Automated Marking of Sleep Spindles using Wavelet Packet Decomposition and Peak Tracking

Abstract

Sleep spindles, along with K-complexes are hallmarks of stage 2 non rapid eye movement (NREM) sleep EEG. Sleep spindles are of significant interest because they are associated with phenomena such as ‘stability’ of sleep, updating of knowledge with new memories, processing of sensorimotor and mnemonic information. Therefore, accurately marking their presence in sleep recordings is essential. Accurate identification of spindles in EEG recordings has proved to be a time consuming task, even with the help of experts. Further, manual detection by different experts introduces disparity and biases due to inter-rater differences. Hence there is a crucial need for an automated detection algorithm. The objective of this paper was to develop a robust algorithm for real-time automated spindle detection based on the wavelet packet decomposition. The developed algorithm replicates the marking methodology used by sleep specialists to identify spindles. Spindles are transient 11–16 Hz oscillations present in NREM with higher amplitude than the background delta waves. To identify the spindles the EEG data was divided into epochs from which appropriate features were extracted to differentiate spindles from the background EEG. The feature vectors used included the level of EOG activity, the quantity of significant peak-to-peak transitions, the wavelet packet energy (WPE) within the frequency band of interest (11–16 Hz) and the presence of K-complexes. EOG activity was tracked to identify NREM sleep sections. Spindles were marked as being present in those epochs in which the WPE and peak-to-peak activity were higher than predetermined thresholds. The thresholds were reduced on detection of K-complexes, mimicking manual scoring. The accuracy of the developed algorithm was verified by comparing to the manual scoring performed by a sleep specialist on the EEG data. The results from the algorithm look promising with a good degree of agreement with the manual scoring. When run on 3 hours of EEG data with 52 manually scored spindles the algorithm successfully detected 42 of them (80.7%) and of the total 21,600 epochs analyzed 290 were falsely detected as containing spindles. It was also observed that the true detection rate increases on varying the thresholds although this introduces further false detections.