Feature-Based Machine Learning Ensemble for the Classification of Harmful Brain Activities

Abstract

The classification of harmful brain activities (HBA) is crucial in treating and curing many neurological illnesses. Electroencephalography (EEG) signals provide a non-invasive approach for capturing cerebral activity. Harvard Medical School supplied the biggest dataset to date for the HBAC using EEG signals, i.e., the HMS dataset. In this study, we used the HMS dataset. In the proposed model, we transformed raw EEG data into spectrograms. We cleaned, denoised, and eliminated artifacts employing Butterworth and Filtfilt filters. Spectrograms are obtained, and the Discrete Cosine Transform is produced in the subsequent step for Mel-Frequency Cepstral Coefficients (MFCC) computation. We partitioned signals into brief overlapping segments before using DCT. Each frame is transformed from the time domain to the time-frequency domain with the Hamming window method to mitigate edge effects. Filters based on the Mel scale are used. We computed the logarithm of the Mel spectrum. The DCT decreases dimensionality while preserving essential coefficients. The first MFCC derivative, Delta 1, was computed to improve model performance. It shows the change in the original features throughout time. Delta 1 computes the variance between MFCC frames. Delta 2 features are derived in a comparable manner; however, delta_1 is produced from MFCC, whereas delta_2 is derived from delta_1. We used machine learning classifiers with delta features to enhance the model's speed. We proposed an ML-based ensemble for the HBAC task and achieved an accuracy of 94.5%, which is comparable to the state-of-the-art methods in the literature.