A Machine Learning Framework for EEG-Based Epileptic Seizure Classification Using Time–Frequency Feature Analysis

Epilepsy affects approximately 50 million people worldwide, and timely, accurate seizure detection is essential for clinical management. Electroencephalogram (EEG) signals carry rich temporal and spectral information about brain dynamics, yet manual interpretation is time-consuming and expert-dependent. This paper presents a complete automated EEG classification pipeline combining signal processing feature engineering with supervised machine learning to distinguish epileptic from non-epileptic brain states. Using the UCI Epileptic Seizure Recognition Dataset (11,500 segments of approximately 1.025 s and five classes at 173.61 Hz), we extract 31 features spanning the time domain (mean, standard deviation, RMS, peak-to-peak, zero-crossing rate, skewness, kurtosis, Hjorth activity, mobility, complexity, line length, mean absolute deviation, and derivative variance) and the frequency domain (DFT magnitude statistics and Welch PSD statistics, spectral entropy, spectral centroid, spectral bandwidth, total power, EEG band powers (delta through gamma), and inter-band power ratios). Four classifiers—logistic regression (LR), random forest (RF), support vector machine (SVM), and gradient boosting (GB)—are trained and evaluated on a binary task (seizure versus non-seizure) and a five-class task. For binary classification, the SVM achieves the highest accuracy of 99.13% with an F1 score of 0.978. For multiclass classification, gradient boosting achieves 81.35% accuracy. Feature importance analysis identifies the Hjorth complexity, spectral centroid, and theta-band power as the most discriminative features. A comprehensive ablation study further confirms that the combined 31-feature representation outperforms either domain in isolation by up to 4.2 percentage points on the multiclass task. Unlike less interpretable deep learning models, the proposed signal processing pipeline is computationally efficient and provides transparent feature-based classification on this benchmark dataset. The complete pipeline is implemented in Python 3.13.15 and is fully reproducible.