1. Introduction

Epilepsy is a chronic brain disorder caused by abnormal hypersynchronous neuronal discharges^[1,2], affecting approximately 50 million people worldwide^[3]. Patients with epilepsy face a premature mortality rate that is two to three times higher than that of the general population, imposing a profound burden on individuals, families, and society^[4,5]. Therefore, accurate seizure prediction is crucial, enabling timely intervention and helping reduce or prevent seizure-related harm. Electroencephalography (EEG) captures abnormal brain rhythms associated with epilepsy, including spikes, slow waves, sharp waves, spike-slow wave complexes, and sharp-slow wave complexes, etc^[6]. EEG-based seizure prediction systems can detect these patterns during the preictal period, enabling timely and accurate seizure forecasting^[7-10].

Numerous methods have emerged for EEG-based seizure prediction, broadly falling into traditional machine learning and deep learning categories. Traditional machine learning approaches^[11,12] for seizure prediction rely on manually designed features and conventional classifiers such as Bayesian classifiers^[13], support vector machines^[5], and k-nearest neighbors^[14]. In recent years, deep learning approaches have been widely adopted due to their data-driven feature learning capability and improved performance^[15-20]. Representative studies include convolutional neural network (CNN)-based frameworks using handcrafted time–frequency or spectral features, such as wavelet packet decomposition with common spatial patterns^[18] and topology-based image encoding of spectral and statistical features^[19].

Despite these advances, seizure prediction performance using CNNs remains limited, primarily due to their reliance on local convolution operations, which restricts their ability to capture global temporal dependencies in EEG signals^[21]. Preictal EEG, however, evolves gradually, exhibiting patterns such as spikes, low-voltage fast discharges (LFD), and high-frequency oscillations (HFOs) that may emerge tens of minutes before seizure onset^[22-24]. To capture such long-range temporal dynamics, bidirectional long short-term memory (BiLSTM) networks have been introduced into seizure prediction tasks^[25]. Nevertheless, existing BiLSTM-based methods still show suboptimal performance^[26,27], partly because redundant or less informative temporal features may obscure critical preictal patterns.

To address these issues, we propose a BiLSTM model enhanced with a channel attention mechanism, termed Attention-BiLSTM. This architecture enables the model to capture global temporal dependencies in preictal EEG while adaptively emphasizing informative temporal features, thereby improving seizure prediction performance. We employ a dense feature transformation at the input stage to provide high-quality EEG representations for effective attention learning. Furthermore, to enhance model interpretability, we analyze the attention weights and feature distribution of the proposed model. Our findings demonstrate that Attention-BiLSTM not only progressively enhances feature discriminability but also directs the attention to the temporal features most relevant to seizure prediction.

2. Methods

2.1 CHB-MIT dataset

We evaluated the proposed model on the CHB-MIT dataset, a widely used and well-established public collection of scalp EEG recordings from Boston Children’s Hospital for seizure prediction^[28-30]. The dataset contains recordings from 23 pediatric patients with refractory epilepsy (5 males, 17 females, and one patient with missing gender/age information). The EEG recordings form 24 cases, as case chb21 comes from re-monitoring the same patient from case chb01 after 1.5 years. A Nihon Kohden EEG-1100A system recorded 23 differential electrode voltages at a sampling rate of 256 Hz with 16-bit resolution. Electrodes were placed according to the International 10-20 system. Table 1 provides detailed information on the CHB-MIT dataset.

