This material is available only on bibliotequer
24 May 2021 
Claude 
Project overview
Epilepsy is a chronic disorder that causes unprovoked, recurrent seizures. It is the second most common neurological disease. Approximately, 2.4 million people of all ages are diagnosed with epilepsy annually [9]. The widely used clinical technique for epilepsy diagnosis is using Electroencephalogram (EEG) records. EGG signals correspond to the electrical activity of the brain. These signals are analyzed by neurologists to detect the disease. Visual inspection of hours or days of EEG data is time-consuming, laborious and require the presence of an expert. Therefore, a lot of research work has been carried out to develop a fast robust computer-aided diagnosis system that detects seizure automatically using EEG signals. Many researchers have proposed different methods to detect the epileptic and the non-epileptic signals using classification algorithms.
Problem statement
Various approaches have been presented to detect epileptic seizure. Some focus on classifying normal and ictal (during seizure) EEG signals, other worked on identifying normal, interictal (between seizure) and ictal cases. Most of the proposed approaches consist of extracting and selecting features from EEG signals, then using different machine learning algorithms. The feature extraction is usually based on traditional signal processing in the time-domain, frequency-domain or a combination of two domains. These hand-design features are sensitive to variation in seizure patterns because the EGG data are non-stationary[6]. The presence of noise and artifact in data influence also the system performance: A study[2] shows that detecting seizures from noisy EEG drop of 10% the system accuracy.
To avoid these problems, researchers start recently to use Deep learning(DL) approaches for epileptic seizure detection. Deep learning algorithms such as convolutional neural network(CNN) has been successfully used in timeseries classification[7][5]. In fact, Deep learning automatically encodes features, which are not data depend and adapt to internal structure of the data. Moreover, features extracted using the DL models have shown to be more discriminative and robust than hand-designed feature[7]. Even though the deep learning may take more time in the training phase than other methods, it is faster in the test phase, which makes DL solutions suitable for real-time clinical setting[1]. In this study, we are going to focus on Deep Learning approaches.
Project goals
In this project, we aim to use deep neural networks to build fast, accurate and robust to noise and artifacts models to detect epileptic seizure using EEG signals. The first model is to classify the normal and seizure EEGs pattern. The second is to identify normal, interictal (between
A second study proposes a pyramidal one-dimensional convolutional(P- 1D-CNN) neural network model[8]. This method claims to involves 61% fewer parameters compared to standard CNN. Another method uses recurrent neural networks(RNN), in particular the Long-short-term-memory (LSTM) neural network to learn a high-level representation of data that is fed into softmax function for classification[6]. Our goal in this project is to compare the performance and the noise impact on these methods. We plan also to try some data augmentation techniques and use bagging algorithm to improve the model’s accuracy.
Methodology
The first step in this project is the literature review to have a good understanding of the EEG signal’s properties and adequate proposed neural networks methods. Then, we will examine and preprocess our data to obtain non-overlapping epochs for training. After that, we will implement different architectures such as CNN, LMTS, P-1D-CNN.., train, validate and evaluate the models to select the most accurate one for both cases: the binary and multi-classification of EEGs patterns. The second part of the project is to improve the performance of these algorithms and try to generate synthetic noise and add them to the clean EEG data[6] to test our model’s robustness. Having good skills in devolving with python, we choose to work with Pytorch library. We will use the public Database collected by Andrzejak et al. at Bonn 2University [4].