An Interpretable Classical Machine Learning Framework for EEG-Based Brain Abnormality Detection

Abstract

Although electroencephalography (EEG) provides an excellent, non-invasive method for monitoring brain activity, the complexity and density of these signals present a significant challenge for automated analysis. Therefore, this paper proposes an intelligent, machine learning-based framework that is both understandable and computationally efficient for classifying these signals and detecting Brain Abnormalities. The study adopted traditional machine learning techniques as a practical alternative to deep learning models, which typically consume enormous amounts of data and computing resources. It started with TUH recordings being processed with frequency filtering and subdivision into overlapping windows in order to deal with signal instability. The research then meticulously extracted the time and frequency domain features manually with emphasis on statistics and power density and used PCA to simplify the data and still maintain its central features. To improve the accuracy of the results and improve the data analysis process, the study used a number of machine learning algorithms. Such algorithms are Support Vector machine (SVM), random forest (RF), k-Nearest neighbors and extreme gradient boosting (XGboost). With experiments on these algorithms, the Supporting Vector Machine (SVM) proved to be the most successful with a fine balance between performance and an accuracy of 88.6 per cent. This shows that the traditional approaches, where they are backed by effective feature extraction, are still very competitive in the area, particularly because the suggested system focuses on clarity, modifiability, and low complexity, thus suit best the intelligent software and decision-making systems.