Epileptic Seizure Detection and Forecasting Ecosystems

Abstract

Epilepsy affects almost 1% of the global population and considerably impacts the quality of life of those patients diagnosed with the disease. Ambulatory EEG monitoring devices that can detect or predict seizures could play an important role for people with intractable epilepsy. Many outstanding studies in detecting and forecasting epileptic seizures using EEG have been developed over the past three decades. Despite this success, their implementations as part of implantable or wearable devices are still limited. To achieve high performance, many of these studies relied on handcraft feature extraction. This approach is not generalizable and requires significant modifications for each new patient. This issue greatly limits the applicability of such methods to hardware implementation. In this thesis, we propose a deep learning-based solution for generalized epileptic seizure detection and forecasting that does not require handcraft feature extraction. The method can be applied to any other patient without the need for manual feature extraction. Secondly, we optimize seizure detection and forecasting systems to reduce computational complexity and power consumption. The optimization is performed from two aspects: algorithm and input signal. In the first aspect, we propose two approaches: automatic channel selection to reduce the number of necessary EEG electrodes; Integer-Net, an integer convolutional neural network, to reduce computational complexity and required memory. In the second aspect, we investigate how sensitive seizure detection algorithms are regarding EEG's resolution. Another problem that we would like to address is the lack of labeled EEG data for epilepsy. Today the process of epileptic seizure identification and data labeling is done by neurologists, which is expensive and time-consuming. We propose an unsupervised learning approach to make use of unlabeled EEG data which is more accessible.