Interpretable EEG-Based Seizure Prediction Using a Hybrid Wavelet CNN-LSTM Model with SHAP

Epilepsy affects nearly 50 million people worldwide, with seizures that are often unpredictable and disruptive. To improve detection and prediction, various ML and DL techniques have been applied to EEG signals. Vieira et al. [11] proposed an explainable AI model using simple classifiers and selected EEG features, achieving over 95% accuracy, but lacking temporal modeling due to the absence of deep learning. Chowdhury and Chowdhury [12] used a quantum machine learning model with MRI and Layer-wise Relevance Propagation (LRP), showing strong performance but limited real-time usability due to high computational cost. Our proposed CNN-LSTM model with SHAP improves both accuracy and interpretability using EEG, optimized for real-time clinical application.

Zhang et al. [13] introduced a DNN with adversarial training and attention mechanisms to improve generalization, though it increased model complexity. Our method achieves similar robustness without adversarial overhead by combining wavelet-based spatial and LSTM-based temporal features. Lo Giudice et al. [14] developed a CNN to differentiate epileptic and psychogenic seizures, but limited data and lack of temporal analysis reduced generalizability. We address this through LSTM layers and larger dataset validation. Al-Hussaini and Mitchell [15] presented SeizFt using wearable EEG, data augmentation, and CatBoost, improving sensitivity but struggling with diverse seizure types. Our model enhances robustness by integrating DWT, CNN, and LSTM for more comprehensive and interpretable predictions.

To capture spatial, frequency, and temporal EEG dynamics, a hybrid model was developed using DWT, CNN, LSTM networks, and SHAP. DWT decomposes EEG signals into sub-bands for frequency-domain features [21], which are processed by CNN layers for spatial learning and LSTM layers for temporal pattern recognition [22]. A dense layer handles binary classification, using ReLU activation, max-pooling, and residual connections for deeper learning [23]. SHAP provides global and local interpretability, revealing feature contributions to predictions (Figure 1).

While the model performed well on the clean Bonn EEG dataset, its limited complexity restricts generalizability [24]. The early plateau in validation loss suggests overfitting. Future work includes testing on more diverse datasets like CHB-MIT [25] and UBMC [17], and applying regularization (e.g., dropout, early stopping). The workflow (Figure 1) includes EEG acquisition, normalization, model training, evaluation, and SHAP analysis. As illustrated in Figure 2, the architecture starts with wavelet decomposition (cA1, cD1–cD3), followed by CNN-LSTM feature extraction, classification, and interpretability. This integrated design boosts predictive accuracy, transparency, and clinical relevance for early seizure prediction.

1.png

Figure 1. Proposed methodology for Hybrid Wavelet CNN-LSTM-SHAP

2.png

Figure 2. Architecture of the proposed Hybrid Wavelet CNN-LSTM model

4.1 Data collection

The EEG dataset used in this study consists of 500 recordings, grouped into five subsets of 100 files each, where each file captures 23.6 seconds of brain activity, represented by 4,097 time-series data points [24]. These recordings were segmented into 23 segments per file, each with 178 data points corresponding to one-second EEG intervals. This results in a matrix of 11,500 rows (segments) and 178 columns (features), with the 179th column representing the class label $y \in \{1.2 .3 .4 .5\}$. The class labels denote specific brain states: Class 1 (epileptic seizure), Class 2 (tumor region), Class 3 (healthy region in tumor patients), Class 4 (eyes closed), and Class 5 (eyes open). Classes 2 to 5 represent non-epileptic activity, and many studies treat the dataset as a binary classification task: seizure (Class 1) vs. non-seizure (Classes 2–5).

The dataset was split randomly into training (70–80%), validation (10–15%), and testing (10–15%) sets. This stratification ensured fair evaluation of model generalization. The training set was used for model learning, the validation set for tuning hyperparameters, and the testing set for final performance assessment on unseen data.

