With the advancement of theoretical principles of computer science and the availability of computational resources, deep learning has outperformed conventional machine learning and achieved better results in many fields (Si et al., 2021). Deep learning algorithms can extract data features that are automated and robust without the biases and complexities involved in feature engineering by hand (LeCun et al., 2015). As a result, some deep learning methods have been used for detecting epileptic seizures, including CNN (Du and Liu, 2022; O’Shea et al., 2020; Abdulwahhab et al., 2024),Long Short-Term Memory (LSTM) (Alharbi et al., 2024; Singh and Malhotra, 2022), Transformer models (Hu et al., 2023), and so on.

Ozdemir et al. (2021) introduced a novel approach that utilizes the Short-Time Fourier Transform (SST) and CNN to detect and forecast epileptic seizure episodes. The proposed method was evaluated on the IKCU and CHB-MIT datasets and achieved high accuracy and precision in detecting seizure episodes based on segments. Albaqami et al. (2023) proposed a novel deep learning network called Multi-Path Seizure Classification Network (MP-SeizNet), which consists of a CNN and a Bidirectional Long Short-Term Memory Neural Network (Bi-LSTM). The proposed method was evaluated on the EEG epilepsy database from the Temple University Hospital and achieved an F1 score of 87.6%. Li suggested a method for predicting seizures in EEG data, which involves using a Transformer-guided CNN (TGCNN) to extract both local and global characteristics efficiently (Li et al., 2022). This method combined different benefits of both CNN and Transformer models. It achieved a sensitivity of 82.2%, an FPR (False Positive Rate) of 0.06/h, and an AUC (Area Under the Curve) of 83.5% on the Kaggle database.

Since Shelhamer et al. (2017) proposed the FCN, it has found diverse applications across multiple domains, such as image segmentation and hyperspectral image classification, achieving excellent results. Jiang introduced a new end-to-end, pixel-level Fully Convolutional Spatial Propagation Network (FCSPN), consisting of a 3D-fully convolutional network (3D-FCN) and a convolutional spatial propagation network (CSPN) (Jiang et al., 2021). This method has achieved advanced performance in hyperspectral image (HSI) classification. Shen et al. (2021) proposed a method that combines a deep FCN with an effective non-local module (ENL-FCN) for hyperspectral HSI classification. The FCN takes the entire HSI as input in the proposed framework and extracts spectral-spatial information within local receptive fields (Shen et al., 2021). The proposed method achieved excellent classification performance while maintaining lower computational costs.

In recent years, FCN has also been applied in medical image processing. For example, Qiu has suggested a hybrid framework that merges FCN and Multi-Layer Perceptron (MLP) to generate high-resolution probability maps of diseases using local brain structures. And generate precise and intuitive individual Alzheimer’s disease risk visualization maps to achieve accurate diagnosis. The diagnostic accuracy achieved by this framework is comparable to that of neurologists (Qiu et al., 2020).

However, FCN has not yet been applied in seizure detection. Meanwhile, Due to the opaque nature of deep learning models, interpreting and comprehending the learned models in an intuitive manner remains a significant challenge (Khan et al., 2024). In particular, different channels of EEG signals contribute differently to seizure detection and the impact of different brain regions on seizure occurrence in the brain. Therefore, a CNN detection algorithm that combines an SE module on top of the FCN was proposed. The SE module offers the advantage of being easily pluggable and has not been previously applied to EEG-based seizure detection algorithms using deep learning. The algorithm not only performs epilepsy detection but also obtains the weights of each EEG signal channel, after which the corresponding brain regions can be found according to the cascade system correspondence. It enhances the model’s performance and interpretability.

The CHB-MIT database was utilized as the EEG database in this research. The details are as follows (Table 1). Records from 22 patients were available for this study, including 5 boys aged between 3 and 22 and 17 girls aged between 1.5 and 19. These records were divided into 23 individual cases, with one (chb21) belonging to the same patient as the case chb01 but collected 1.5 years later. The specific details of each subject can be found in Table 2. The chb24 case was included in the dataset in December 2010. However, the exact age and gender of this individual are not accurate. Each patient has between 9 and 42 consecutive.edf files, with the majority of these files containing precise one-hour signals. Each patient has an average of 5.75 seizure episodes of varying durations. In this database, the EEG signals were universally sampled at 256 Hz with a 16-bit resolution. Most EEG recordings include 18–23 bipolar EEG electrodes arranged according to the international 10–20 system (Abdallah et al., 2023; Liu et al., 2022).

