Neurodynamic EEG Seizure Analysis Using CWT Scaleograms and Modified DenseNet

Technical Summary: Neurodynamic EEG Seizure Analysis Using CWT Scaleograms and Modified DenseNet

Problem Statement
Epilepsy, affecting approximately 50 million people globally, is characterized by abnormal neuronal discharges leading to recurrent seizures. While Electroencephalography (EEG) is the primary non-invasive tool for detection due to its high temporal resolution, analyzing these signals remains challenging. Seizure activity manifests as non-stationary time-varying oscillatory activity, abnormal synchronization, and spectral energy shifts across brain rhythms (delta to gamma bands). Traditional Computer-Aided Diagnosis (CAD) systems relying on handcrafted features (e.g., wavelet transforms, entropy, PCA) and classical classifiers (e.g., SVM, Random Forest) often suffer from poor generalization, noise sensitivity, and heavy dependency on expert knowledge. Conversely, deep learning approaches (CNNs, RNNs, Transformers) offer automated feature learning but face challenges regarding data scarcity, the need for large datasets, and a lack of interpretability for clinical deployment.

Methodology
The authors propose an explainable neurodynamic framework designed for binary seizure classification (seizure vs. non-seizure) using the Bonn dataset. The methodology follows a structured pipeline:

1. Data Preprocessing:

   • Source: 500 one-dimensional EEG segments (4097 samples each, ~23.6 seconds) from the Bonn dataset.
   • Cleaning: Signals undergo a 50 Hz IIR notch filter to remove power-line interference, followed by baseline correction (−0.5 s to −0.1 s).
   • Artifact Removal: FastICA (Independent Component Analysis) is applied to remove artifacts.
   • Filtering: A 1–30 Hz band-pass filter retains relevant EEG rhythms. Power Spectral Density (PSD) analysis confirms effective attenuation of the 50 Hz component and harmonics.

2. Neurodynamic Representation (CWT Scaleograms):

   • To address the non-stationary nature of EEG, the preprocessed 1D signals are transformed into 2D time-frequency images using the Continuous Wavelet Transform (CWT).
   • A Gaussian mother wavelet is utilized to generate scaleograms. This transformation captures transient oscillations, synchronization patterns, and energy concentrations associated with ictal (seizure) activity, preserving localized spectral-temporal patterns that 1D signals obscure.

3. Modified DenseNet Architecture:

   • The core classification model is a modified DenseNet tailored for 2D scaleogram inputs.
   • Structural Changes: Unlike standard DenseNet implementations, this variant removes fully connected (FC) layers to reduce overfitting and parameter count.
   • Mechanism: It employs a fully convolutional structure featuring convolutional upsampling and Global Average Pooling (GAP). This design retains spatial information within the scaleogram domain while ensuring efficient feature reuse through dense connectivity.
   • Training Configuration: The model uses the Adam optimizer (learning rate 0.001), batch size 32, 20 epochs, dropout (0.25), ReLU activation, and binary cross-entropy loss with a sigmoid output for binary classification.

4. Explainability (Grad-CAM):

   • To enhance clinical transparency, Gradient-weighted Class Activation Mapping (Grad-CAM) is applied to the final convolutional layer. This generates heatmaps visualizing the specific time-frequency regions that contribute most to the model's decision, verifying that the model focuses on clinically relevant spectral bursts.

Key Contributions

• Novel Representation: The study introduces a CWT-based 2D scaleogram representation specifically for EEG seizure analysis, transforming 1D neurodynamic signals into a format suitable for efficient convolutional processing.
• Architectural Innovation: A modified DenseNet architecture is proposed that eliminates fully connected layers in favor of a fully convolutional design with upsampling and GAP. This modification aims to improve generalization on small datasets and reduce overfitting.
• Explainable AI (XAI): The integration of Grad-CAM provides visual interpretability, allowing clinicians to verify that the model's decisions are based on discriminative spectral-temporal regions rather than noise.
• Comprehensive Benchmarking: The proposed model is rigorously compared against a wide range of state-of-the-art architectures, including VGG16, VGG19, ResNet-50, standard CNNs, BiLSTM, EfficientNetB0, and Transformers.

Results
The model was evaluated using an 80:20 train-test split with stratified 5-fold cross-validation and class-weighted loss. The Modified DenseNet achieved the following performance metrics:

• Accuracy: 95.1% (95% CI: 92.4%–97.8%)
• Precision: 97.0%
• Recall (Sensitivity): 93.5%
• Specificity: 95.0%
• F1-Score: 95.2%
• AUC: 0.97

Comparative Performance:
The proposed model outperformed all baseline models by a margin of at least 4.1% in accuracy.

• Transformers achieved 90.2% accuracy.
• EfficientNetB0 achieved 89.8% accuracy.
• ResNet-50 achieved 91% accuracy.
• BiLSTM performed relatively lower (87.4%) on the scaleogram inputs.
• VGG16/VGG19 and standard CNNs scored between 84% and 86%.

Grad-CAM visualizations confirmed that the model correctly identified concentrated high-frequency bursts in seizure cases, whereas non-seizure cases showed diffuse, low activations.

Significance and Claims
The paper claims that the combination of CWT-based neurodynamic representation and a modified, fully convolutional DenseNet provides an effective, data-efficient approach for EEG seizure classification. The authors assert that this framework addresses the limitations of traditional CAD systems (noise sensitivity) and standard deep learning models (data hunger and lack of interpretability). By achieving superior accuracy on a small, single-channel dataset while providing visual explanations via Grad-CAM, the study suggests a pathway toward clinically transparent seizure detection tools.

Limitations Acknowledged by the Authors
The authors explicitly note that the study is limited by:

• Evaluation on a small, single-channel dataset (Bonn), which may restrict generalization to real-world, multi-channel clinical EEG.
• Potential loss of temporal and spatial information during the conversion to 2D scaleograms.
• Coarse interpretability provided by Grad-CAM.
• Lack of validation on multi-channel, real-time, and noisy clinical recordings.

Future work is proposed to focus on multi-channel data, transformer-based models, and graph-based neurodynamic analysis.