4.2 Data pre-processing and balancing

Data pre-processing enhances learning efficiency by modifying input features to a standardized scale. In this study, Min-Max normalization was applied to rescale EEG signals to a range of 0 to 1. This approach reduces computational complexity, accelerates convergence, and preserves relative amplitude variations critical for detecting seizure-related anomalies. Unlike Z-score normalization, which may obscure these patterns, Min-Max scaling ensures that all features contribute proportionally to the model [5, 11]. The normalization is defined by:

$\begin{aligned} u_0=\left(\left(y_0-y_{0_{\text {min}}}\right)\right. & (\max -\min)) /\left(\left(y_{0_{\text {max}}}-y_{0_{\text {min}}}\right)+\min \right)\end{aligned}$          (1)

where, u₀ is the normalized value, y₀ is the original input, and (min, max) represent the target range.

Although the CHB-MIT dataset is commonly used for real-world clinical evaluations, we employed the Bonn EEG dataset due to its clean, balanced structure and high signal quality, which make it suitable for benchmarking deep learning models. To address class imbalance—especially the underrepresentation of seizure cases (Class 1)—the Synthetic Minority Over-sampling Technique (SMOTE) was applied. As shown in Table 1 (before SMOTE and after SMOTE), this technique balanced the class distribution, enhancing recall and F1-score for seizure detection while maintaining high precision across all classes. This data preparation pipeline supports the robustness and generalizability of the proposed Wavelet CNN-LSTM-SHAP model and establishes a foundation for future testing on more complex datasets like CHB-MIT [26, 27].

SMOTE (Synthetic Minority Over-sampling Technique) is a popular technique for balancing imbalanced data by creating synthetic samples for minority classes. In the pre-processing step, SMOTE selects a sample from a minority class and works to find its k-nearest neighbours. For each selected model, an artificial model is generated by interpolation between the model and its neighbours. For each selected model, an artificial model is generated by interpolation between the model and its neighbours. The equation for the artificial sample x_n is given by:

$x_n=x_i+\lambda \times\left(x_{n n}-x_i\right)$

where, $x_i$ is a subclass sample, $x_n$ is one of its nearest neighbours k, and λ is a random number ranging from 0 to 1. This process is repeated until equilibrium is reached, contributing to machine learning modelling variety is effective. More classes are adequately represented during training. 

The Bonn EEG dataset was selected for its clean signals, balanced class labels, and controlled conditions—ideal for validating deep learning models. In contrast, the CHB-MIT dataset offers more clinical realism but includes high noise, patient variability, and temporal imbalance, requiring extensive preprocessing and personalized models [24, 25]. Thus, the proposed method establishes a strong baseline for future application to more complex EEG datasets. 

Table 1. Comparison of model performance before and after SMOTE balancing

4.3 Feature extraction and classification using Hybrid Wavelet CNN-LSTM-SHAP

The DWT is a mathematical technique that decomposes signals into time-frequency components, offering advantages over traditional Fourier transforms by capturing both frequency and location information. DWT is particularly effective for multi-scale analysis of complex sequences, making it suitable for EEG signal processing.

Using wavelet decomposition, the EEG signal is split into two parts: approximation coefficients (low-frequency components) and detail coefficients (high-frequency components). The approximation captures the primary structure of the signal, while the detail contains transient or noisy information. This allows for effective noise reduction by isolating and discarding high-frequency fluctuations, improving feature quality for downstream analysis using the hybrid CNN-LSTM-SHAP model.

To compute a set of wavelets $\psi^{\prime}{ }_{j, k}(t)$ and binary scalefunctions $\varphi^{\prime}{ }_{j, k}(t)$ for a given wavelet's mother function $\psi^{\prime}(t)$ and its associated scaling function $\phi^{\prime}(t)$, use the Eqs. (2) and (3):

$\psi_{j, k}^{\prime}(t)=2^{\frac{j}{2}} \psi^{\prime}\left(2^j t-k\right)$           (2)

$\varphi_{j, k}^{\prime}(t)=2^{\frac{j}{2}} \varphi^{\prime}\left(2^j t-k\right)$           (3)

in where $t, j$, and $k$ signify the time indices, scaling variables, and translating variables. The initial sequences $o s^{\prime}(t)$ may be represented as shown in Eq. (4):

