Robust Epileptic Diagnosis Using EMD Extractor and Voting-Hybrid with Optimization algorithms

The neurological disorder epilepsy affects millions of people worldwide, so early, accurate diagnosis remains essential for effective treatment and improved quality of life. The diagnosis of epileptic seizures through EEG signals faces significant challenges because brain activity demonstrates non-stationary characteristics and high dimensionality. Current diagnostic approaches rely on fixed feature extraction techniques and single optimization methods, yet these methods do not perform well across different patients. In our work, we will present an automated epileptic diagnosis system that starts by extracting statistical and nonlinear features from different combinations of the Bonn EEG signals dataset through Empirical Mode Decomposition (EMD). Then we will test four optimization methods, like Snake Optimization (SOA), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO), and conduct independent feature selection for the most relevant features from these three methods. Additionally, a voting-based decision system counts feature frequencies, identifying features that appear in two or more of the previous three algorithms. Finally, we test three classifiers, such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Random Forest (RF), to classify the signals and give the final accuracy. The proposed system achieved high performance across all classifiers. Extensive experiments conducted on multiple binary and multiclass combinations of the Bonn EEG dataset demonstrate the effectiveness of the proposed framework. When combined with PSO, the KNN classifier achieves the highest overall performance, reaching an average accuracy of 96.15%, and a specificity of 97.61%. The Random Forest classifier also exhibits strong, stable performance, particularly when optimized using the GA, achieving average accuracies of 92.94%, 93.18%, and 96.98% for sensitivity, specificity, and overall accuracy, respectively. SVM shows comparatively lower performance, with PSO yielding average accuracies of 87.56% and 93.56% for specificity. Although the SOA delivers competitive results across several binary and moderately complex cases, its performance declines in higher-dimensional multiclass scenarios. The proposed voting-based hybrid metaheuristic selector improves feature selection robustness in certain configurations but demonstrates inconsistent performance in complex classification tasks.