Applying machine learning techniques to brain electroencephalogram signals to forecast epileptic seizures

Epileptic seizure forecasting could greatly improve the quality of life of people suffering from epilepsy. Modern forecasting systems leverage Artificial Intelligence (AI) techniques to automatically analyze neurophysiological data, such as the electroencephalogram (EEG), in order to predict upcoming epileptic events. However, most of the research in this field has evaluated the robustness of automated forecasting methods through randomized cross-validation techniques, while clinical applications would require much more stringent validation based on patient-independent testing. In this research work, we first systematically reviewed studies that focus on human scalp EEG signals with cross-patient validation methods. Then, we used two long-term continuous multichannel EEG datasets (CHB-MIT and a private dataset) to investigate the use of AI techniques and obtain experimental results. We first reported methodological issues in the evaluation of forecasting models with a drop in accuracy during a transition to more rigorous validation methods, from approximately 80% to 50% (chance level). Afterwards, we demonstrated the superior performance of deep learning algorithms compared to more traditional machine learning approaches. Our investigations also indicated that giving the recordings from all electrodes as input allows us to exploit useful channel correlations to learn more robust predictive features, compared to treating each channel independently. Eventually, we proposed a novel calibration method to fine-tune deep learning classifiers, which significantly improved forecasting models even in independent patients who have never been seen during the training phase. Although the development of an automated forecasting system remains a challenging task, we conclude that leveraging a large cohort of data and applying more realistic validation methods could pave the way for promising results in future studies.