Publication

Abstract : Electroencephalography (EEG) analysis is vital for diagnosing neurological disorders and understanding brain function, yet traditional methods often struggle with the high dimensionality and inherent noise in EEG signals. This project introduces an integrated AI pipeline that preprocesses EEG data (originally in Parquet format) by extracting statistical features (mean, variance, kurtosis, skewness) and frequency-domain characteristics (delta, theta, alpha, beta, gamma power) to reduce complexity and enhance interpretability. Techniques such as SMOTE and PCA were employed to address class imbalance and optimize feature space quality. We evaluated several machine learning models—Random Forest, XGBoost, CatBoost, Light-GBM, SVM, and KNN—individually and in sequential stacking configurations. The optimal stacking combination of Random Forest and LightGBM achieved 85% classification accuracy, outperforming all individual models. This streamlined pipeline demonstrates the effectiveness of model stacking in improving EEG classification and holds promise for advancing diagnostic and cognitive neuroscience applications.