2.2 Pre-processing

In general, the interictal period is defined as the time interval starting at least 4 hours after the end of the last seizure and ending at least 4 hours before the start of the next seizure^[31]. For cases chb12 and chb13, the electrode types, sequences, and numbers, and reference schemes varied multiple times. Therefore, we excluded cases chb12 and chb13 from the analysis. Following the recommendation of Maiwald et al.^[32], we set the preictal horizon to 30 minutes to ensure sufficient time for seizure prediction. Within the preictal and interictal periods, EEG signals were segmented into non-overlapping 3-second windows to increase the number of samples^[33]. For each patient included in this study, a total of 23 differential electrode voltages were recorded: FP1–F7, F7–T7, T7–P7, P7–O1, FP1–F3, F3–C3, C3–P3, P3–O1, FP2–F4, F4–C4, C4–P4, P4–O2, FP2–F8, F8–T8, T8–P8, P8–O2, FZ–CZ, CZ–PZ, P7–T7, T7–FT9, FT9–FT10, FT10–T8, and T8–P8. As the channel T8–P8 appeared twice, duplicate channels were removed, resulting in 22 channels corresponding to the maximum set of commonly available channels shared across 21 subjects. To suppress power-line interference introduced during EEG acquisition, a notch filter centered at 60 Hz was applied to remove line noise and its harmonics.

2.3 Attention-BiLSTM model

In our data-driven seizure prediction approach, let $X^{\left(i\right)}\in R^{N_p\times N_{ch}}$ represents the i-th training sample, i = 1, 2, ···, N_t, where N_t is the number of training samples, N_p is the number of sample points, and N_ch is the number of channels. We structure this section into three main modules: BiLSTM, Channel Attention, and Attention-BiLSTM. Figure 1 illustrates the overall architecture.

Display Full Size

Figure 1. Flow chart of the Attention-BiLSTM architecture. BiLSTM: bidirectional long short-term memory; EEG: electroencephalogram; BN: batch normalization.

2.3.1 BiLSTM

LSTM is a type of recurrent neural network designed to capture long-term dependencies in sequential data and time-series prediction. BiLSTM extends LSTM by processing sequences in both forward and backward directions, enabling it to capture temporal dependencies from both past and future EEG signals, which is particularly important for modeling gradual preictal changes that precede seizures. The BiLSTM module incorporates both forward and backward LSTM layers, each with 256 hidden units. The core component of an LSTM layer is the memory unit, which stores information across time steps via three gating mechanisms: the input gate i_t, the forget gate f_t, and the output gate o_t. These gates determine what information to add, retain, or output at each time step. By combining the memory unit with gating mechanisms, LSTM can selectively retain, update, and propagate information over long sequences, enabling effective modeling of long-term temporal dependencies. The gating mechanisms and memory state updates of LSTM can be expressed as follows:

(1)$f_t=\sigma (W_{fh}\cdot h_{t-1}+W_{fx}\cdot x_t+b_f)$

(2)$i_t=\sigma (W_{ih}\cdot h_{t-1}+W_{ix}\cdot x_t+b_i)$

(3)$\widetilde{c_t}=\tanh (W_{ch}\cdot h_{t-1}+W_{cx}\cdot x_t+b_c)$

(4)$c_t=f_t\odot \cdot c_{t-1}+i_t\odot {\tilde{c}}_t$

(5)$o_t=\sigma (W_{oh}\cdot h_{t-1}+W_{ox}\cdot x_t+b_o)$

(6)$h_t=o_t\odot \tanh \left(c_t\right)$

where ${\tilde{c}}_{{}_t}$ denotes the candidate memory at time step t, c_t denotes the memory cell state, h_t denotes the hidden state. For Equations (1)-(6), W is the weight matrix, b is the bias vector, and σ represents the sigmoid activation function. The forget gate f_t in Equation (1) determines how much information from the previous memory cell should be retained or discarded. The input gate i_t in Equation (2) represents how much of the current memory cell is written into the current memory cell. The candidate memory ${\tilde{c}}_{{}_t}$ in Equation (3) calculates the new information generated from the current input and previous hidden state. The current memory cell c_t in Equation (4) is then updated by combining retained past information and new candidate information. The output gate o_t in Equation (5) controls how much the current memory cell contributes to the hidden state. Finally, the current hidden state h_t is calculated by filtering the activated memory cell through the output gate, as shown in Equation (6). Based on the above memory unit and gating mechanisms, LSTM captures long-term temporal dependencies in EEG signals from preictal and interictal periods. The BiLSTM further extends this capability by processing sequences in both forward and backward directions, enabling effective extraction of discriminative temporal features critical for seizure prediction.

