Methods for Nonconvulsive Epileptic Seizure Monitoring

In this thesis a patient-specific method is presented and evaluated for real-time NCES detection from scalp EEG applying a combination of tensor analysis and machine learning algorithms. The automated NCES detection facilitates alerting the doctors about the patient risk to evolve to NCSE. In particular, this proposal attempts to identify the NCES based on their similarity to the first NCES detected by the physician on the EEG. The analysis is performed by expanding the EEG signal in third-order tensors, with dimensions “frequency× time × channels”. The components extracted for each dimension are used as input to a binary classification problem to detect the NCES. This approach is shown to outperform several methods proposed in the literature for NCES displaying accuracy values over 98%. Yet, a decrease in performance was observed during evaluation when EEG pattern morphology underwent changes