| Abstract [eng] |
This research paper examines the development and analysis of automated systems designed to detect epileptic seizures using electroencephalogram (EEG) signals. The study used EEG data collected at Boston Children's Hospital and the Massachusetts Institute of Technology (MIT), covering long-term recordings of 24 pediatric patients. Several different methods were used to detect epileptic seizures. First, features were extracted from the EEG signals using discrete wavelet transform (DWT) and then classified using a support vector machine (SVM) and k-nearest neighbors (KNN) method. Second, spectrograms obtained from short-time Fourier transform (STFT) were used as input to various convolutional neural network (CNN) architectures and pre-trained models: ResNet50, ResNet50V2, EfficientNetB0, and EfficientNetV2B0. The best results among SVM models were achieved by the configuration with the Haar wavelet function and 5-level decomposition, which showed 0.86 accuracy, 0.91 precision, 0.80 recall, and 0.85 F1 score. The KNN algorithm performed best using the bior6.8 forest, but with level 2 decomposition—this model achieved an accuracy of 0.87, sensitivity of 0.80, precision of 0.98, and an F1 score of 0.88. Among the pre-trained models, ResNet50V2 demonstrated the best results, with classification metrics of 0.687 accuracy, 0.702 precision, 0.687 sensitivity, and 0.682 F1 score. Summarizing all the results obtained, the best epilepsy seizure detection efficiency was achieved using a combination of DTW and KNN methods. |