$\begin{gathered}o s^{\prime}(t)=\sum_{k=1}^n c_{j, k}^{\prime} \varphi_{j, k}^{\prime}(t)+ \sum_{j=1}^J \sum_{k=1}^n d_{j, k}^{\prime} \psi_{j, k}^{\prime}(t)\end{gathered}$          (4)

In this equation, $c^{\prime}{ }_{j, k}$ represents the approximate coefficient for scale $j$ and location $k, d^{\prime}{ }_{j, k}$ represents the comprehensive coefficients at scale $j$ and locations $k, n$ is the initial sequence size, and $J$ is the breakdown levels. According to the rapid DWT, the approximation sequence and the complete sequence within a specific WD level may be derived using numerous low and high-pass filters.

The proposed Hybrid Wavelet CNN-LSTM model (Figure 2) combines DWT, CNN, and LSTM to extract spatial, spectral, and temporal features from EEG signals for early seizure prediction. DWT, using the Daubechies 4 (db4) wavelet, is selected for its effective time-frequency localization and proven utility in biomedical signal analysis [21, 22]. The EEG signals are decomposed into three levels, yielding one approximation (cA3) and three detail coefficients (cD3, cD2, cD1) that preserve essential signal characteristics linked to seizure onset. These coefficients are then passed into a CNN with two convolutional layers (32 and 64 filters, 3×3 kernels), followed by ReLU activations, batch normalization, dropout (rate = 0.25), and 2×2 max-pooling. Residual connections are added to retain low-level spatial features and improve training stability [23].

The CNN output is flattened and passed to a stacked LSTM network for modeling temporal dependencies. The first LSTM layer (128 units, return_sequences=True) and the second (64 units) capture sequential patterns in the EEG. A final dense layer with a sigmoid activation performs binary classification (seizure vs. non-seizure). The model is trained using the Adam optimizer (learning rate = 0.001), binary cross-entropy loss, batch size of 64, and 100 epochs. Performance is measured using accuracy, precision, recall, and F1-score. Compared to Short-Time Fourier Transform (STFT), DWT offers adaptive resolution and improved localization of transient seizure activity [11]. For interpretability, SHAP quantifies each feature’s influence on predictions, providing both global and local insights to support clinical transparency and trust [18].

4.3.1 Convolutional layer

The process of convolution is characterized as a particular linear approach for extracting local patterns in temporal domains and identifying local correlations in the input sequence. Eqs. (5) and (6) define the fundamental sequence input S as well as the filter sequences FS. Vectors are shown in bold according to the standard.

$S^{\prime}=\left[s_1^{\prime}, s_2^{\prime}, s_3^{\prime}, \ldots s_L^{\prime}\right]$           (5)

$F S^{\prime}=\left[\omega_1^{\prime}, \omega_2^{\prime}, \omega_3^{\prime}, \ldots \omega_K^{\prime}\right]$         (6)

In this example, $s_i^{\prime} \in R$ represents a sequence of data points arranged by time, while $\omega_j^{\prime} \in R^{\prime(m \times 1)}$ represents the filtering vectors. $L$ the lengths of initial sequence inputs $S$, whereas $K$ is the total amount of filtering in the convolutional layers. Eq. (7) defines the convolution process as a product of a filter vector $\omega_j^{\prime}$ and a combination of vectors $s_{i: i+m-1}^{\prime}$.

$s_{i: i+m-1}^{\prime}=s_i^{\prime} \oplus s_{i+1}^{\prime} \oplus s_{i+2}^{\prime} \oplus s_{i+m-1}^{\prime}$           (7)

The combination operator is $\bigoplus$, and $s_{i: i+m-1}^{\prime}$ represents a window of $m$ continual timing steps beginning with the $i$-th step. Furthermore, the term "bias" $b \in R$ would be addressed throughout the convolution procedure. Therefore, the last calculating equation is expressed in Eq. (8).

