EEG signal analysis for brain injury detection
Detection of brain injuries is possible based on EEG measurements if injuryrelated changes in the EEG can be accurately described. By modeling EEG signal as a linear AR process, two important parameters, AR model order and AR model coefficients, can be used to describe the EEG signal. The optimal order of EEG signal is found to be unreliable for brain injury detection. Euclidean distance based on AR model coefficients, Itakura distance, generalized Itakura distance, AR model spectral distance, modified AR model spectral distance, and Kullback-Leibler information distance are used to describe the changes in EEG signal in this thesis. The Itakura distance does not have any big difference between the EEG during anesthesia (baseline) and EEG during hypoxia, but the Itakura distance could be employed to effectively and efficiently quantify severe injury-related changes in the EEG. A generalization of the Itakura distance is proposed to allow the comparison of EEG segments with different optimal AR model orders. The performance of the Itakura distance is compared with Euclidean distances of different parameters, spectral distances, as well as the Kullback-Leibler information distance. The AR model-based spectral distances are effective to detect non-severe injury changes (hypoxia). Analysis of residue error energy ratio reveals that EEG segments from an injured brain can no longer be modeled accurately by a linear AR process of any order. A natural extension of linear AR models is to include nonlinear functions of past values as a part of the regression vector. Such a nonlinear AR model may result in a more accurate description of the injury and is a suggestion for future work.