MASc Seminar: A Multi-Tier Distributed fog-based Architecture for Early Prediction of Epileptic Seizures

Monday, January 22, 2018 2:00 pm - 2:00 pm EST (GMT -05:00)

Candidate: Huda Diab Abdulgalil

Title: A Multi-Tier Distributed fog-based Architecture for Early Prediction of Epileptic Seizures

Date: January 22, 2018

Time: 2:00pm

Place: E5-5047

Supervisor(s): Basir, Otman - Hilal, Allaa R. (Adjunct)

Abstract:

Epilepsy is the fourth most common neurological problem. With 50 million people living with epilepsy worldwide, about one in 26 people will continue experiencing recurring seizures during their lifetime. Epileptic seizures are characterized by uncontrollable movements and can cause loss of awareness. Despite the optimal use of antiepileptic medications, seizures are still difficult to control due to its sudden and unpredictable nature. Such seizures can put the lives of patients and others at risk. For example, seizure attacks while patients are driving could affect their ability to control a vehicle and could result in injuries to the patients as well as others. Notifying patients before the onset of seizures can enable them to avoid risks and minimize accidents, thus, save their lives. Early and accurate prediction of seizures can play a significant role in improving patients’ quality of life and helping doctors to administer medications through providing a historical overview of patient's condition over time.

The individual variability and the dynamic disparity in differentiating between the pre-ictal phase (a period before the onset of the seizure) and other seizures phases make the early prediction of seizure a challenging task. Although several research projects have focused on developing a reliable seizure prediction model, numerous challenges still exist and need to be addressed. Most of the existing approaches are not suitable for real-time settings, which requires bio-signals collection and analysis in real-time. Various methods were developed based on the analysis of EEG signals without considering the notification latency and computational cost to support monitoring of multiple patients. Limited approaches were designed based on the analysis of ECG signals. ECG signals can be collected using consumer wearable devices and are suitable for light- weight real-time analysis. Moreover, existing prediction methods were developed based on the analysis of seizure state and ignored the investigation of pre-ictal state. The analysis of the pre- ictal state is essential in the prediction of seizures at an early stage. Therefore, there is a crucial need to design a novel computing model for early prediction of epileptic seizures. This model would greatly assist in improving the patients' quality of lives.

This work proposes a multi-tier architecture for early prediction of seizures based on the analysis of two vital signs, namely, Electrocardiography (ECG) and Electroencephalogram (EEG) signals. The proposed architecture comprises of three tiers: (1) sensing at the first tier, (2) lightweight analysis based on ECG signals at the second tier, and (3) deep analysis based on EEG signals at the third tier. The proposed architecture is developed to leverage the potential of fog computing technology at the second tier for a real-time signal analytics and ubiquitous response. The proposed architecture can enable the early prediction of epileptic seizures, reduce the notification latency, and minimize the energy consumption on real-time data transmissions. Moreover, the proposed architecture is designed to allow for both lightweight and extensive analytics, thus make accurate and reliable decisions. The proposed lightweight model is formulated using the analysis of ECG signals to detect the pre-ictal state. The lightweight model utilizes the Least Squares Support Vector Machines (LS-SVM) classifier, while the proposed extensive analytics model analyzes EEG signals and utilizes Deep Belief Network (DBN) to provide an accurate classification of the patient’s state.

The performance of the proposed architecture is evaluated in terms of latency minimization and energy consumption in comparison with the cloud. Moreover, the performance of the proposed prediction models is evaluated using three datasets. Various performance metrics were used to investigate the prediction model performance, including: accuracy, sensitivity, specificity, and F1- Measure. The results illustrate the merits of the proposed architecture and show significant improvement in the early prediction of seizures in terms of accuracy, sensitivity, and specificity.