Epilepsy is a chronic disease which affects over 40 million people worldwide. The majority of these people are able to keep the symptoms under control; however, some still suffer with frequent seizures. As such, methods for detecting the occurrences of these episodes are required. The main source of data for detecting seizures is the electroencephalogram (EEG) signal. Machine Learning algorithms, for instance, can be applied to these signals to perform such a task. Moreover, Deep Learning techniques, such as Convolutional Neural Networks, have been proven to be effective to the problem with some results achieving an accuracy of 93% or more. Nevertheless, there is little exploration towards applying these types of networks in an embedded system environment and even less with the aim of building a wearable device, which could forewarn a patient of impeding episodes. This article presents the steps taken to evaluate an architecture that can be applied to intra-patient data considering a low resource environment. We consider the application of said architecture in an embedded platform with estimates for time and memory consumption to build a wearable device. An architectural overview of such a device is also discussed.
Paper
Full text
Using deep-learning to automatically detect epileptic seizure in a wearable device
Semantic Scholar · Medicine · 2020
Abstract
Epilepsy is a chronic disease which affects over 40 million people worldwide. The majority of these people are able to keep the symptoms under control; however, some still suffer with frequent seizures. As such, methods for detecting the occurrences of these episodes are required. The main source of data for detecting seizures is the electroencephalogram (EEG) signal. Machine Learning algorithms, for instance, can be applied to these signals to perform such a task. Moreover, Deep Learning techniques, such as Convolutional Neural Networks, have been proven to be effective to the problem with some results achieving an accuracy of 93% or more. Nevertheless, there is little exploration towards applying these types of networks in an embedded system environment and even less with the aim of building a wearable device, which could forewarn a patient of impeding episodes. This article presents the steps taken to evaluate an architecture that can be applied to intra-patient data considering a low resource environment. We consider the application of said architecture in an embedded platform with estimates for time and memory consumption to build a wearable device. An architectural overview of such a device is also discussed.