Monitoring Depth of Hypnosis under Propofol General Anaesthesia
Granger Causality and Hidden Markov Models
Nicoletta Nicolaou
1,2,3
and Julius Georgiou
1,2,3
1
Holistic Electronics Research Lab, University of Cyprus, Nicosia, Cyprus
2
Dept. of Electrical and Computer Engineering, University of Cyprus, Nicosia, Cyprus
3
KIOS Research Centre, University of Cyprus, Nicosia, Cyprus
Keywords:
Brain-Computer Interface, EEG, Anaesthesia, Awareness.
Abstract:
Intra-operative awareness is experienced when a patient regains consciousness during surgery. This work
presents a Brain-Computer Interface system that can be used as part of routine surgery for monitoring the
patient state of hypnosis in order to prevent intra-operative awareness. The underlying state of hypnosis is
estimated using causality-based features extracted from the spontaneous electrical brain activity (EEG) of the
patient and a probabilistic classification framework (Hidden Markov Models). The proposed method is applied
to EEG activity from 20 patients under propofol anaesthesia. The mean discrimination performance obtained
was 98% and 85% for wakefulness and anaesthesia respectively, with an overall performance accuracy of 92%.
The use of a probabilistic framework increases the anaesthetist’s confidence on the estimated state of hypnosis
based on the marginal probabilities of the underlying state.
1 INTRODUCTION
Intra-operative awareness occurs in approximately
0.1-0.8 % of surgical patients (Bruhn et al., 2006).
The real incidence of awareness, however, is likely
to be much higher due to the amnesic effect of cer-
tain anaesthetics resulting in some patients having
no recollection of regaining awareness. The impor-
tance of monitoring depth of anaesthesia in order to
prevent intra-operative awareness is apparent consid-
ering that, given the estimated 234.2 million major
surgical procedures undertaken annually worldwide
(Weiser et al., 2008), at least 1,873,600 people are
likely to have experienced intraoperative awareness.
The use of the electrical brain activity (EEG) is
currently the preferred method for monitoring anaes-
thetic depth. Devices that monitor the patient state
of hypnosis during anaesthesia are a form of a Brain-
Computer Interface (BCI) system. Even though the
idea of using the spontaneous patient EEG for mon-
itoring purposes is not new (McEwen et al., 1975),
a relatively small number of devices are currently
commercially available for this purpose. In these de-
vices, the patient’s spontaneous EEG activity is con-
tinuously monitored for changes that signal regaining
of awareness. In incidences of intra-operative aware-
ness the patient cannot communicate this to the anaes-
thetist due to the co-administration of neuromuscular
blockers with the anaesthetic agents. Thus, the patient
is essentially put in a chemically-induced ‘locked-in’
state whereby communication via conventional means
is impaired. Despite the existence of commercial
EEG-based devices for monitoring patient state dur-
ing anaesthesia, they are not considered as part of rou-
tine anaesthetic practice in the majority of hospitals
worldwide. This is mainly attributed to issues relat-
ing to robustness and inter-subject variability (Voss
and Sleigh, 2007). Inter-subject variability in partic-
ular is an issue that cannot be addressed with current
systems, as the systems have a universal state indi-
cator scale (from 0-100; 100: fully conscious, 40-60:
surgical anaesthesia, 0: no activity) that cannot be cal-
ibrated for each individual patient and, thus, does not
take into account patient specifics.
In this study the anaesthetic-induced EEG changes
are modelled using causality-based features and the
underlying state (wakefulness / anaesthesia) is es-
timated using a probabilistic framework (Hidden
Markov Models). The use of causality features in-
creases the system robustness, as it has been shown
that such features capture general mechanisms of
anaesthetic administration regardless of the particu-
lar anaesthetic protocol (Nicolaou et al., 2012; Barrett
et al., 2012), and with a high discriminative ability
256
Nicolaou N. and Georgiou J..
Monitoring Depth of Hypnosis under Propofol General Anaesthesia - Granger Causality and Hidden Markov Models.
DOI: 10.5220/0004679402560261
In Proceedings of the International Congress on Neurotechnology, Electronics and Informatics (BrainRehab-2013), pages 256-261
ISBN: 978-989-8565-80-8
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
(Nicolaou and Georgiou, 2013). The use of a proba-
bilistic framework in such a system is advantageous,
as the anaesthetist can also assess the reliability of
a state estimation through the marginal state proba-
bilities. Learning an HMM model for each patient
also ensures that the system is calibrated for each pa-
tient and the estimated states are not affected by inter-
subject variability.
2 METHODS
2.1 Dataset
The data is a subset of EEG data collected from pa-
tients during surgery at the Nicosia General Hospi-
tal, Cyprus. In this particular study data from 20
male patients (mean age 41.8 ± 20.6) were analysed.