2.3.2 Channel attention

The channel attention module enhances the representation of discriminative features by modeling interdependencies among BiLSTM hidden units. It adaptively assigns importance weights to individual feature channels according to their relevance to seizure prediction, thereby emphasizing informative patterns while maintaining parameter efficiency. Unlike generic attention mechanisms that treat all features equally or attend only across time steps, channel attention explicitly models interdependencies among feature channels, allowing the network to weight feature channels according to their relevance for the task. Specifically, the BiLSTM outputs $h\in R^{N_p\times N_f}$, where N_f denotes the number of hidden units (also interpreted as feature channels), are transformed into an attention score matrix $u\in R^{N_p\times N_f}$ via a fully connected layer:

(7)$u^T=f(W_h\cdot h^T+b_h)$

where $W_h\in R^{N_f\times N_f}$ is the learnable weight matrix, $b_h\in R^{N_f}$ is the bias vector, and f denotes a nonlinear activation function. The attention scores are then normalized across all feature channels using the SoftMax function, producing the attention weight matrix $\hat{u}$:

(8)$\widehat{u_{i,j}}=\frac{e^{u_{i,j}}}{\sum _{j=1}^{N_f}{e}^{u_{i,j}}}$

where $\widehat{u_{i,j}}$ represents the attention weight for the j-th feature channel at the i-th sample point. Finally, the computed attention weights are applied to the original BiLSTM output via element-wise multiplication, yielding the refined feature representation as follows.

(9)$\hat{h}=\hat{u}\odot h$

Through this process, the channel attention module amplifies feature channels that contribute most to seizure prediction while suppressing less relevant or redundant ones. During training, the attention weights are iteratively refined via backpropagation, enabling the model to dynamically focus on the most discriminative temporal characteristics. As a result, this module complements the BiLSTM by enhancing the extraction of salient temporal features from EEG signals, supporting more effective modeling of both short-term and long-term dependencies essential for accurate seizure prediction.

2.3.3 Attention-BiLSTM

At the beginning of the proposed model, we employ dense preprocessing to provide high-quality EEG representations for subsequent effective attention learning. Then, the Attention-BiLSTM module integrates the refined temporal representations from the BiLSTM layer with the feature-wise emphasis provided by the channel attention mechanism. This integration creates optimized feature representations specifically tailored for seizure prediction. The processing pipeline begins by applying a 1D max-pooling layer to the weighted feature representation, which reduces feature dimensionality while preserving the most salient activations. The pooled features are then flattened into a vector format suitable for subsequent fully connected layers. To stabilize training and accelerate convergence, batch normalization is applied, ensuring more stable feature distributions. A dense layer performs a non-linear transformation and dimensionality reduction, projecting the normalized features into a compact representation space optimized for classification. Finally, the network concludes with an output layer that generates predictions for the two target classes: preictal and interictal states. This integrated architecture enables the model to effectively combine long-range temporal dependencies captured by the BiLSTM with the feature-channel-wise discriminative information emphasized by the attention mechanism. The complete model architecture, including output shapes and parameter settings for each layer, is summarized in Table 2.

2.4 Network training and evaluation

We implemented the proposed approach using the TensorFlow framework with the Keras API. Model training was conducted on a computing system equipped with an Intel Core i5-13600KF CPU and an NVIDIA GeForce RTX 4090 GPU. We optimized the model parameters using the Adam optimizer with a cross-entropy loss function. To prevent overfitting, we employed an early stopping strategy and retained the model weights from the epoch with the lowest validation loss for final evaluation. For performance assessment, we evaluated accuracy, sensitivity, specificity, and area under the curve (AUC) using subject-specific ten-fold cross-validation. For each subject, the preictal and interictal data were first divided into 10 folds separately and then combined for cross-validation. We report the mean and standard deviation of these metrics across 21 patients. The evaluation metrics were calculated as follows:

(10)$\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$

(11)$\text{Sensitivity}=\frac{TP}{TP+FN}$

(12)$\text{Specificity}=\frac{TN}{TN+FP}$

where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. We used paired t-tests to assess the statistical significance of performance differences between compared methods.

