| Individuals with epilepsy experience seizure disabilities, injuries, impairment of productivity, and disabling side effects from medications. The goal of this research is to detect seizure onset as early as possible with maximal accuracy and with a minimum number of false negatives and false positives, ultimately developing an automatic system allowing patients to take appropriate precautions. The system designed is based on a pattern recognition approach that encompasses the stages of preprocessing, processing, classification, and validation.; The extraction and selection of features was performed within a designed methodological environment that also allowed a parallel investigation of seizure prediction. For the classification stage, a probabilistic neural network was chosen as the decision block. A crossvalidation scheme was used to validate the results. The overall detector was evaluated for 119 one-hour records from all the patients with a three-dimensional feature vector. The average delay for detecting the seizure electrographic onset with the system developed was 1.76 seconds with zero false negatives and an average of 1.02 false positives per hour, resulting in an average clinical onset prediction time of 11.26 seconds. In addition, from the features investigated, the accumulated energy was found as a promising indicator for seizure electrographic onset prediction, yielding 85.19% of accuracy with an average prediction time of 18.49 ± 13.42 minutes.; The main contribution of this research is the development of a systematic methodology for tackling the seizure onset detection problem on a patient basis, as well as the actual software implementation of the overall seizure detection system. Within this methodology, the most relevant steps toward advancing the field are: the development of an original technique to determine an “optimal” window length for every feature on each patient; the establishment of the best feature vector according to a measure of class-separability; a comparison for the first time of fractal dimension algorithms which demonstrated that factors like the window length used, noise level in the data set, and fractal dimension range of the data, can greatly affect the accuracy and performance of the algorithm used; the finding of the accumulated energy as a very promising feature for forecasting seizures; and the design of an original linear performance metric to evaluate the classification results adapted to this particular application. |