$c_i^{\prime}=f\left(\omega_j^{\prime T} s_{i: i+m-1}^{\prime}+b\right)$         (8)

$\omega_j^{\prime T}$ is the transposed of the filtering matrix $\omega_j^{\prime}$, and $f$ is a nonlinear activation function. Furthermore, indices $i$ specifies the $i$-th timing steps, while index $j$ represents the $j$-th filters.

The incorporation of activated functions aims to improve the models' ability to acquire big and complex functions, hence increasing prediction performance. Using an effective activation function may not just accelerate convergence but additionally improve model complexity. Rectified Linear Units are utilized in simulations because they exceed other types of activation mechanisms.

An activation function called Rectified Linear Unit (ReLU) comes after each convolutional layer. Data can go from the first to final layers to residual learning blocks. Moreover, these blocks are utilized to optimize CNN loss by concatenating the retrieved features from several convolutional layers. One of the most important parts of the residual learning block is depth concatenation [23]. The feature map's depth is increased by using depth concatenation. A convolutional layer, a batch normalization layer, and a ReLU function comprise each convolutional block.

4.3.2 Pooling layer

The preceding sample only shows the specific convolution operations technique among a single filter and the input sequences. A single filter can produce one features map. Numerous filters are used in the convolutional layer to effectively extract the main properties of input data. The convolutional layer has K filters having a windows size of m, as previously assumed. In Eqs. (6) and (8), every vector $\omega_j^{\prime}$ denotes a filter, and the single value $c_i^{\prime}$ indicates the window activations.

The convolutional operation across the full sequence input is carried out by moving filtering windows from the first to the last timing step. As a result, the features map related to that filter may be represented using a vector, as shown in the Eq. (9).

$F_j^{\prime}=\left[c_1^{\prime}, c_2^{\prime}, c_3^{\prime}, \ldots c_{L-m+1}^{\prime}\right]$            (9)

The items in $F_j^{\prime}$ represent multi-windows as $\left\{s_{1: m}^{\prime}, s_{2: m}^{\prime}, \ldots s_{1-m+1: L}^{\prime}\right\}$. Index $j$ represents the $j$-th filter.

Pooling is equivalent to sub sampling since it subsects the result of a convolutional layers depending on a certain pooling size $p$. This indicates that the pooling layers may efficiently condense the length of the features map, thereby reducing the amount of model parameters. The model's max-pooling method yields the compressed vector of features $F_j^{\prime}-$compress, as shown below. Also, the max function requires a max functional over the p successive elements in the features map $F_j^{\prime}$ in Eq. (10).

$F_j^{\prime}-$ compress $=\left[h_1^{\prime}, h_2^{\prime}, h_3^{\prime}, \ldots h_{\frac{L-m}{p}+1}^{\prime}\right]$            (10)

where, $h_j^{\prime}=\max \left(c_{(j-1) p}^{\prime}, c_{(j-1) p+1}^{\prime}, \ldots, c_{j p-1}^{\prime}\right)$.

The CNN-based extractors of features can produce accurate and relevant data than the original sequencing input. Furthermore, compressing the duration of the sequences of inputs improves the capacity of future LSTM models to collect temporal information.

The CNN output, which contains spatial information gleaned from EEG, is fed into the LSTM layers. This allows the model to analyse long-term trends in EEG data and produce precise predictions by capturing temporal dependencies that are essential for seizure prediction.

The conventional neural networks structures are differentiated by fully connected among neighbouring levels, that may translate the present input into target vectors. Yet, RNN can maps the target vectors utilizing the complete past of prior inputs. RNN outperforms traditional neural networks for simulating dynamics in sequential data. Overall, RNN joins units from guided cycles and remembers previous inputs via internal states. RNN results at time intervals t−1 may have an influence on RNN results at time steps t. This permits RNN to form temporal links between the current patterns and the previous ones [22].

The consecutive vectors $X=[x(0), x(1), x(2)]$ are fed into RNN one by one based on the timing step. This is distinct from a standard feed-forward networks, where all sequencing vectors are supplied into the framework at the same time. The applicable formula can be stated as follows in Eq. (11).

