Effective Detection of Epileptic Seizure

Abstract

The prevalence of neurological and psychiatric disorders, like epilepsy, is increasing newlinedue to factors such as mental stress and behavioral changes. Epilepsy is characterized newlineby abnormal brain electrical activity, often monitored via electroencephalography newline(EEG) and electrocorticography (ECoG). These signals are non-stationary, non-linear, newlinenoisy, and complex, making seizure detection challenging. Traditionally, doctors newlinediagnose epilepsy by reviewing medical history and symptoms, along with visually newlineanalyzing EEG signals. However, this is time-consuming and prone to human error. newlineTo address these challenges, this research proposes two methods for detecting newlineepileptic seizures by distinguishing between focal and non-focal EEG signals. A key newlineissue with machine learning models for seizure detection is the need for large datasets, newlineas models like KNN and SVM often suffer from overfitting while using small datasets newlinefor training. When the model is unable to generalize, it is said to be overfit, especially newlinein the presence of noise. This research utilizes the Hidden Markov Model (HMM), newlinewhich requires fewer samples for training, reducing overfitting. Additionally, Human newlineLearning Optimization (HLO) algorithm is applied to select the most discriminative newlinefeatures and optimize the model further. newlineThe first model, a Human Learning Optimization (HLO)-trained Hidden Markov newlineModel (HMM), optimizes electrode selection for feature extraction, reduces model newlinecomplexity, and enhances feature discrimination. The HLO algorithm incorporates newlinehuman learning to accelerate convergence, improving performance. On the CHB-MIT newlinedataset, the model attained 99.158% accuracy, 98.086% sensitivity, and 99.155% newlinespecificity. For the Siena dataset, the model attained 94.231% accuracy, 92.075% newlinesensitivity, and 99.926% specificity, proving to be effective in detecting epileptic newlineseizures. newlineThe second, more advanced model is a deep dual adaptive Convolutional Neural newlineNetwork-Hidden Markov Model (CNN-HMM) designed to have better EEG signal newlineprocessing. It uses wavelet decomposition for noise reduction, sub-band splitting, newlineHLO for electrode selection, and features like Tunable Q-Wavelet Transform newline(TQWT) and Hjorth for feature extraction. TQWT improves low-energy component newlineextraction, outperforming Empirical Mode Decomposition (EMD). This model was newlinetested on four datasets and showed superior performance. Using TP, dataset 1 newlineachieves 99.46%, 98.48%, 99.46%, 99.90%, and 99.58%; using 10-fold, it achieves newline98.13%, 98.46%, 97.56%, 99.88%, and 99.56%. With TP, dataset 2 achieves 94.53%, newline92.37%, 99.94%, 93.11%, and 93.60%; with 10-fold, it achieves 90.84%, 91.17%, newline90.27%, 93.09%, and 93.58%. In the same way, dataset 3 achieves 94.54%, 94.68%, newline96.87%, 95.46%, and 96.45% with 10-fold, and 94.48%, 94.62%, 96.82%, 95.41%, newlineand 96.40% with TP. With TP, dataset 4 achieves 99.13%, 98.72%, 98.00%, 96.73%, newlineand 97.72%; with 10-fold, it achieves 99.28%, 99.32%, 99.22%, 98.85%, and newline98.92%, it is more effective compared to other methods newline