Electroencephalogram (EEG) is used as a primary tool by neurophysiologists in monitoring and diagnosing seizures in newborns. However, analysis of EEG by human experts has many limitations. As a consequence, growing attention has been focused on the development of computerized methods for automatic newborn seizure detection based on the EEG. However, the neonatal seizure recognition remains a very challenging task. Several researchers have found an association between seizures and changes in heart rate and electrocardiogram (ECG) rhythm. However there has been little work done with signals other than EEG for detecting seizures in the newborn. These facts gave me motivation to investigate the use of heart rate variability (HRV) to provide additional information to improve on the existing newborn EEG seizure detection.
In this thesis, I propose and compare feature-level and decision-level fusion of EEG and HRV for newborn seizure detection. Both the proposed technique consists of a sequence of processing steps, namely: preprocessing of the signals, feature extraction of the signals, feature selection and finally the combination process. The preprocessing steps remove extraneous data and prepare the data for the feature extraction stage.
Feature extraction extracts those features of the signals during seizure and non-seizure states that are suitable for classification. A number of time-frequency features from the newborn HRV related to seizure and non-seizure epochs were proposed and extracted. Furthermore, it has been shown that accurate components of HRV can be obtained using a recently developed time-frequency based multicomponent instantaneous frequency estimation technique. The number, locations, time durations, bandwidth and the amplitude (energy) of HRV components corresponding to seizure and non-seizure epoch were clearly exhibited by the method. Based on this information, seizure activity can be distinguished from non-seizure activity. Features of newborn EEG seizure and non-seizure epochs were derived from the existing methods in literature.
Feature selection step selects an optimal subset of HRV and EEG features from the larger set extracted by removing irrelevant and redundant data that do not contribute in class distinction. Using a filter approach, a feature selection method based on discriminant and redundancy analysis has been proposed for this purpose. The proposed technique is able to significantly reduce the number of features used and achieve high classification performance compared to the original feature set. In addition, it has been shown that the proposed filter can be used as a preprocessing stage for a wrapper-based feature selection algorithm. This has the advantage of reducing the computation load and the severity of the search operations of the wrapper-based feature selection algorithms.
Lastly, two different fusion methods, namely; feature level fusion and decision level fusion were considered to combine the selected EEG and HRV information. Feature-level fusion for newborn seizure detection is investigated in form of EEG-HRV feature vector concatenation. Decision/classifier-level fusion is studied by fusing the independent decisions from individual classifiers of EEG and HRV. In both fusion schemes, I have shown that the proposed integrated approach leads to improved performance of newborn seizure detection compared to either EEG or HRV based seizure detectors.
In order to achieve the EEG and HRV fusion scheme, new EEG multi-channel fusion model for newborn seizure detection is proposed. I investigate and evaluate the two fusion strategies (feature fusion and decision fusion) mentioned before, to combine simultaneously recorded multi-channel EEG features. It has been shown that the detection method based on feature fusion outperforms the one based on classifier fusion. It has also been shown that the new newborn seizure detection based on multi-channel EEG outperforms some existing well-documented newborn EEG seizure detection technique. For these reasons, the multi-channel EEG feature fusion configuration was used as a component in the seizure detection process that combines information from EEG and HRV.