$S^{\prime}(t)=\sigma\left(U \cdot x^{\prime}(t)+W \cdot S^{\prime}(t-1)+b\right)$          (11)

$y^{\prime}(t)=\sigma\left(V \cdot s^{\prime}(t)+c\right)$           (12)

The equation shows that $x^{\prime}(t)$ is the initial variables at the t time steps. $W, U$ and $V$ are weight matrices. $b$ and $c$ are biased vectors, $\sigma$ is a function of activation, and $y^{\prime}(t)$ is the result that is anticipated at a $t$ times in Eq. (12).

While RNN excels in simulating dynamics in sequential data, it could be impacted by the gradient vanishing and inflating issue during backpropagation-based training of models when analysing longer sequences. Consider the inherent shortcomings of standard RNN, its enhanced form, termed LSTM, is used in this study, as described in the following section.

The LSTM network is a kind of RNN that integrates representations learning and model construction without needing additional domain expertise. The improved LSTM architecture helps to eliminate gradient disappearance and explosive issues in standard RNN. This demonstrates that LSTM is more successful in collecting long-term connections and simulating nonlinear structures when dealing with sequential data that is longer in period.

LSTM is specifically developed to avoid the issue of gradient vanishing, allowing the connection among vectors in the shorter and longer terms to be retained. In an LSTM cell, $h^{\prime}(t)$ represents a short-term state, whereas $c^{\prime}(t)$ represents a longer-term state. The major feature of LSTM is its ability understand which should be preserved in the long term., which should be rejected, and which should be read. When the $c^{\prime}(t-1)$ point reaches the cell, it first travels via a forget gates to remove memory; next the new memory are inserted into it through an input gate; at last, a novel outcome $y^{\prime}(t)$ is produced and processed by the resultant gate. The mechanism of where new recollections originates those gates operate is demonstrated below.

(1) Forget Gate

This section describes that LSTM determines what kind of data is allowed into the memories cell. After passing through the sigmoid function, $h^{\prime}(t-1)$ and $x^{\prime}(t)$ create a value $f^{\prime}(t)$ ranging from 0 to 1. A value of 1 indicates that $h^{\prime}(t-1)$ would be completely incorporated in the cells state $c^{\prime}(t-1)$. If the current value is 0 , cell state $c^{\prime}(t-1)$ will forsake $h^{\prime}(t-1)$. The formula for this procedure is presented below in Eq. (13):

$f^{\prime}(t)=\sigma\left(W_f^{\prime} \cdot\left[h^{\prime}(t-1), x^{\prime}(t)\right]+b_f^{\prime}\right)$             (13)

in which $W_f^{\prime}$: weighted matrix, $b_f^{\prime}$: bias vectors, and $\sigma$ is the activation factor.

(2) Store Gate

This section explains that LSTM determines which types of data can be kept within the cell's state. The sigmoid function is used to transform $h^{0 \prime}(t-1)$ into a value ranging from 0 to 1. The $\tanh$ function is then used to transform $h^{0 \prime}(t-1)$ into an alternative potential value, $g^{0^{\prime}}(t)$. Finally, the two outputs described above are combined to update the prior state, as shown in Eqs. (14) and (15).

$i^{0 \prime}(t)=\sigma\left(w_i^{\prime} \cdot\left[h^{0 \prime}(t-1), x^{\prime}(t)\right]+b_i^{\prime}\right)$           (14)

$g^{0 \prime}(t)=\tanh \left(w_g^{\prime} \cdot\left[h^{0 \prime}(t-1), x^{\prime}(t)\right]+b_g^{\prime}\right)$           (15)

The prior cell state $c^{0 \prime}(t-1)$ decides which information to discard and store before creating the next cell state $c^{0 \prime}(t)$. This procedure may be expressed as follows in the Eq. (16).

$c^{0 \prime}(t)=f^{\prime}(t) \cdot c^{0 \prime}(t-1)+i^{0 \prime}(t) \cdot g^{0 \prime}(t)$          (16)

(3) Output Gate

The LSTM outputs is dependent on the modified cell state $c^{\prime}(t)$. Initially, use the sigmoid functions to create a value $o^{\prime}(t)$ to regulate output. The cell state $h^{\prime}(t)$ is generated by using $\tanh$ and the outcome of the sigmoid function $o^{\prime}(t)$. After the aforementioned process, produce $y^{\prime}(t)$, as illustrated in the two Eqs. (17) and (18).

