Abstract:Recently, most seizure prediction methods mainly utilize pure CNN or Transformer model, which cannot extract local and global features simultaneously. To this end, we propose an Electroencephalogram (EEG) seizure prediction method based on Transformer guided CNN (TGCNN), which combines the complementary advantages of CNN and Transformer. The proposed method first use short-time Fourier transform (STFT) to extract time–frequency features from EEG signals. Then, these features are fed into the alternating structure to model both local feature and long-distance dependencies, which can overcome both the deficiency of long distance dependence in CNN and the lack of local features in Transformer. Finally, the prediction result is obtained through a global average pooling layer and fully connected layer. The proposed method achieves sensitivity of 91.5%, false prediction rate (FPR) of 0.145/h, and area under curve (AUC) of 93.5% on CHB-MIT database and 82.2% sensitivity, 0.06/h FPR, and 83.5% AUC on Kaggle dataset.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that the existing methods for epileptic seizure prediction have limitations in feature extraction, specifically as follows:
1. **Pure CNN model**: Although it can extract local features well, it is difficult to capture long - distance dependencies (i.e., global features), which limits its accuracy in epileptic seizure prediction.
2. **Pure Transformer model**: Although it is good at capturing long - distance dependencies, it performs poorly in extracting local features.
To solve the above problems, the author proposes a method based on Transformer - Guided Convolutional Neural Network (TGCNN). This method combines the advantages of CNN and Transformer and aims to simultaneously extract local features and long - distance dependencies, thereby improving the accuracy of epileptic seizure prediction.
### Specific Problem Description
Epilepsy is a chronic non - infectious brain disease that affects people of all ages. Approximately 25% of patients cannot achieve a seizure - free state through surgery or drugs. For these patients, accurate prediction of epileptic seizures can significantly reduce various risks caused by epilepsy, thereby improving their quality of life. Therefore, it is of great significance to study how to accurately predict epileptic seizures from EEG signals.
### Limitations of Existing Methods
1. **Traditional machine learning methods**: They require a great deal of expert experience and feature engineering, and have poor generalization ability when data is limited.
2. **Deep learning methods**:
- **CNN**: It mainly uses local features and ignores global information or long - distance dependencies.
- **Transformer**: Although it can capture long - distance dependencies, it has insufficient extraction of local features, and may be affected by time - relatedness during training, resulting in the model learning time - relatedness rather than the real distinguishing features.
### Proposed Solution
The author proposes a deep - learning framework named Transformer - Guided Convolutional Neural Network (TGCNN), aiming to combine the complementary advantages of CNN and Transformer. Specifically:
- **STFT (Short - Time Fourier Transform)**: It is used to extract time - frequency features from EEG signals.
- **Alternating structure**: First, use CNN units for preliminary local feature extraction, and then input these features into an alternating structure containing CNN and Transformer to further model local features and long - distance dependencies.
- **Multi - scale feature extraction**: Gradually reduce model parameters through phased design to reduce the computational burden.
- **Classification layer**: Use the global average pooling layer and the fully - connected layer for final classification.
### Experimental Results
The experimental results on two public datasets, CHB - MIT and Kaggle, show that the proposed TGCNN method is superior to the existing CNN and ResNet models in terms of sensitivity, false positive rate (FPR), and AUC. The specific indicators are as follows:
- **CHB - MIT dataset**: Sensitivity 91.5%, FPR 0.145/h, AUC 93.5%.
- **Kaggle dataset**: Sensitivity 82.2%, FPR 0.06/h, AUC 83.5%.
These results indicate that TGCNN can significantly improve the accuracy of epileptic seizure prediction while maintaining a low false positive rate.