=Paper=
{{Paper
|id=Vol-2491/abstract69
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2491/abstract69.pdf
|volume=Vol-2491
|dblpUrl=https://dblp.org/rec/conf/bnaic/SeeuwsAHN19
}}
==None==
https://ceur-ws.org/Vol-2491/abstract69.pdf
Unsupervised Deep Feature Extraction for
Neonatal Sleep Stage Classification
Nick Seeuws1 , Amir Hossein Ansari1 , Sabine Van Huffel1 , and Gunnar
Naulaers2
1
Department of Electrical Engineering (ESAT), STADIUS, KU Leuven, Belgium
2
Department of Development and Regeneration, University Hospitals Leuven,
Neonatal Intensive Care Unit, KU Leuven, Leuven, Belgium
Preterm birth carries many risks for an infant. These infants are accomo-
dated in a specialized unit of the hospital: the Neonatal Intensive Care Unit,
NICU. The NICU utilizes many tools to provide care and assess maturation of
these infants. Sleep stage monitoring is such an important tool. EEG provides
a non-invasive way of performing the monitoring task. However, correctly iden-
tifying sleep stages from EEG recordings is a challenging and time-consuming
task even for experts.
Machine learning offers clinicians a way of performing continuous sleep stage
monitoring with a minimum of human intervention. A big drawback, however,
is the need of labeled examples for training machine learning models. Due to the
complexity of the task it is assumed that a substantial amount of training data
is needed placing a heavy burden on the experts.
This work aims to extend the usability of machine learning models for sleep
stage classification in preterm newborns. To this end, it investigates the potential
of leveraging unlabeled EEG recordings as additional training data. By doing
so this allows for lowering the human effort required for further improvement of
classification models. The focus was on distinguishing quiet sleep from non-quiet
sleep. Two approaches are considered: an unsupervised feature extraction model
utilizing all data, labeled and unlabeled, followed by a supervised classifier mak-
ing use of the extracted features and the corresponding labels as a first approach
and a semi-supervised model jointly utilizing the labeled and unlabeled infor-
mation as a second approach.
The EEG data was recorded at the Neonatal Intensive Care Unit of the Uni-
versity Hospitals in Leuven, Belgium. The Ethics Committee of the University
Hospitals provided approval for the recordings and informed parental consent
was obtained. The labeled dataset consisted of recordings of 26 preterm infants
born before 32 weeks of gestation. 97 recordings taken between 27 and 42 weeks
of postmenstrual age make up the dataset for a total of 492 hours of EEG data.
Recordings were split up into 30 second segments and labeled as quiet or non-
quiet. A full description of the dataset can be found in the work by Dereymaeker
Copyright c 2019 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0).
2 N. Seeuws et al.
et al. [3] The unlabeled dataset consists of seven additional recordings providing
an additional 109 hours of EEG recordings. This data was recorded under the
same modalities as the labeled dataset. The signal is filtered and downsampled
to 30Hz from 250Hz for use in this work. Altough the full dataset provides a
multichannel signal, this work focused on single channel bipolar EEG data to
simplify implementation and experimentation. It used the difference between the
C3 and C4 electrode based on the international 10-20 system[2] as a signal.
The unsupervised approach used a Variational Auto-Encoder, VAE, as in-
troduced by Kingma et al.[5] to extract features from the EEG segments. After
training on the combined dataset of EEG recordings, the VAE defines a poste-
rior probability distribution in latent space given an EEG segment and the mean
value of this posterior was used as the extracted feature vector of a segment. A
gradient boosting classifier performed the final classification using the extracted
features of the labeled segments.
The semi-supervised approach made use of a Generative Adversarial Net-
work, GAN, as introduced by Goodfellow et al.[4] The discriminator of the
GAN was extended with an additional output acting as the desired classifier
as proposed by Salimans et al.[7] This model can make direct use of labeled and
unlabeled information and does not need an additional classifier to detect sleep
stages.
The results are benchmarked against two other approaches. The first makes
use of a set of features deemed relevant for sleep stage identification as proposed
by Piryatinska et al.[6] mainly focused on spectral information of the EEG seg-
ments. The features are classified by a gradient boosting model. The second ap-
proach makes use of a convolutional neural network in a traditional supervised
setting proposed by Ansari et al.[1] Several performance metrics were computed
but models were mainly compared using Cohen’s kappa coefficient. The VAE
based model scored 0.47 on the test set while the GAN based model scored 0.64.
Comparing these kappa values to the 0.43 scored by the classifier making use of
spectral features[6] and the 0.60 reported for the supervised CNN by Ansari et
al.[1] one can see the performance of a sleep stage classifier can be improved by
making use of unlabeled data.
The VAE based model succeeded at extracting slightly better features com-
pared to classical spectral features based on the improved classification per-
formance but failed to improve upon the performance of an end-to-end deep
learning approach. The GAN based model did show improved performance to
supervised deep learning when making use of unlabeled information.
A further investigation of the feature space for both the VAE based model
and the GAN based model leads to the conclusion that an unsupervised model
struggles to separate factors of variation corresponding to sleep stage where a
semi-supervised approach does identify such relevant factors.
Title Suppressed Due to Excessive Length 3
The results show a possibility to improve sleep stage classification perfor-
mance by leveraging unlabeled recordings. Increased performance is currently
only observed for semi-supervised models.
References
1. Ansari, A., De Wel, O., Pillay, K., Dereymaeker, A., Jansen, K., Van Huffel, S.,
Naulaers, G., De Vos, M.: A convolutional neural network outperforming state-of-
the-art sleep staging algorithms for both preterm and term infants. Internal Report
19-65, ESAT-STADIUS, KU Leuven (Leuven,Belgium) (2019 (submitted for publi-
cation))
2. Cherian, P.J., Swarte, R.M., Visser, G.H.: Technical standards for recording and in-
terpretation of neonatal electroencephalogram in clinical practice. Annals of Indian
Academy of Neurology 12(1), 58 (2009)
3. Dereymaeker, A., Pillay, K., Vervisch, J., Van Huffel, S., Naulaers, G., Jansen, K.,
De Vos, M.: An automated quiet sleep detection approach in preterm infants as a
gateway to assess brain maturation. International journal of neural systems 27(06),
1750023 (2017)
4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural infor-
mation processing systems. pp. 2672–2680 (2014)
5. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint
arXiv:1312.6114 (2013)
6. Piryatinska, A., Terdik, G., Woyczynski, W.A., Loparo, K.A., Scher, M.S., Zlot-
nik, A.: Automated detection of neonate eeg sleep stages. Computer methods and
programs in biomedicine 95(1), 31–46 (2009)
7. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Im-
proved techniques for training gans. In: Advances in neural information processing
systems. pp. 2234–2242 (2016)