Contents lists available at ScienceDirect
Biomedical Signal Processing and Control
journal homepage: www.elsevier.com/locate/bspc
Automated sleep spindle detection with mixed EEG features
Peilu Chen
a
, Dan Chen
a,
<
,LeiZhang
a
,YunboTang
a
,XiaoliLi
b
a
School of Computer Science, Wuhan University, Wuhan 430072, China
b
National Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing 100875, China
ARTICLE INFO
Keywords:
Spindle detection
Sleep EEG
Convolutional Neural Network
Deep features
Entropy
ABSTRACT
Detection of sleep spindles, a special type of burst brainwaves recordable with electroencephalography (EEG),
is critical in examining sleep-related brain functions from memory consolidation to cortical development. It
has long been an onerous and highly professional task to visually position individual sleep spindles and label
their onset & offset. Automated spindle detection (template- and classifier-based) is experiencing performance
bottleneck due to uncertain variances between spindles in both duration & formation.
This study then develops a generic framework based on Deep Neural Network for accurate spindle detection
by mixing the deep (micro-scale) features and the entropy (macro-scale) of sleep EEG. First, an ‘‘elastic’’ time
window applies to adapt to the significantly varied durations of spindles in EEG, after which regulated deep
features of EEG epochs with variable-lengths are obtained via a compact Convolutional Neural Network (CNN)
with spatial pyramid pooling. Second, these deep features are mixed with the entropy of EEG epochs to support
spindle classification. Focal loss applies to ease the severe imbalance between spindles and other epochs.
Finally, elastic EEG epochs are set to capture the individual spindles.
Experimental results on a public sleep EEG dataset (DREAMS) with the proposed framework against the
state-of-the-art counterparts indicate that (1) it outperforms the counterparts with an F1-score of 0.66(0.11)
while introducing entropy information gains 0.034(0.02) i n this process; (2) it incurs less errors in identifying
the onset & offset of spindles. Overall, the core design of the framework paves the way for detection of
complicated EEG waveforms or time series in general.
1. Introduction
Characteristic electroencephalography (EEG) waveforms often func-
tion as the crucial ‘‘bio-markers’’ in neuro-science & engineering tasks
such as evaluation of brain development status, consciousness mea-
surement, and cognitive neurological ergonomics [1,2]. It is well-
known that spikes and sharp waveforms (two interictal epileptiform
discharges, IEDs) signify the hallmarks of epilepsy [3] and other dis-
order with neural networks leading to cognitive issues especially with
children [4,5]. Spindle is a special type of burst waveforms observable
in sleep EEG [6], routinely characterized by a frequency range 11–
16 Hz with a duration 0.5–2 seconds (see Fig. 1,Rechtschafenand
Kales (R&K) and American Academy of Sleep Medicine (AASM) [7,8]).
Spindles are believed to mediate many sleep-related brain functions,
from memory consolidation to cortical development [9]. Sleep spindle
detection has become a fundamental task to sustain examination of
these functions.
It has long been an onerous and highly professional task to visually
position each individual sleep spindle and label its onset & offset [10,
<
Corresponding author.
E-mail address: dan.chen@whu.edu.cn (D. Chen).
11]. The situation is even worse due to a lack of inter-expert agree-
ment and uncertain biases in manual inspections [12,13]. Automated
approaches to spindle detection from sleep EEG have recently emerged
from the signal processing community and gained successes especially
with the booming of computational intelligence [14,15].
Traditionally, template-based approaches largely make use of tem-
plate matching rule for the detection of sleep spindle. Early approaches
are based on a bandpass filtering and amplitude threshold [16,17].
Later, time–frequency analysis has also applied to decompose oscil-
latory and transient components to form the envelope plane to be
checked with constant or adaptive thresholds [18]. Those methods
manifest superiority in simplicity of modeling and strong interpretabil-
ity of corresponding applications. However, the performance of these
approaches is generally limited due to uncertain variances between
spindles in both duration & formation (e.g., density and amplitude)
and significant variances amongst individuals. Adaptive parameter set-
tings has yet been generically enabled to adapt to such intensive
variance [12,19].
Later, classifier-based approaches regard the detection of sleep
spindle as a classification task, that is to extract multiple features
https://doi.org/10.1016/j.bspc.2021.103026
Received 12 April 2021; Received in revised form 20 July 2021; Accepted 30 July 2021