Epileptic seizure prediction using artificial intelligence based on EEG and ECG signals

Epilepsy is one of the most common neurological disorders in the world, characterized by sudden-onset seizures. Clinically, the two most preferred treatment methods are surgical procedures and pharmacological therapies. However, neither of these approaches can be considered the gold standard due to their limited applicability and relatively low success rates across patients. Therefore, developing seizure prediction algorithms that can estimate seizures before they occur may enable the establishment of early warning systems. In this study, the Siena Scalp EEG Database was used for early epileptic seizure prediction. All analyses and model implementations were conducted in MATLAB. Within the scope of this thesis, machine learning and deep learning algorithms were developed based on EEG and ECG biosignals. Initially, the biosignals were preprocessed and segmented into 30-second epochs. For machine learning algorithms, 19 features were extracted from EEG and 20 from ECG signals, derived from both time and frequency domains. The five most successful models were evaluated, and average and median early prediction times were reported in minutes. Among EEG-based machine learning models, the highest performance was achieved by the Efficient Linear SVM with an accuracy of 90.3%, sensitivity of 87%, and specificity of 92%. This model also demonstrated strong early warning capability by predicting seizures approximately 18.5 minutes before onset. In ECG-based classification, the best performance was obtained using the Linear SVM model (Accuracy: 85.2%; Sensitivity: 81%; Specificity: 88%), which predicted the preictal state approximately 14.8 minutes in advance, with a median warning time of 13 minutes. In the deep learning part, convolutional neural networks (CNN) were employed. Seizure prediction algorithms were developed using architectures commonly used in the literature. Spectrogram images generated from signal epochs were utilized for 2D architectures, including ResNet-18, MobileNetV2, and EfficientNet-B0. Additionally, 1D models were developed using raw temporal signals with ResNet-1D, MobileNet-1D, and InceptionTime. Deep learning models trained with EEG data yielded higher performance compared to those trained with ECG data. The highest accuracy, 95.4%, was achieved using ResNet-18 (2D) with EEG spectrogram inputs. Among 1D temporal models, the highest accuracy was again achieved with EEG signals using ResNet-1D at 93.8%. In terms of early prediction, the median times ranged between 10.5–9.0 minutes for 2D models and 9.2–8.6 minutes for 1D models, which are considered sufficient for timely clinical intervention. It is inferred that the relatively shorter early prediction times in deep learning compared to machine learning are due to the models' tendency to provide more cautious and precise predictions.