR, AlexNet, RCF, LSTM, VGG) when using the cwPLI+SaEn feature.
- Efficiency: The proposed method requires only about 10 minutes for EEG signal collection and 4.62 seconds for model running time, significantly faster than traditional methods.
Limitations and Future Work
- Dataset Size and Diversity: The study was conducted on a private dataset (ZJU4H) with a relatively small number of subjects (20 stroke patients, 19 controls). The authors acknowledge that this limits generalizability.
- Channel Dependency: The proposed cwPLI feature extraction method is noted to be dependent on the specific EEG channel locations used (10-20 system).
- Future Work:
- Larger and Diverse Datasets: Collect more diverse EEG data from a larger population to validate and improve the model's robustness.
- Model Optimization: Further refine the MSE-VGG model to enhance its recognition performance.
- Real-time Application: Explore and optimize the model for real-time clinical deployment.
Summary updated