3. Results

Table 3 presents the stratified classification performance on the CHB-MIT dataset. Mean represents the average value, and SD represents the standard deviation. We conducted a stratified analysis by dividing patients into high- and low-seizure-frequency groups based on the average number of seizures in the dataset. The results indicate that the high-seizure-frequency group achieves higher average accuracy and sensitivity, whereas the low-seizure-frequency group exhibits relatively higher specificity and AUC, suggesting that seizure frequency influences different aspects of classification performance.

Across 21 patients, the model achieved an average accuracy of 94.77%, sensitivity of 94.58%, specificity of 94.97%, and AUC of 98.38%. The model exhibited particularly strong performance for patient chb15, reaching 98.31% accuracy, 99.04% sensitivity, 99.58% specificity, and 99.88% AUC. Relatively inferior performance observed in a small number of patients (e.g., chb06, chb10, chb14, and chb21) was not attributed to channel-related factors, as no channel selection was performed and an identical set of EEG channels was used for all patients. Instead, this variability is more likely due to inter-subject differences in seizure characteristics and the discriminability between preictal and interictal EEG patterns.

To further assess the effectiveness of the main component, we performed ablation studies involving BiLSTM (without dense1 + channel attention), BiLSTM + Dense (without channel attention), BiLSTM + Attention (without dense1), and the proposed model, as illustrated in Figure 2. Table 4 presents the results under different ablation conditions. Specifically, paired t-tests revealed that the proposed model showed significant enhancements over BiLSTM on accuracy (p = 0.0189 < 0.05), specificity (p = 0.0005 < 0.001), and AUC (p = 0.0152 < 0.05). The proposed model achieved significant improvements over BiLSTM + Dense across all metrics, including accuracy (p = 0.0002 < 0.001), sensitivity (p = 0.0071 < 0.01), specificity (p = 0.0004 < 0.001), and AUC (p = 0.0056 < 0.01). Similarly, the proposed model significantly outperformed BiLSTM + Attention in terms of accuracy (p = 0.0020 < 0.01), sensitivity (p = 0.0063 < 0.01), specificity (p = 0.0268 < 0.05), and AUC (p = 0.0228 < 0.05). Overall, these comparative results demonstrate the effectiveness of the main components in the network design.

Display Full Size

Figure 2. Classification performance comparison between BiLSTM, BiLSTM + Dense, BiLSTM + Attention, and the proposed model. Results represent averages across patients. Note that * represents p < 0.05, ** represents p < 0.01, *** represents p < 0.001 according to paired t-tests. Error bars indicate standard errors. BiLSTM: bidirectional long short-term memory.

Table 5 provides a comprehensive comparison between our method and recently published seizure prediction approaches using CNNs on the CHB-MIT dataset^{[18,19,26,34-36]}. While no single method consistently outperforms all others across every metric, our Attention-BiLSTM achieves the highest values in accuracy, specificity, and AUC, demonstrating its competitive advantage in seizure prediction tasks.

3.1 Visualization analysis

To gain deeper insights into the model’s behavior, we performed visualization analysis focusing on the attention weight distribution learned by the channel attention module. Figure 3 illustrates the evolution of feature selection across 512 hidden units.

Display Full Size

Figure 3. Analysis of attention weight distributions. (a) Initial distribution at the beginning of training; (b) Optimized distribution after training.

Figure 3a shows the initial attention weight distribution at the beginning of training, which appears nearly uniform, indicating no inherent preference for specific feature channels. In contrast, Figure 3b displays the optimized attention weight distribution after training, revealing a more heterogeneous pattern where the model selectively emphasizes discriminative feature channels while suppressing less informative ones. This contrast demonstrates the channel attention module’s capability to enhance feature representation for seizure prediction.

