Background: Automated EEG-based seizure detection can reduce subjectivity in clinical workflows but often suffers from overfitting and unclear validation. Objectives: The proposed study will help to improve the prediction of epileptic seizures based on the application of advanced mathematical descriptors and machine learning (ML) models. It aims to help eliminate disadvantages of manual drawing and subjective evaluation of EEGs by automating the process of seizure classification, and using statistically relevant, domain-combining features. Methods: The paper involved the use of EEG data in Bonn Dataset, where seven major features were extracted which are Detrended Fluctuation Analysis (DFA), Hjorth Fractal Dimension (HFD), SVD Entropy, Spectral Entropy, Fisher Information, Petrosian Fractal Dimension (PFD), and mean amplitude. To extract the frequency sub-bands, Daubechies and Biorthogonal wavelets were used to decompose a Wavelet (DWT). Classification of such features was done with the help of several ML algorithms and Artificial Neural Network (ANN) trained through Levenberg-Marquardt algorithm. To mitigate optimism bias, we adopt nested stratified cross-validation (outer 10-fold; inner 5-fold) with blocked splits by original record, plus a held-out test set. Metrics include accuracy, precision, recall, F1-score, ROC-AUC, MCC, and 95% CIs; model calibration (Brier score) and permutation-based feature importance are reported. Findings: ANN model demonstrated the accuracy of classification of 99%, where Random Forest and XGBoost demonstrated highest performance equally (99.3 % and 99.1%, respectively). The assessment on confusion matrix, ROC curve as well as F1-score showed that the proposed models performance exceeded the former state-of-the-art methods by 3-5 percent. The Kruskal-Wallis was used to remove redundant features allowing minimization of overfitting and enhancing generalization across classes of patients. Novelty: The proposed work integrates multi-domain statistical and nonlinear EEG features with validated hybrid deep convolutional neural network-based classifiers. The proposed work provides an innovative approach to integrate multi-domain statistical and nonlinear EEG features which includes time, frequency, and nonlinear signal behavior with optimal learning structures. Compared to the previous works that use single-domain or end-to-end black-box models, the current study incorporates the interpretable and computationally feasible modules and establishes a new bar in the seizure detection task that is based on EEG signals. Interpretation: Performance remains high under stringent validation, indicating low overfitting. The feature-engineered pipeline is computationally efficient and interpretable. Limitations: Bonn lacks subject identifiers; therefore cross-dataset evaluation is recommended in future work. Keywords: Epileptic Seizure Prediction, EEG Analysis, Machine Learning, Artificial Neural Network, Biomedical Signal Processing
Chaturvedi et al. (Sat,) studied this question.