Modelling of stage 2 sleep EEG data

Abstract

Humans spend approximately a third of their lives sleeping. Undoubtedly, sleep is essential to human health and sleep research continues to reveal more about the characteristics and structures of sleep. During a night’s sleep, brain activity cycles through a number of stages, each with its own characteristics that can be clearly extracted from an electroencephalogram (EEG), which records the brain electrical signals from the human scalp. EEG recordings for stage two sleep contain two hallmark events known as sleep spindles and K-complexes. Spindles have a strong clinical significance because they tend to change with age and atypical spindling is associated with a range of disorders and diseases. In particular, spindles hold promise as a biomarker of dementia. Sleep spindles are generally extracted manually by human experts from voluminous sleep EEG recordings. Since this process is time consuming and prone to human bias, many studies have recently tried to implement automatic spindle detectors which label spindle activity in an EEG recording. This dissertation investigates the operation of two different spindle detectors and compares their performance when scoring spindles in two sleep EEG databases, one of which is open access. One of the detectors is a root-mean-square (RMS) amplitude detector which is commonly used for discussion and comparison in the literature. It identifies spindles based on the temporal characteristics of the EEG signal. The second detector is an autoregressive switching multiple model (AR-SMM) detector which consists of a number of mathematical models representing different modes of the EEG signal: background EEG and spindle activity. These models are trained on pre-scored data and are then used to score spindles in new, incoming EEG data. This work has shown that overall the RMS detector exhibited better performance over the two EEG datasets tested and was found to be less sensitive to the amount of data used to extract the necessary detector parameters. The lower AR-SMM detector performance may have been due to the quality of the data used for training and thus future work can investigate how this can be improved using data representative of fundamental spindle characteristics and not marred by noise, disturbances or artefacts.