The processed data was passed through a sixth-order Butterworth bandpass filter from 1 to 60 Hz to eliminate baseline noise and high-frequency interference. The raw EEG signals were resampled to a frequency of 128 Hz to reduce the data in the deep learning model. Because epileptic seizures have an intermittent nature, during seizures, EEG recordings are usually of shorter duration compared to non-seizure periods. Therefore, following previous research, the filtered EEG data was segmented into 2-s epochs using a sliding window analysis (Park et al., 2020). All the operations above were performed using the MNE-Python library in Python (Jaiswal et al., 2020).

A lightweight CNN combined with an SE module was used to build a seizure detection algorithm. Figure 1 shows a schematic representation of the proposed method’s framework, which draws inspiration from the FCN architecture for the model design, mainly utilizing its skip connections and deconvolution (upsampling) approach.

The EEG segment was initially configured with 32 channels. Due to each component containing 2 s of EEG data with a sampling rate of 128 Hz, the shape of the input EEG segment would be (32, 256). The data segment is passed through a two-dimensional convolutional layer in the next step. This layer consists of 32 filters with a kernel size of (1, 4) and a stride of (1, 2). Batch normalization and ReLU activation extract low-level features and minimize data superfluity. Following that, the data is passed through a convolutional layer with 64 filters, a kernel size of (1, 4), and a stride of (1, 2). Batch normalization and ReLU activation are applied again. The core idea of batch normalization is to perform standardization on the inputs of each layer in the network. Batch normalization layers can maintain the stability of the input data distribution while preserving the network’s nonlinear expressive power. The data is then passed through another convolutional layer with 128 filters, a kernel size of (1, 4), and a stride of (1, 2). At this stage, the output feature map shape would be (32 × 32 × 128). The output features are passed through the SE module, which focuses on channel relationships. The goal of the SE module is to enable the model to learn the relative significance of diverse channel features automatically (Hu et al., 2020).

The feature maps are then passed through a convolutional layer with two filters, a kernel size of (1, 1) and a stride of (1, 1). Next, they enter the first deconvolutional layer with two filters, a kernel size of (1, 4) and a stride of (1, 2). At this stage, the output feature map shape would be (32 × 64 × 2). The second convolutional layer has two filters, a kernel size of (1, 1) and a stride of (1, 1). The output features from the first convolutional layer are simultaneously fed into a deconvolutional layer with two filters, a kernel size of (1, 4), and a stride of (1, 2). At this stage, the output feature map shape would be (32 × 128 × 2). The output features from the first convolutional layer are passed through a convolutional layer with two filters, a kernel size of (1, 1), and a stride of (1, 1). Simultaneously, the second deconvolutional layer’s output features are fed into a deconvolutional layer with two filters, a kernel size of (1, 4), and a stride of (1, 2). At this stage, the output feature map shape would be (32 × 256 × 2). Finally, the output features are passed through a flattened layer and a Multilayer Perceptron (MLP) to output probabilities of epileptic and non-epileptic seizures (Table 3).

The details of SE are shown in Figure 2A. In SE, the first step is to perform a Squeeze operation (F_sq(⋅)) on the input feature U, resulting in z_c. This operation (1) decreases the number of dimensions in the feature map by taking the average value and extracting global feature information.

Next, the Excitation operation (F_ex(⋅)) is performed, where z represents the previous step’s output, and W1 and W2 denote linear layers. The computed s (2) in this step is the module’s core, meaning the weights for each channel.

Finally, the Scale operation [F_scale(⋅)] is performed. This results in the final output Xˇ of the SE module.

Where u_c represents a channel in U, and s_c means the weight of the channel. Therefore, it is equivalent to multiplying the value of each channel by its weight.

A leave-one-patient-out cross-validation approach was implemented to assess the effectiveness of the method, specifically designed for medical utilization. In the dataset of N patients, one patient’s data served as the test set, while the data from the remaining N-1 patients were employed as the training and validation sets. Each patient’s data was sequentially utilized as the test set, repeating this process N times to ensure comprehensive coverage (Park et al., 2020). Within the data of N-1 patients, 20% was arbitrarily picked as the validation set, while the remaining 80% was utilized for training the model.

The specifics of the training and testing approach are illustrated in Figure 3. During the training process, a batch size of 32 was utilized, and the model was trained for 100 epochs. The model that achieved the highest validation accuracy was chosen for the testing stage. The model was developed using Python 3.10 and Keras 2.4.3, and it is configured to operate on an NVIDIA GeForce RTX 3080 Ti GPU.

