Seizure Detection with 2D Spectrogram Using CNN and SVM Integration

Complex neurological disorders such as epilepsy are often diagnosed by the detailed analysis of electroencephalogram (EEG) signals. In this work, we investigate seizure classification using the 2DCNN-SVM model on EEG data. Our comprehensive evaluation, centered on EEG signals from the Bonn dataset, incorporates essential preprocessing steps like baseline correction and Short-Time Fourier Transform (STFT) analysis. EEG data can be thoroughly examined and analyzed in both the frequency and time domains via the STFT. It generates a representation based on spectrograms that have significant time-frequency details. This representation becomes the foundation for effective feature extraction using a 2D Convolutional Neural Network (2D CNN), while seizure detection is orchestrated through Support Vector Machines (SVM). Our model demonstrates its accuracy and versatility by successfully handling two-class (A vs E, B vs E, $C$ vs E, D against E) and four-class (ABCD Vs E) classification problems. After a comprehensive comparative performance analysis, the model's performance in terms of accuracy and sensitivity shows the ability to identify seizures within EEG signals. This study contributes to epilepsy diagnosis and EEG-based seizure detection systems.

I. Introduction

The recurrent neurological condition of the brain known as epilepsy affects about 65 million people worldwide. It occurs due to abnormal electrical impulses in the brain that have the potential to produce seizures. It is reported that 1 in 26 people suffers from epilepsy [1]. The symptoms of seizures, which include convulsions, hallucinations, and attention shortfalls, impact daily activities [2]. Thus, there is a great need to create a reliable approach for classifying seizures to help epileptic patients for emergency medical attention. An electroencephalogram (EEG) is a device used to monitor and record brain electrical activity. It is frequently used to identify epilepsy [3]. However, it is a challenging and time-consuming operation to manually analyze EEG signals. For this reason, accurate epilepsy diagnosis and efficient clinical screening depend on automated EEG signal analysis. The manual analysis of EEG data used in current diagnostic techniques takes a long time and specialized expertise. Automated and effective seizure detection technologies are therefore urgently required. This need is driven by the potential to improve diagnostic processes, enable timely intervention, and personalize treatment approaches, significantly enhancing patient care and outcomes. Automated EEG data processing to identify epileptic episodes has been the subject of significant study in the past several years. The analysis of these results was carried out using the EEG dataset from the University of Bonn, a commonly used standard database for identifying seizures [4]. Research on detecting epileptic seizures through EEG primarily falls into two groups: traditional approaches and deep learning methods. Conventional methods typically involve applying classification algorithms after features are extracted from unprocessed EEG data. The effectiveness of these conventional methods substantially depends on the choice of classifiers and carefully constructed feature extraction methods. Early detection of these seizures can help patients seek treatment on time and enable the immediate use of preventive measures [5]. A wide range of machine learning methods for seizure detection has been applied, encompassing statistical, temporal, frequency, time-frequency domain, and nonlinear characteristics for comprehensive analysis [6].