You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This severely limits data throughput, and human subjectivity is likely to introduce systematic bias.
43
46
Indeed, it is expected that scoring will be performed differently by each expert, making result difficult to reproduce independently.
44
47
Often, two experts score the same data, in order to ensure satisfying agreement.
45
-
Although, manual scorers are generally reported as being very consensual with each other\citationneeded{},
46
-
it can be argued that experts most likely work in the same laboratory and trained one another, or were trained by the same third person.
48
+
Although, manual scorers are generally reported as being very consensual with
49
+
each other\cite{costa-miserachs_automated_2003,sen_comparative_2014}, it can be
50
+
argued that experts most likely work in the same laboratory and trained one another, or were trained by the same third person.
47
51
In this context, agreement between experts does not account for the variability between communities of researchers, and cannot be used to assess reproducibility.
48
52
49
-
In order to overcome both speed and subjectivity limitations, efforts have been directed towards automation of sleep scoring.
53
+
In order to overcome both speed and subjectivity limitations, efforts have long
However, little adoption has occurred and very few available implementations, in the form of software that biologists could use, have been developed.
51
57
Typically, two different approaches to classification have been followed: unsupervised or supervised learning.
52
58
53
-
Unsupervised learning has the advantage of making no assumption about the nature of the different vigilance states, and how they should be defined.
59
+
Unsupervised learning \cite{l&xe4_sleep_2012,sunagawa_faster:_2013} has the
60
+
advantage of making no assumption about the nature of the different vigilance states, and how they should be defined.
54
61
Therefore, this approach can lead to the discovery of truly new states.
55
-
One issue is that the choice of the variables used for clustering is very critical.
56
-
Often, variables such as frequency domain variables are in fact chosen in order to generate clusters that will match human defined clusters.
57
-
In addition, unsupervised methods may lack robustness in so far as the cannot easily include covariates explaining, for instance, variability between recording equipments.
62
+
One issue is that the choice of the variables used for clustering is very
63
+
critical.
64
+
Often, variables such as frequency domain variables are in fact chosen in order
65
+
to generate clusters that will match human defined clusters.
66
+
In addition, unsupervised methods may lack robustness\cite{sunagawa_faster:_2013} in so far as the
67
+
cannot easily include covariates explaining, for instance, variability between recording equipments.
58
68
59
69
Another approach is to assume human annotations are, although imperfect, biologically relevant and generally consistent,
60
-
and therefore to use supervised learning techniques.
61
-
Of course, if human decisions were biased, such a method may suffer from the same bias.
Of course, if human decisions were biased, such a method may suffer from the same bias.
62
73
However, a vast corpus of experimental work has provided hypothesis about function of these states which tends to validate the actual `existence' of these discrete vigilance states.
63
74
Building a classifier that would produce a consensual prediction of vigilance states could be seen as an attempt to formalised and rationalise the definition of such states.
64
75
This would improve future research without denying decades of sleep neurobiology.
65
76
66
-
Many supervised learning techniques such as from \glspl{svm}, \glspl{ann}, to \glspl{hmm} have been investigated.
77
+
Many supervised learning techniques such as from
78
+
\glspl{svm}\cite{crisler_sleep-stage_2008},
79
+
\glspl{ann}\cite{ventouras_performance_2012},
80
+
to
81
+
\glspl{hmm}\cite{doroshenkov_classification_2007,pan_transition-constrained_2012} have been investigated.
67
82
In general, the first step is to compute features on consecutive segments of annotated electrophysiological signals know as epochs.
68
83
Then, the relation between the response variable(annotation) and the independent variables (features) can be modelled.
69
84
Either epochs are considered to be independent from one another or time-dependent structures are explicitly modelled (\eg{} using \glspl{hmm}).
70
85
Time aware modelling has the advantage of accounting for the interdependence of consecutive epochs (see fig.~\ref{fig:sleep_description}B).
71
86
However, it generally does not model non-linear relationships between large numbers of predictors as well as classical classifiers.
72
87
73
88
Recently, promising results were obtained for automatic scoring of human sleep stages by performing an exhaustive
74
-
feature extraction, including variables resulting from discrete wavelet decomposition.
75
-
Then, the authors compared several classifiers and found that random forest were, overall, the most accurate predictors.
89
+
feature extraction, including variables resulting from discrete wavelet
90
+
decomposition\cite{sen_comparative_2014}.
91
+
Then, the authors compared several classifiers and found that random
92
+
forest\cite{breiman_random_2001} were, overall, the most accurate predictors.
76
93
77
94
The study herein bases itself on these promising results by computing an even larger number of features.
78
-
An important addition was the computation of time-aware features which significantly improved accuracy.
79
-
Furthermore, rigorous stratified cross-validation procedure and comparisons of sleep structure were performed.
95
+
An important addition was the computation of time-aware
96
+
features\cite{dietterich_machine_2002,deng_time_2013} which significantly improved
97
+
accuracy.
98
+
Furthermore, rigorous stratified cross-validation\cite{ding_querying_2008} procedure and
99
+
comparisons of sleep structure were performed.
80
100
These improvement altogether contributed to achieve a very satisfying overall accuracy of 92\%.
81
101
In order to pave the way to an implementation of an ubiquitous sleep scoring software.
82
102
\pr, a new \py{} package was also build to facilitate efficient feature extraction.
0 commit comments