Automated spike detection using cascade of simple classifiers

The diagnosis of epilepsy heavily depends on the detection of interictal epileptiform spikes in EEG recordings of patients. However, the traditional visual inspection is time-consuming and subjective. The infinite variety of spike morphologies and the similarity of spikes to normal EEG and artifacts also make the detection of spikes difficult. As a result, automated spike detection methods are in great demand. To this end, we propose a cascade of simple classifiers to detect spikes by rejecting non-spikes (or background waveforms) partially at each step of the cascade. We utilize EEG recordings from 9 patients with epilepsy as training data to train various support vector machines (SVMs), and to build the cascades based on this SVMs. We test the performance of the cascades of SVMs on a new patient. In our experiments, the cascade consisting of 122 different SVMs achieved 95.76% accuracy, 98.64% sensitivity, and 95.58% specificity. To validate its performance, we further compare the results of cascade of simple classifiers with background rejection by thresholds of feature values [1]. After several steps of background rejection, the cascade of simple classifiers may receive better performance than the background rejection.Master of Science (Computer Control and Automation