Analysis of human polysomnography (PSG) for automatic sleep event detection using Hidden Markov Model
MetadataShow full metadata
This thesis work evaluates our proposed methodology for automated detection of sleep events from Polysomnographic (PSG) data. The sleep data was collected during real sleep studies using Profusion PSG3. The event detection tasks used a Hidden Markov Model (HMM) to achieve signal classification for sleep event detection. The Hilbert transform (envelope) was used to extract features for input to the HMM. HMM was selected as our classification method of choice, due to the fact that it was able to capture the temporal variations of the biosignals collected through PSG. In this work, we detected sleep motion events, such as rapid eye movements (REM) and leg movements, and breathing events like obstructive apnea, hypopnea and snore. The task of detecting events of interest was achieved using a sliding window approach, and classifying each signal segment as containing an event or not, hence, leading to a binary classification problem for each type of event. Our experimental results show that our proposed approach can be successfully used for sleep event detection, to assist experts in sleep quality assessment, however, the big imbalance between the number of segments that contain a positive event and the ones that do not, often negatively affects the performance of our classification method.