We further employed t-distributed stochastic neighbor embedding (t-SNE) to visualize the feature distribution learned by different model components, as shown in Figure 4. Each point represents an EEG sample, with orange indicating preictal and blue indicating interictal states. Figure 4a displays the embeddings derived from raw EEG samples, showing substantial overlap between the two classes. Figure 4b presents the feature embeddings learned by the BiLSTM layer, which exhibit more organized clustering with tighter intra-class grouping and increased inter-class separation. Figure 4c shows the embeddings from the complete Attention-BiLSTM, demonstrating further improved separation between preictal and interictal samples, with more compact and distinct clusters. These visualizations collectively illustrate the progressive enhancement in class separability achieved by our approach.

Display Full Size

Figure 4. t-SNE visualization of feature embeddings: (a) embeddings derived from raw EEG samples; (b) Embeddings learned by BiLSTM layer; (c) Embeddings learned by Attention-BiLSTM. t-SNE: t-distributed stochastic neighbor embedding; EEG: electroencephalogram; BiLSTM: bidirectional long short-term memory.

3.2 Complexity analysis

We further present the parameter count, floating-point operations per second (FLOPs), memory usage, training time, and inference latency of the proposed model, evaluated in Python on a hardware platform equipped with an Intel(R) Xeon(R) Platinum 8352V CPU@2.10 GHz and 24.0 GB RAM. As summarized in Table 6, the proposed model contains 2.47 M parameters. The average training time per batch is 107.2 ms. The inference latency is 39.7 ms per batch and 0.155 ms per sample. The per-sample inference latency is shorter than the duration of the input EEG segment (3 s). In addition, the computational cost of the proposed model is 416.55 M FLOPs. For reference, the classic MobileNet network deployed on mobile devices has 569 M FLOPs^[37].

4. Discussion

4.1 Temporal neural mechanism of preictal EEG

Our study introduces a BiLSTM-enhanced architecture designed to capture temporal dependencies in EEG signals, leveraging the progressive dynamics of preictal EEG as biomarkers for seizure prediction. Substantial evidence indicates that EEG network alterations evolve across multiple timescales prior to seizure onset^[38,39]. On longer timescales, Medina et al.^[38] observed that recurrent, short-lasting (0.6 s) high-connectivity network configurations emerging hours prior to seizures, suggesting that preictal network reorganization may commence well in advance. On shorter timescales closer to onset, both degree and betweenness centrality increase in seizure onset zone channels approximately 37.0  ±  2.8 s before seizure initiation, followed by a rise in degree across all channels about 8.2  ±  2.2 s prior to onset, indicating a gradual intensification of network connectivity^[39]. Concurrently, EEG recordings reveal widespread oscillatory alterations^{[22-24,40-42]}, most commonly manifesting as LFD. Preceding LFD, EEG changes may include preictal spikes, spike-wave trains, or slow-wave complexes^[22]. Additionally, short-duration (< 100 ms) HFOs above 80 Hz have emerged as crucial preictal biomarkers over the past decade^[40-42], with certain HFO characteristics evolving within 30 minutes before seizure onset in specific individuals^[23,24]. Together, these findings demonstrate that the temporally progressive dynamics of preictal EEG provide rich, multiscale information that can significantly enhance accurate seizure prediction.

4.2 Effectiveness of the channel attention mechanism

Existing seizure prediction studies focusing on temporal dependencies, including Wang et al.^[26], Li et al.^[43], and Daoud et al.^[44], frequently overlook the presence of redundant temporal information within the extracted features. While some models achieve relatively high seizure prediction performance, they often suffer from overfitting and lack interpretability. For instance, Li et al. developed an STS-HGCN-AL model to capture spatio-temporal dynamics across different cortical rhythms, but its reliance on temporal spectral information necessitates a highly complex network architecture that compromises interpretability. To address these limitations, we incorporated a channel attention mechanism that provides targeted improvements in both feature representation and model interpretability. This mechanism adaptively emphasizes informative temporal features through a simple yet efficient architecture, effectively reducing information redundancy. Simultaneously, the explicit attention weights enhance interpretability by quantifying the importance of different temporal components. Our experimental results validate these advantages.

