ANÁLISE DA CLASSIFICAÇÃO DE CRISES EPILÉPTICAS COM BASE EM CARACTERÍSTICAS ESTATÍSTICAS DOS MOMENTOS CONJUNTOS DA DISTRIBUIÇÃO TEMPO-FREQUÊNCIA DE SINAIS EEG UTILIZANDO TRANSFORMADA WAVELET CONTÍNUA E REDES NEURAIS PROFUNDAS

Abstract

Epilepsy is a chronic neurological condition characterized by recurrent seizures resulting from abnormal, excessive, and synchronous brain discharges. Accurate classification of these seizures is essential for diagnostic precision and the establishment of effective therapeutic strategies. However, the visual analysis of electroencephalogram (EEG) signals remains a laborious process, subject to subjectivity and inter-observer variability. In this context, this dissertation proposes an automated deep learning-based approach for classifying four types of epileptic seizures: Generalized Non-Specific Seizure (GNSZ), Complex Partial Seizure (CPSZ), Focal Non-Specific Seizure (FNSZ), and Tonic-Clonic Seizure (TCSZ). Statistical features were extracted from time-frequency representations obtained through the Continuous Wavelet Transform (CWT), applied to signals from the public Temple University Hospital Seizure Corpus (TUSZ) database. Three wavelet functions (Morse, Morlet, and Bump) were evaluated, from which statistical moments of mean, variance, skewness, and kurtosis were calculated in temporal, spectral, and joint domains. These features were used as input for models based on Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and a hybrid CNN-LSTM architecture. The results showed that the combination of the Morse wavelet with the CNN-LSTM architecture and the variance and kurtosis features achieved the best overall performance, with an average accuracy of 96.8%, precision of 96.5%, sensitivity of 96.9%, and F1-score of 96.7%. The TCSZ class stood out, with accuracy of 98.2% and F1-score of 98.0%. Among the evaluated wavelets, Morse achieved the best performance (95.2%), outperforming Bump (93.7%) and Morlet (91.8%), a result attributed to its greater ability to represent low-frequency components. Regarding statistical features, variance demonstrated the highest discriminative power (94.3%), followed by kurtosis (92.7%), while the mean showed inferior performance (accuracy below 84%). These findings highlight the potential of integrating time-frequency analysis and deep neural networks in the development of intelligent systems to support the clinical diagnosis of epilepsy.