Detection of epileptic seizures in EEG signals: Analysis of linear time-domain features using classical machine learning techniques

Detecting epileptic seizures from EEG signals allows for earlier and more accurate diagnosis of patients with epilepsy. This study evaluates the performance of three traditional machine learning (ML) models: support vector machines (SVM), random forests (RF), and K-nearest neighbors (KNN) for the binary classification of EEG segments with or without seizures. The CHB-MIT database was used, without preprocessing, to mimic a real clinical setting. Linear temporal features were extracted in three configurations: 9 widely cited features, 21 novel features, and a combination of all 30 features. The results show that RF is the most robust and reliable classifier; with 30 features, it achieved 96.75% precision, 94.30% accuracy, 91.71% recall, 94.14% F1 score, and 96.88% specificity. These findings emphasize the importance of proper feature selection and suggest new possibilities for improving the accuracy of traditional ML models in detecting epileptic seizures from EEG signals.