Data-Driven Seizure Prediction Using EEG and ECG Signals

Drug-resistant epilepsy affects approximately 30% of the 50 million people living with epilepsy worldwide. Despite extensive research, seizure prediction remains a major challenge, with no system yet integrated into clinical practice. One of the main obstacles to progress lies in the number of a priori assumptions that researchers often introduce during the system creation process. These choices are frequently driven by practical constraints and empirical decisions rather than objective optimization and physiological data, potentially obscuring optimal solutions. This work presents a fully automated, data-driven toolbox designed to process both EEG and ECG signals while systematically optimizing key parameters to maximize final classification performance. The system is structured as an end-to-end pipeline, including signal preprocessing, feature extraction, dataset adjustments and classifier training. By minimizing manual parameter selection, the approach ensures truly data-optimized processing, enhancing both performance and reproducibility. The system leverages Neural Network Intelligence (NNI) for hyperparameter tuning, allowing for a broader and more efficient exploration of parameter combinations. The proposed system was evaluated using data from the Siena Scalp EEG Database, which includes multimodal recordings from 14 patients. These recordings were used either completely or partially, depending on the trial, with the possibility of treating each patient independently or combining them. To assign the final classification, both traditional machine learning and deep learning approaches were explored. The machine learning models were trained using feature vectors, while the deep learning models operated on structured feature matrices, preserving the temporal dependencies of the data. To prevent data leakage, strict dataset partitioning and Z-score normalization were implemented, ensuring unbiased model evaluation. Additionally, a hybrid balancing strategy was employed to address the dataset class imbalance, combining oversampling of preictal windows with undersampling of interictal windows. Different trials were conducted: both signals were tested using either machine learning or neural networks, and both patient-specific and general approaches were tried. Some promising results were obtained even when using only ECG combined with ML algorithms, in particular in the patient-specific case: the system achieved an overall accuracy of approximately 76%, with a sensitivity of 77% and a specificity of 75% in distinguishing preictal from interictal states. The percentage of time under false alarm remained below 9%, while the prediction horizon ranged from 15 to 20 minutes. In the general case, the system achieved an overall accuracy of approximately 58%, with a sensitivity of 91% and a specificity of 27% in distinguishing preictal from interictal states. The percentage of time under false alarm reached 4.3%, while the prediction horizon was 20 minutes. Some limitations are still present, mainly due to the small size of the dataset and the final metrics are still not ready for a clinical routine implementation. However, they highlight the potential of this automated, data-driven system