Studying the Use of Hidden Markov Models in the Detection and Classification of EEG Epileptiform Transients using LPC features

The process of identifying the presence of an AEP (Abnormal Epileptiform Paroxysmal) in a subject\u27s EEG, normally done by neurologist experts, is a particularly long one and involves considerable financial expenses. This research aims to pave an automatic method of detecting and classifying streams of EEGs as to whether or not it has any AEPs present in it. This is a two step process, where step 1 is the classification problem and step 2 is the detection problem. There are many different activities on the EEGs, and the classification task helps to identify which of these activities are AEPs. So, this task involves training 2 HMMs to classify all given artifacts into 2 classes, AEP or NonAEP. LPC features extracted from the spike have been used to train the HMMs. The detection task is to find out the presence of ETs (Epileptiform Transients) from a patient\u27s EEG. For detection, two HMMs have been trained on examples taken from two classes, the ETs and the Non-ETs. The ETs class is all the Yellow Boxed annotations provided to us by the experts. The Non-ET class data has been formed by taking into consideration all the data which has not been marked as an ET. In this task, LPC features extracted from the spike and the contextual information has seen to provide good results. For validation of the system, a cascaded structure of four HMMs is formed. The first two HMMs are for detection and the next two classify the detected ETs. Test EEG signals, having both AEPs and NonAEPs are passed through this system, and the AEPs are marked and identied. The results have been compared to the annotations marked by experts