oa Selecting the maximum relevant translated EEG time-frequency features with a minimum redundancy using the mutual information measure for newborn seizure detection and classification

Background Analyzing the Electroencephalogram (EEG) is a standard approach for detecting newborn seizure. The manual detection of newborn seizure by visual scanning of EEG recordings is time consuming especially with long recordings. It also requires skilled interpreters, i.e. a neurophysiologist, resulting in possible subjective judgment and error. Hence, the EEG signal parameters extracted and analyzed using computer based digital signal processing techniques are highly useful in diagnostics and more suitable for detecting newborn EEG seizures and other abnormalities. Objective This work aims to select the maximum relevant translated EEG time-frequency features with a minimum redundancy to improve the classification accuracy performance of newborn seizure detection and classification systems. Method The automatic newborn EEG seizure detection and classification system includes pre-processing of EEG signals, finding their optimal time-frequency distributions (TFDs), extracting features from the TFDs, and finally allocating the T-F features to the relevant class. Based on this classification system, a new approach is proposed to improve their performance, and includes the following stages: Defining T-F features by translating some relevant time features and/or frequency features in the T-F domain; selecting the maximum relevant translated T-F features with a minimum redundancy using the mutual information measure and using the selected relevant features to characterize and classify different newborn EEG seizures. Results The experimental results show that the selection of a minimum set of relevant translated EEG T-F features according to a combined minimal-redundancy and maximal-relevance criterion significantly improve the performance of the newborn seizure classification system based on the use of T-F features extracted using both signal and image processing techniques, by up to 4% for 100 real newborn EEG segments. The better results are obtained using the modified-B distribution and the spectrogram distribution with the multi-class SVM classifier. Also, the use of the selected features reduces the computation cost of the classification system. Conclusion The improvement obtained is dependent on the choice of relevant translated EEG T-F features. The latter can also be extended and applied to detect other newborn EEG abnormalities.