Accelerometer-based Convulsive Seizure Detection using CNN

Abstract

Convulsive seizures contribute to a significant portion of the seizure-associated injuries, accidents, and sudden unexpected deaths in Epilepsy (SUDEP). An ambulatory seizure detection system may prevent such accidents and improve the quality of life. Conventional seizure detection methods require specialized approaches such as video or EEG analysis, which are frequently ineffective in non-clinical settings such as during daily activities. In recent years, a couple of wearable accelerometer-based seizure detection systems have been proposed. But the common problem these devices face is low specificity and high False Alarm Rate (FAR). In this work, we proposed an improved way to study and classify accelerometer data using Convolutional Neural Network (CNN) to detect General Tonic Clonic Seizures (GTCS), also known as Convulsive Seizures. Due to the unavailability of a dataset of accelerometer data related to seizure movements, an accelerometer-based wrist-worn data acquisition device was constructed to develop a dataset mimicking seizure-like movement. The accelerometer data were then converted to RGB images for training and testing with three different CNN architectures: DenseNet, ResNet-50, and VGG16, to determine which architecture is best suited for these types of data. Among these three, the DenseNet architecture achieved the highest accuracy of 99.2%, sensitivity of 98.4%, and specificity of 100%. Hence, an algorithm was developed based on the DenseNet model to detect convulsive seizures with a feature to tune according to the patient’s seizure type. The proposed method can be implemented to develop an ambulatory seizure monitoring device to detect seizures before accidents occur.