چکیده:
Introduction: Epileptic seizure prediction is an essential clinical goal, as unpredictable seizures severely affect patients’ safety and quality of life. Although the electroencephalogram (EEG) has traditionally been used for seizure prediction, unimodal approaches often suffer from high false-positive rates. Recent evidence suggests that electrocardiogram (ECG) features may provide complementary information. This study aimed to design and evaluate a multimodal deep learning framework that integrates EEG and ECG signals for improved seizure prediction.
Methods: Thirty patients with drug-resistant epilepsy underwent simultaneous EEG and ECG monitoring. Signals were preprocessed using band-pass filtering, independent component analysis, and normalization. Statistical, spectral, and nonlinear features were extracted, and a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model was trained using stratified 4-fold cross-validation. The classification task was defined as binary discrimination between preictal (seizure-predictive) and interictal (non-seizure or baseline) states, allowing the model to identify physiological changes preceding seizure onset. Performance metrics included accuracy, sensitivity, specificity, false-positive rate, and area under the curve (AUC).
Results: The multimodal CNN–LSTM system achieved 92.6 ± 0.5% accuracy, 90.4 ± 0.5% sensitivity, 94.1 ± 0.3% specificity, and an AUC of 0.95 ± 0.01, with a false-positive rate of 7.2 ± 0.8%. These results significantly outperformed unimodal EEG-only and ECG-only models, which demonstrated accuracies of 84.3% and 77.5%, respectively. Statistical analysis confirmed the superiority of the multimodal approach, F(2, 6) = 8.73, p = .006.
Conclusion: Integrating EEG and ECG features in a multimodal CNN–LSTM framework enhances seizure prediction accuracy and reduces false alarms compared with unimodal models. The findings underscore the translational potential of multimodal deep learning for real-time early-warning systems and wearable applications in epilepsy management.
Methods: Thirty patients with drug-resistant epilepsy underwent simultaneous EEG and ECG monitoring. Signals were preprocessed using band-pass filtering, independent component analysis, and normalization. Statistical, spectral, and nonlinear features were extracted, and a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model was trained using stratified 4-fold cross-validation. The classification task was defined as binary discrimination between preictal (seizure-predictive) and interictal (non-seizure or baseline) states, allowing the model to identify physiological changes preceding seizure onset. Performance metrics included accuracy, sensitivity, specificity, false-positive rate, and area under the curve (AUC).
Results: The multimodal CNN–LSTM system achieved 92.6 ± 0.5% accuracy, 90.4 ± 0.5% sensitivity, 94.1 ± 0.3% specificity, and an AUC of 0.95 ± 0.01, with a false-positive rate of 7.2 ± 0.8%. These results significantly outperformed unimodal EEG-only and ECG-only models, which demonstrated accuracies of 84.3% and 77.5%, respectively. Statistical analysis confirmed the superiority of the multimodal approach, F(2, 6) = 8.73, p = .006.
Conclusion: Integrating EEG and ECG features in a multimodal CNN–LSTM framework enhances seizure prediction accuracy and reduces false alarms compared with unimodal models. The findings underscore the translational potential of multimodal deep learning for real-time early-warning systems and wearable applications in epilepsy management.
| بازنشر اطلاعات | |
 |
این مقاله تحت شرایط Creative Commons Attribution-NonCommercial 4.0 International License قابل بازنشر است. |
