Design of a simulator for neonatal multichannel EEG: Application to time-frequency approaches for automatic artifact removal and seizure detection

The electroencephalogram (EEG) is used to noninvasively monitor brain activities; it is the most utilized tool to detect abnormalities such as seizures. In recent studies, detection of neonatal EEG seizures has been automated to assist neurophysiologists in diagnosing EEG as manual detection is time consuming and subjective; however it still lacks the necessary robustness that is required for clinical implementation. Moreover, as EEG is intended to record the cerebral activities, extra-cerebral activities external to the brain are also recorded; these are called “artifacts” and can seriously degrade the accuracy of seizure detection. Seizures are one of the most common neurologic problems managed by hospitals occurring in 0.1%-0.5% livebirths. Neonates with seizures are at higher risk for mortality and are reported to be 55-70 times more likely to have severe cerebral-palsy. Therefore, early and accurate detection of neonatal seizures is important to prevent long-term neurological damage. Several attempts in modelling the neonatal EEG and artifacts have been done, but most did not consider the multichannel case. Furthermore, these models were used to test artifact or seizure detection separately, but not together. This study aims to design synthetic models that generate clean or corrupted multichannel EEG to test the accuracy of available artifact and seizure detection algorithms in a controlled environment. In this thesis, synthetic neonatal EEG model is constructed by using; single-channel EEG simulators, head model, 21-electrodes, and propagation equations, to produce clean multichannel EEG. Furthermore, neonatal EEG artifact model is designed using synthetic signals to corrupt EEG waveforms. After that, an automated EEG artifact detection and removal system is designed in both time and time-frequency domains. Artifact detection is optimised and removal performance is evaluated. Finally, an automated seizure detection technique is developed, utilising fused and extended multichannel features along a cross-validated SVM classifier. Results show that the synthetic EEG model mimics real neonatal EEG with 0.62 average correlation, and corrupted-EEG can degrade seizure detection average accuracy from 100% to 70.9%. They also show that using artifact detection and removal enhances the average accuracy to 89.6%, and utilising the extended features enhances it to 97.4% and strengthened its robustness.