Abstract
Epilepsy is a neurologic condition characterized by recurring seizures resulting from aberrant brain activity. It is crucial to promptly and precisely detect epileptic seizures to ensure efficient treatment. The gold standard electroencephalography (EEG) accurately records the brain's electrical activity in real time. The intent of this study is to precisely detect epileptic episodes by leveraging machine learning and deep learning algorithms on EEG inputs. The proposed approach aims to evaluate the feasibility of developing a novel technique that utilizes the Hurst exponent to identify EEG signal properties that could be crucial for classification. The idea posits that the prolonged duration of EEG in epileptic patients and those who are not experiencing seizures can differentiate between the two groups. To achieve this, we analyzed the long-term memory characteristics of EEG by employing time-dependent Hurst analysis. Together, the Hurst exponent and the Daubechies 4 discrete wavelet transformation constitute the basis of this unique feature extraction. We utilize the ANOVA test and random forest regression as feature selection techniques. Our approach creates and evaluates support vector machine, random forest classifier, and long short-term memory network machine learning models to classify seizures using EEG inputs. The highlight of our research approach is that it examines the efficacy of the aforementioned models in classifying seizures utilizing single-channel EEG with minimally handcrafted features. The random forest classifier outperforms other options, with an accuracy of 97% and a sensitivity of 97.20%. Additionally, the proposed model's capacity to generalize unobserved data is evaluated on the CHB-MIT scalp EEG database, showing remarkable outcomes. Since this framework is computationally efficient, it can be implemented on edge hardware. This strategy can redefine epilepsy diagnoses and hence provide individualized regimens and improve patient outcomes.