Abstract

Long-term EEG monitoring of epileptic patients makes automatic seizure detection necessary, because it is hard for clinicians to interact with the patients or view the recordings continuously. The problem of seizure detection is inherently difficult because seizure EEG activity consists of a variety of morphologies. It is generally difficult to design a single method that can detect all types of seizures in all patients. In most patients however, one or sometimes two or three types of seizures tend to occur repeatedly. In these cases, the electrographic seizures of each type are similar to each other. Based on this observation, we propose our model-based seizure detection method. In this thesis, a model-based seizure detection method using statistically optimal null filters (SONF) is presented. A template seizure from a patient is first selected, and a set of basis functions that model the template seizure is derived using the proposed modeling methods. Subsequent electroencephalogram (EEG) recording is processed by the SONF and the output represents the noise-free estimate of the seizure. The energy ratio between the output and the input of the SONF is calculated and processed, and used as the test statistic for the seizure detection. Simulation result shows that the modeling Method 4 (Sinusoidal wavelet basis functions) has better performance than other modeling methods. Experiments using 100 hours real SEEG recordings from 5 patients show that the model-based seizure detection method using SONF can lower the false detection rate, and it is most effective for long rhythmic seizures with a clear pattern.