The EEG is an often utilized medical technique for diagnosing epilepsy (Duan et al., 2022). In the past few years, extensive interest has been in seizure detection based on EEG. EEGNet, proposed by Lawhern, is a compact CNN used for brain-computer interface based on EEG (Lawhern et al., 2018). Xu et al. (2020) proposed a 1D CNN-LSTM model for automatically identifying and Detecting epileptic seizures by analyzing EEG signals. Zuochen Wei proposed a CNN method for automatically detecting epileptic EEG signal segments and seizure events (Craley et al., 2021). These methods have demonstrated excellent performance.

However, the comparative results on the CHB-MIT dataset between the approach and these methods indicate that the designed CNN-based epileptic seizure detection algorithm exhibits superior advancement, stability, and reliability. This is because their CNN algorithms overlooked the importance of different channels in the feature maps during the convolution and pooling processes, leading to a performance loss. The approach incorporated the SE module, which helps reduce this loss by considering the significance of different channels. The SE module explicitly models interdependencies among convolutional feature channels to enhance the network’s representation capability, allowing feature recalibration to be performed (Hu et al., 2020). With this mechanism, global information can selectively emphasize enlightening characteristics and constrain usefulness (Chen et al., 2022). Therefore, introducing the SE module can improve performance metrics such as accuracy (Khan et al., 2022).

Moreover, skip connections were adopted from the FCN architecture, involving the connection of fundamental and advanced attributes. This enables the model better to capture crucial elements at different scales (Liu et al., 2021). Enhancing discriminating epileptic abnormal signals improves classification accuracy, ensuring the model’s robustness and precision. The skip connections facilitate the gradient propagation of shallow features to deeper layers, thereby accelerating the network’s training, learning, and convergence speed (Huang et al., 2022). It helps reduce training time and enhance the training and testing efficacy of the model. Indeed, the CNN model detection algorithm, which combines the SE module with the FCN base, enables more accurate identification of epileptic seizures in different patients. This approach effectively addresses the issue of doctors having to spend significant time and effort reviewing long-term EEG recordings.

In addition, the seizure onset zone of epilepsy can be further explored. Dissanayake et al. (2022) proposed two geometric deep learning (GDL) techniques. One approach involved utilizing graphs generated from the physical connections in the electroencephalogram (EEG) grid, while the other employed deep neural network-synthesized diagrams for predicting seizures in epilepsy. This method aids in the localization of epileptic seizures based on scalp EEG (Dissanayake et al., 2022). Narasimhan et al. (2020) utilized non-directional resting-state connectivity measurements to evaluate the connectivity between brain regions in patients and discovered that areas with increased epileptogenicity exhibited higher functional connectivity.

Similarly, the approach identifies seizures in different patients and considers the impact of brain regions on seizure occurrence. The system utilizes the FCN architecture with skip connections and deconvolution (upsampling) method. Deconvolution (upsampling) enlarges the shrunken feature maps to their original dimensions. If upsampling is only applied to the final layer’s feature map to obtain the actual image size, the results are often unsatisfactory. This is because the last layer’s feature map is too small, resulting in the loss of a significant amount of detail. Therefore, global and local effects can be balanced by combining the final layer (which contains global information) with shallower layers (which contain local information) through skip connections.

Visualizing the upsampled feature map of the last layer and normalizing it helps us obtain a weight matrix. In this weight matrix, the value at each position represents the importance of that position for epileptic seizures. By doing this, the classification can be extended from the image-level to the pixel-level. Each row of pixels forms a channel, allowing us to examine each row channel’s weight distribution, thus quantifying each row channel’s contribution to the occurrence of epileptic seizures. Each channel can correspond to specific brain regions using the international 10–20 electrode placement system. This allows us to investigate the significant impact of brain regions such as FZ, CZ, PZ, FT9, FT10, and T8 on the occurrence of epileptic seizures.

This paper proposes a compact CNN model on top of the FCN architecture for epileptic seizure detection, integrating the SE module. This algorithm is a method for independent epileptic seizure detection within the same dataset. Nevertheless, building an epileptic seizure detection algorithm that can generalize across different datasets is a more significant challenge. Developing an advanced algorithm for cross-dataset epileptic seizure detection is the next objective, demanding enhanced generalization capabilities from the algorithm.

In addition, while the contribution of EEG channel weights to epileptic seizure detection has been quantified, it is also being explored which brain regions play a significant role in the onset of epileptic seizures. However, the contrast of the weights in the EEG channel map needs to be more evident. Therefore, the algorithm still needs improvement to achieve better visualization and enhance the interpretability of the model.

It has been observed that effective channels are not the same across subjects and they may vary. In order to explore the variation in effective channels and to make the proposed model more generalizable, In the future, a dedicated dataset will be utilized for multicenter cross-validation.