Detection and Classification of Abnormality in Electroencephalogram Signals Using Deep Learning and Convolutional Neural Networks رسالة ماجستير

Abstract

The human body is combined of a group of subsystems that communicate with each other, where the brain is the control unit for these systems. The brain function through a group of small parts called neurons, groups of neurons is connected to create a neural network to perform some tasks. When these neurons have issues or give wrong signals the brain will start to confuse the body systems and that can lead to neurological problems such as Epilepsy. To understand and diagnose such issues a test called Electroencephalogram (EEG) is performed. It records the electrical activities of the brain so that each part of this record can refer to an activity or group of activities. The issue with EEG is that it produces a huge amount of data for a short recording time, and due to the complexity of the brain signals and human errors, doctors are taking lots of time to diagnose the records, and many patients are misdiagnosed. This leads to the need to have a computerized system that can reduce these problems.

Many systems were proposed in the previous years, and with the advantages of AI and Machine Learning, much research was done in this field. Some applications were created using rule-based systems, others using Multilayer Perceptron (MLPs). When Deep Learning Networks such as Convolutional Neural Networks (CNNs) starts to be popular, many applications also were طbسنlt‎ based on them it. One main challenge of EEG signal processing is that the patterns are not necessarily unique; where the same signal for different patients can mean different things, which makes it very hard to create a generic model for EEG signal processing.

In this research, a new approach is proposed where we take advantage of the CNN abilities to extract features and handle complex time-series signals, combine it with wavelet signal

decomposition along with preprocessing steps to create a robust model to analyze the EEG data

and detect the epileptic seizures. The main strength of this work is that it’s built at the patient level; since the EEG test provides a huge amount of data it’s possible to tune the model for each patient. In this work, a 1D CNN model with ConvlD, Long Short-Term Memory (LSTM), and MLPS layers was created. A global dataset for several patients who are suffering from epilepsy was usعd.‎ In this research, a comparison between our work and other regular algorithms such as Regression Tree, K-Nearest Neighbors (KNN(, Support Vector Machines )SVM), and Ensemble was done. The algorithms we selected were based on the data structure and the studies that were done in this field. The proposed preprocessing algorithm enhances the data and made it easier to analyze, whereas the proposed detection model provided much better detecting accuracy when it’s used with the processed data. Also, the regular algorithms did so based on the previous studies that were reviewed. The final average epileptic signals detection accuracy is 97.14%, with the highest accuracy of 99.2%. Originally this research was targeting the local Palestinian data, unfortunately,

such kind of data was not available and due to that, a global data set was used.