On prediction and detection of epileptic seizures by means of genetic programming artificial features

This work presents a novel, general-purpose algorithm called Genetic Programming Artificial Features (GPAF), which consists of a genetic programming (GP) algorithm and a k-nearest neighbor classifier, and which surpasses the performance of another recently published method called Genetically Found, Neurally Computed Artificial Features for addressing similar classes of problems. Unlike conventional features, which are designed based on human knowledge, experience, and/or intuition, the artificial features ( i.e., features that are computer-crafted and may not have a known physical meaning) are systematically and automatically designed by a computer from data provided. In this dissertation, we apply the GPAF algorithm to one of the most puzzling brain-disorder problems: the prediction and detection of epileptic seizures. Epilepsy is a neurological condition that makes people susceptible to brief electrical disturbance in the brain thus producing a change in sensation, awareness, and/or behavior; and is characterized by recurrent seizures. It affects up to 1% of the worldwide population, or sixty million people, and 25% cannot be fully controlled by current pharmacological or surgical treatment. The possibility that an implantable device might eventually warn patients of an impending seizure is of utmost importance, allowing on-the-spot medication or safety measures.; Epileptic electroencephalographic (EEG) signals were treated from a chaos theory perspective. First, we reconstructed the EEG state-space trajectories via a delay-embedding scheme. Then these pseudo-state-space vectors were input to a genetic programming algorithm, which designed one or more (non)linear features providing an artificial space where the baseline (nonseizure data) and preictal (preseizure data, or ictal data in case of detection) classes are sufficiently separated for a classifier to achieve better accuracy than using principal components analysis, our benchmark feature extractor. The GPAF algorithm was applied to data segments extracted from 730 hours of EEG recording obtained from seven patients. The machine automatically discovered one or more patient-specific features that predicted epileptic seizures with a time horizon from one to five minutes before the unequivocal electrographic onset of each seizure. Results showed that 43 of 55 seizures were correctly predicted, for a 78.19% correct classification rate, while 55 epochs out of 59 representative of baseline conditions were classified correctly, for a low false positive rate per hour of 0.0508. In the case of detection, a low false-positive-per-hour-rate and a high detection rate were also achieved. A generic (cross-patient) model for prediction of epileptic seizures was also found, at the expense of decreased performance with an average of 69.09% sensitivity. The GPAF algorithm was additionally investigated to design seizure detectors. Evaluating 730 hours of EEG recording showed that with customized, artificially designed detectors, 83 of 86 seizures were detected. Seven previously unreported seizures were also detected in this work.