Beruflich Dokumente
Kultur Dokumente
FrP2A1.18
I. INTRODUCTION
3188
and Narcotrend are used in the context of depth-ofanesthesia monitoring. Each of these monitors is based on
the calculation of one or more quantitative EEG measures [8,
9], that are thought to contain anesthesia/sedation-specific
information. General anesthesia is very deep sedation with
the absence of a cortical response to pain. Some of these
devices [particularly the Bispectral or BIS monitor
(Aspect Medical)] have been studied in the context of light
sedation and are unable to reliably distinguish between light
and deep sedation [1]. We set out to determine if the
combination of a number of quantitative EEG measures or
features could be used to automatically determine a
patients level of sedation.
Grade
0
1
Criterion
Awake alpha rhythm
Decrease in amplitude and drop out
/ activity theta activity
2
Theta wave activity, sleep spindles / K
complexes
3
Generalised delta activity 20-50%
epoch duration
Table 1: EEG criteria for assignment of sedation grade
II. METHOD
A. Data set
We used an independently validated dataset of EEG
recordings from 12 healthy patients aged 37+/-9yrs
(M:F::4:8), who underwent sedation prior to anesthesia using
propofol (2,6-disopropylphenol). The study protocol was
approved by the Clinical Research Ethics Committee of the
Cork Teaching Hospitals. All patients gave prior written
informed consent.
B. Clinical Protocol
A target-effect site propofol infusion was commenced to
provide a target-effect site concentration of 0.5 gml-1. The
target-effect site was brain. The target effect concentration
was increased in 0.5 gml-1 increments every four minutes to
a maximum of 2 gml-1. A period of 4 minutes was allowed
at each concentration level to allow adequate equilibrium to
be achieved across the blood-brain-barrier. For each patient
the data acquisition period was thus 16 minutes.
C. EEG Acquisition
Nineteen channels of EEG were recorded for each patient
using a NicVue digital EEG machine. For the purposes of
analysis and comparison we calculated each quantitative
EEG measure or feature in the following bipolar channels:
'P4-O2'
'P3-O1'
'Fp1-Fp2'
'F3-C3'
'F4-C4'
'F3-F4'
Even numbers by convention refer to electrodes placed on
the right side of the scalp. The occipito-parietal leads were
chosen as alpha rhythm which originates in the posterior
cortex is best seen in these leads. Fp1 Fp2 represents a
prefrontal location likely to be contaminated by EMG
artifact from the frontalis muscle. Activity in the three other
channels is representative of frontal and central lobe activity.
D. Sedation Measurement
The EEG recordings were retrospectively assessed by a
clinical neurophysiologist blinded to the clinical sedation
score. A sedation grade was assigned to each time period of
four minutes corresponding to a set propofol concentration.
The sedation grade was assigned according to preset criteria
set out in table 1. For the purposes of this analysis any four
minute period in which the sedation score was greater then
zero was deemed to be sedated. There were 48 four minute
intervals in twelve patients. Twenty nine had evidence of
sedation with 19 non-sedated. Each record contained 19
EEG channels and was sampled at 250Hz. Records had a
mean duration of 19.8 minutes. The dataset contained a total
of 192 minutes of 19 channel EEG. 116 minutes of EEG
were assigned the labeled sedated while 76 minutes of
EEG were assigned the label non-sedated. Table 2
summarizes the characteristics of each of the recordings.
Patient
#
1
2
3
4
5
6
7
8
9
10
11
12
Record
Length
(mins)
16.3
16.6
16.6
16.2
16.3
17.6
22.6
21.6
17.0
19.8
22.7
34.9
Mean:
19.8
Table 2: Record information
Sedated
Time (mins)
16
4
4
8
12
8
16
12
0
12
12
12
Total:
116
Nonsedated
time (mins)
0
12
12
8
4
8
0
4
16
4
4
4
Total:
76
E. Feature Extraction
The EEG for each channel was low-pass filtered using a
type II Chebyshev IIR filter with a corner frequency of 34Hz
to remove power line noise along with out-of-band noise.
The EEG for each channel was then considered in epochs of
2 seconds duration. The following features were extracted
3189
HS ( X ) =
1
log N f
P ( X ) log
f
Pf ( X ) (1)
Combined
SEF
Acc (%)
77.45
61.28
67.89
70.27
66.21
69.38
60.90
Sens (%)
74.70
62.32
75.28
69.07
65.49
70.98
65.52
Spec (%)
81.67
59.70
56.55
72.10
67.32
66.92
53.81
ROC Area
0.86
0.63
0.71
0.77
0.71
0.77
0.63
Measure
Table 3: Overall performance results for each feature taken individually as well as all features combined together and classified
using a linear discriminant classifier model.
3190
REFERENCES
Patient Independent ROC curve
[1]
100
90
80
[2]
Sensitivity [%]
70
60
50
[3]
40
30
20
[4]
10
0
10
20
30
40
50
60
Specificity [%]
70
80
90
100
[5]
[6]
Acc (%)
Sens (%)
[8]
Spec (%)
82.18
82.18
85.92
58.82
94.96
76.42
20.83
95.21
77.47
73.95
81.01
66.18
59.22
87.29
67.78
55.65
79.92
98.12
98.12
78.24
71.31
99.16
74.58
74.58
10
75.10
78.55
64.71
11
71.85
72.91
68.64
12
75.52
86.87
41.67
[7]
[9]
[10]
[11]
[12]
[13]
[14]
IV. DISCUSSION
A system for the automated estimation of a patients level
of consciousness is presented here. The EEG for each patient
was
dichotomized
by
an
experienced
Clinical
Neurophysiologist into two classes, sedated and non-sedated.
Six quantitative EEG measures per EEG channel were used
in this study. Each quantitative EEG measure has been used
previously in depth of sedation / anesthesia research.
Classifying each epoch using a LD classifier model led to a
sedation sensitivity of 74.70% with an associated specificity
of 81.67%.
3191
[15]
[16]
[17]
[18]