Uma nova abordagem para a predição de crises epilépticas baseada nas técnicas de padrões espaciais comuns e aprendizagem de máquina

Abstract

Epilepsy is one of the most common neurological diseases characterized by recurrent seizures caused by brief disturbances in the brain's electrical functions. In 30% of the cases, this condition cannot be successfully treated by medication or resection, directly affecting the quality of life of these individuals. Thus, there is a significant interest in developing reliable tools for predicting seizures, enabling decision making, or alerting patients to be prepared when a seizure is approaching. The proposed method for seizure prediction is based on time-frequency analysis of the scalp electroencephalogram (EEG) and spatial filtering techniques to extract features capable of discriminating the interictal and preictal activities. The coefficients of the theta, alpha, and beta EEG rhythms, obtained by the decomposition of the Wavelet Discrete Transform, are subjected to the Common Spatial Patterns filtering technique. Statistical and entropy-related attributes are extracted, and then features are selected and applied to the SVM classifier with Gaussian kernel to discriminate cerebral state as preictal or non-preictal. The proposed algorithm is evaluated on multichannel surface recordings of 17 subjects with refractory epilepsy from the Children's Hospital Boston and Massachusetts Institute of Technology (CHB-MIT) database. Two techniques, namely Kalman Filter and Median Filter, are used to smooth the classifier's outputs. A final decision of each EEG epoch is yielded after a thresholding process. The best results have shown an average precision of 68.8% for sample classification. The alarm generator reported a false-positive rate of 0.334 per hour.