Detectores de alto desempenho para crises epilépticas por técnicas de aprendizado de máquina

Among the various research niches in epilepsy, scientists and computer engineers have been contributing with methods for the detection and prediction of seizures, as well as with the location of epileptic foci. A common challenge is represented by the inherent variability of the disease and the scarcity of data containing seizure onset time and length, usually making most optimized solutions found for a particular patient of poor performance when applied to other patients. It highlights the need for developing more robust feature extraction strategies and predictive intelligent algorithms, capable of discriminating cerebral states under diverse conditions, as well as, ideally, being able to convey information across disjoint predictive tasks. This work addresses, first, a comparison of multiple representations of electroencephalographic signals. These representations yield feature vectors of the underlying phenomena which have intrinsic distinct interpretations. We compare Fourier, wavelet and graph-based decompositions of the incoming signal, which correspond to features encoding energies of the signal at various time scales, the similarity of the waveform to that of a given motherwave function and the correlation between multiple cerebral areas, respectively. Finally, we evaluate multiple methods that could prove to be useful when experimental conditions are not ideal, in which there is a lack of properly labeled data. This difficulty is initially handled by an autoencoder, which solely learns the profiles of normal cerebral states, not containing seizure timestamps, and detects seizures through a measure of their divergences with respect to normal timestamps, thus avoiding the need for experimental training sets that contain seizure labels. Algorithms that are part of a community of transfer learning methods, in computer science, are also studied in contexts in which we try to use distinct patients' labeled instances aiming at improving a detection task for a given target patient (AU)