Evaluation on the CHB-MIT dataset demonstrates that Attention-BiLSTM achieves significant improvements over the baseline BiLSTM without channel attention across all metrics: accuracy, sensitivity, specificity, and AUC. Analysis of attention weight distributions further reveals that the Attention-BiLSTM adaptively assigns higher weights to discriminative feature channels while suppressing less relevant ones, indicating effective selective feature representation. Corresponding t-SNE visualizations show progressively improved separation between preictal and interictal samples, from raw inputs to BiLSTM features and further to Attention-BiLSTM features. Together, these findings confirm that the channel attention mechanism not only reduces information redundancy and enhances feature discriminability but also provides clearer insights into the model’s decision-making process. The improved separation between preictal and interictal samples could guide dynamic, patient-specific thresholds in real-world seizure prediction systems, providing a basis for early warning decisions. Furthermore, by capturing temporal dependencies in EEG signals, the proposed model also holds potential for extension to other biomedical signal classification tasks relying on time-resolved features, such as Alzheimer’s disease detection^[45,46], driver fatigue evaluation^[47], sleep stage detection^[48], mental state classification^[49], and heart disease monitoring^[50,51].

4.3 Limitations and future directions

Although the proposed seizure predictor achieves a promising forecasting performance, two limitations still remain in the current work. First, classification performance in this study may be overestimated due to temporal autocorrelation. Second, preictal EEG patterns are highly patient-specific, influenced by factors such as epileptogenic zone location, seizure types, and disease severity, the proposed subject-specific approach may overfit to individual patients and struggle to capture these heterogeneous preictal signatures, which makes it not directly applicable to cross-subject scenarios. To address these limitations, future work will focus on developing cross-subject adaptive or domain-generalizable frameworks, incorporating personalized feature adaptation or meta-learning strategies, evaluating the approach on larger and more diverse datasets, and investigating the effects of different signal-level normalization strategies (e.g., z-score, min–max, per-channel versus global) on model performance. These efforts aim to better capture patient-specific preictal patterns and heterogeneous preictal signatures, ultimately improving the practical applicability of seizure prediction systems.

5. Conclusion

This study proposes a BiLSTM architecture enhanced with a channel attention mechanism for seizure prediction. The proposed architecture adaptively captures key temporal features in preictal EEG signals while effectively suppressing redundant information. Comprehensive evaluation on the CHB-MIT dataset demonstrated that Attention-BiLSTM outperforms the baseline BiLSTM without channel attention, achieving an average accuracy of 94.77%, sensitivity of 94.58%, specificity of 94.97%, and an AUC of 98.38%. Beyond these quantitative improvements, an analysis of attention weights and feature distributions confirms that the model successfully directs computational focus toward the most relevant temporal features while progressively enhancing feature discriminability. These findings provide valuable insights to support the future development of scalable and generalizable seizure prediction systems.

Authors contribution

Yu H: Conceptualization, methodology, investigation, formal analysis, writing-original draft, funding acquisition.

Shan B: Conceptualization, methodology, investigation, formal analysis, writing-original draft.

Shi Q: Conceptualization, methodology, writing-original draft.

Meng J, Yi W, Huang Y, He F, Jung TP: Formal analysis, writing-review & editing.

Xu M, Ming D: Conceptualization, investigation, writing-review & editing, supervision, funding acquisition.

Conflicts of interest

The authors declare no conflicts of interest.

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Availability of data and materials

Data could be obtained from the corresponding author upon reasonable request.

Funding

This work was supported in part by the National Key Research and Development Program of China under Grant No. 2023YFF1203700, the National Natural Science Foundation of China under Grant No. 62406220, No. 62576242, No. 82330064, China Postdoctoral Science Foundation under Grant No. 2023M742604, and Postdoctoral Fellowship Program of CPSF under Grant No. GZC20231916.

Copyright

© The Author(s) 2026.