$o^{\prime}(t)=\sigma\left(w_o^{\prime} .\left[h^{\prime}(t-1), x^{\prime}(t)\right]+b_o^{\prime}\right)$          (17)

$y^{\prime}(t)=h^{\prime}(t)=o^{\prime}(t) * \tanh \left(C^{\prime}(t)\right)$            (18)

The EEG analysis pipeline integrates signal processing and deep learning to enable effective seizure prediction. It begins with collecting raw EEG data, which is decomposed using Wavelet Transform into four frequency components: cD3 (high-frequency), cD2 (mid-frequency), cD1 (low-frequency), and cA1 (approximation). These components are then passed to a CNN-LSTM model, where the convolutional layer extracts spatial features and LSTM layers handle temporal patterns. The model processes these components sequentially, and its output is derived by summing the learned sub-sequences.

This architecture first applies Wavelet Transform to separate the EEG signal into discrete detail (cD) and approximation (cA) coefficients. These are input into the CNN-LSTM model, where CNN layers extract local patterns and LSTM layers capture sequential dependencies while filtering irrelevant information. The extracted features are combined and aggregated to form a comprehensive representation of the EEG signal. This hybrid approach effectively leverages both wavelet decomposition and deep learning for EEG analysis, making it well-suited for tasks like epileptic seizure prediction.

4.4 Explainable AI (XAI)

Interpreting ML model predictions is essential for clinical trust and decision-making. While simpler models offer transparency, they often lack the accuracy of deep learning approaches. To balance this trade-off, SHAP provide a unified method to interpret outputs across different ML models, including complex neural networks [18].

SHAP extends game theory’s Shapley values to quantify each feature's contribution to a prediction. It supports global interpretability by ranking overall feature importance and local interpretability by explaining individual outcomes. Unlike traditional techniques, SHAP delivers case-specific insights, improving transparency and trust. It is also model-agnostic, supporting both linear and non-linear architectures.

Integrated into the proposed Wavelet CNN-LSTM model, SHAP highlights the most influential EEG features in seizure prediction. It supports various explanation modes:

• Text-based (feature importance scores),
• Local (impact of small input changes),
• Representative (training data influence), and
• Visual (feature effect plots).

This study integrates SHAP into the Wavelet CNN-LSTM model to provide both global and local interpretability, enabling transparent identification of the most influential EEG features in seizure prediction through text, local, representative, and visual explanation modes.

The proposed Wavelet-CNN-LSTM model achieved 98% accuracy, 98% recall, and 98.7% precision (1798 TP, 24 FP; Figure 11), demonstrating strong potential for seizure prediction with superior accuracy and interpretability over previous methods. The pipeline comprises EEG data collection, normalization, Wavelet CNN for spatial-frequency feature extraction, and LSTM for capturing temporal dependencies. SHAP analysis enhances model transparency by highlighting key predictive features, enabling global and local interpretability. The model was evaluated using standard metrics and benchmarked against existing techniques. Although high accuracy was achieved on the Bonn EEG dataset, its artifact-free and balanced nature limits clinical generalizability [24]. The early plateau in validation loss (Figure 13) indicates potential overfitting. Despite applying dropout and SMOTE, further validation using complex datasets such as CHB-MIT [25], UBMC [17], and TUH EEG is necessary to ensure robustness across real-world conditions.

Clinically, the model shows promise with a low false positive rate (1.3%) and high recall (98%), ensuring minimal missed seizures, which is crucial for patient safety. This effective balance between sensitivity and specificity supports reliable and ethically responsible deployment. SHAP-driven insights improve clinical trust by revealing EEG features influencing predictions, helping physicians make informed treatment decisions and enabling more personalized care. While the model is computationally intensive and depends on high-quality labeled data, its ability to capture spatial, temporal, and frequency-domain patterns positions it as a viable tool for seizure prediction and broader medical time-series analysis.