| Abstract |
Neonatal seizures are a condition happening in early childhood years and it is accounting for several deaths and severe problems on newborn neonates. Despite the early advancements on the treatment of this condition, the main problem concerning the physicians is the difficulty to identify and characterize a seizure, as one a small percentage gets detected in neonatal intensive care units (NICU). Multi-channel EEG signal analysis is the gold standard for seizure detection. However, the interpretation of such signals presents a great challenge, since only experienced pediatric neurologists who have emphasized in neonatal EEG analysis can perform this task. Machine learning methods can become a useful tool in the interpretation of EEG signals and in the assignment of seizure classification and regression tasks.
Various studies exist in the literature that have also employed supervised machine learning methods for neonatal seizure classification. However, an important step before proceeding with seizure classification, is rejecting the multiple artefacts that exist throughout the whole EEG signal. Especially in neonatal EEG analysis, where there are more artifacts compared to adult EEG signals, further steps of preprocessing need to be considered.
In our study, we included an extra step, besides the basic frequency filtering steps proposed in the literature, of a signal decomposition to its independent signal sources, by using independent component analysis (ICA). This way, and by computing some statistical measures as thresholds for component rejection, we managed to isolate the independent noise sources that were present throughout the whole frequency spectrum and reject them upon confirming their noisy nature.
Having artefact-free signal sources, we performed wavelet analysis to extract features both in time and frequency domain, which would serve as classifiers for the supervised classification models. The basic brain rhythm frequency bands were extracted, along with some additional statistical measures, as suggested by the literature.
Two seizure classification models were trained on two-class labeled datasets, containing seizure and non-seizure windows. An SVM and a random forest classifier were cross validated and used for the classification step and the features were finally reduced by performing feature selection to remove the redundant ones. The whole process was repeated in four different trials, where seizure and non-seizure windows of varying length were used to observe the impact of the different window size on our models.
Both classification models were tested on independent datasets and yielded great accuracy scores of more than 82% for SVM and more than 95% for random forest. This thesis contributes two classification models for neonatal seizure detection, as well as six selected features (delta_meanEnergy, gammaLow_meanEnergy, gammaHigh_meanEnergy, Shannon_entropy, Renyi_entropy and Kurtosis) which yielded high accuracy scores. The importance of thorough artefact rejection is discussed, as well as the differences between the two classification models and the impact of the varying window size on their performance.
|
Design of a clinical trial simulator for studying infant seizures