The study has been approved by the Cyprus National
Bioethics Committee and patients involved gave writ-
ten informed consent for their participation. The ex-
perimental protocol details are described elsewhere
(e.g. see (Nicolaou et al., 2012)). In summary, anaes-
thesia was induced with a propofol bolus and main-
tained with constant intravenous propofol administra-
tion. In most patients this was titrated with an in-
travenous administration of remifentanil hydrochlo-
ride. Lungs were ventilated with an air-oxygen or air-
oxygen-N
2
O mixture. During surgery boluses of neu-
romuscular blocking agents and other drugs, such as
antibiotics, were administered as required. EEG data
were obtained from 19 electrodes based on the inter-
national 10/20 system, with a sampling rate of 256
Hz. Data recording was performed throughout the en-
tire surgical duration (awake pre-induction, induction,
surgical anaesthesia and recovery of consciousness).
Since the exact point at which loss of consciousness
occurs after patient induction is not known, the point
at which the anaesthetic bolus was administered was
considered instead. Recovery of consciousness was
defined as the point at which the patient responded
(either via voluntary muscular movement or a verbal
response) to verbal commands or tactile stimuli by the
anaesthetist.
2.2 Granger Causality
Granger Causality (GC) is defined by Wiener as fol-
lows: ‘for two simultaneously measured signals, if
one can predict the first signal better by incorporat-
ing the past information from the second signal than
using only information from the first one, then the
second signal can be called causal to the first one’
(Wiener, 1956). Mathematically, causality was de-
fined by Granger through the use of regression: for
two time series, X
1
, and X
2
, if X
1
is influenced by
X
2
, then the addition of past values of X
2
in the re-
gression of X
1
will improve its prediction (Granger,
1980). The performance of the prediction can be as-
sessed through the variances of the fitted regression
models. Thus, GC is defined as:
GC
X
2
X
1
= ln
σ
2
X
1
/X
1
σ
2
X
1
/X
1
X
2
(1)
where σ
2
X
1
/X
1
and σ
2
X
1
/X
1
X
2
are the variances of the
regression errors, e
x
j
and e
x
1
x
2
respectively, obtained
from the following (auto)regression models ( j = 1, 2):
x
j
(t) =
P
i=1
a
ix
j
x
j
(t i) + e
x
j
(t) (2)
x
1
(t) =
P
i=1
a
ix
1
x
2
x
1
(t i)+
P
i=1
b
ix
1
x
2
x
2
(t i)+e
x
1
x
2
(t)
(3)
The variables a
ix
j
, a
ix
1
x
2
and b
ix
1
x
2
are the regression
coefficients for models of order P. The main idea is
that if the prediction of X
1
is improved by using past
values of X
2
in its prediction, then σ
2
X
1
/X
1
X
2
< σ
2
X
1
/X
1
and, therefore GC
X
2
X
1
increases. If, however, the
past of X
2
does not improve the prediction of X
1
, then
GC
X
2
X
1
will be close to zero. Therefore, by defini-
tion, GC
X
2
X
1
= 0 when there is no causality between
the signals, and GC
X
2
X
1
> 0 otherwise. Similarly,
we can also define GC in the opposite direction. A
number of factors must be taken into account when
considering GC-based analysis, such as the choice of
the regression model order, data stationarity and data
filtering (more details can be found in (Bressler and
Seth, 2010)).
2.3 Hidden Markov Models
Hidden Markov Models (HMMs) belong to the family
of Bayesian networks (Rabiner, 1986). They are used
to model systems that are assumed to be a Markov
process with states that are not directly visible to the
observer. The observer can observe only the output of
the system. However, the sequence of outputs gives
some information about the invisible sequence of dis-
crete states, as each state has a probability distribu-
tion associated with it. Transitions between states
are also probabilistically described through the tran-
sition probability matrix. The HMM model makes
two assumptions: (i) the Markov assumption, which
states that the current state is dependent only on the
previous state; and (ii) the independence assumption,
MonitoringDepthofHypnosisunderPropofolGeneralAnaesthesia-GrangerCausalityandHiddenMarkovModels
257
which states that the output observed at a particular
time is independent of past outputs and states and de-
pends only on the current state. An HMM can also
perform learning, whereby the model parameters that
best describe a process can be estimated from a set of
examples from the particular process. Training can be
both supervised or unsupervised. In supervised learn-
ing, the model outputs are equated to the inputs, and
the outputs to the corresponding states. The model pa-
rameters can then be estimated using maximum likeli-
hood estimation. In our particular case, the output ob-
servations correspond to the GC-based features, while
the unobserved underlying states are ‘wakefulness’
and ‘anaesthesia’. The HMM is then used to answer
the question: ”What is the most likely sequence of
wakefulness/anaesthesia states that could have gener-
ated the observed GC values?”.
2.4 Methodology
The ability to separate consciousness and anaesthesia
using the spontaneous changes in the patient’s EEG
activity was investigated. The particular methodology
consists of the following steps:
1. For each patient, the 19 electrodes were split into
the following five grids: left frontal (LF: elec-
trodes Fp1, F7, F3, T3, C3), right frontal (RF:
Fp2, F8, F4, C4, T4), left posterior (LP: T5, P3,
O1), right posterior (RP: T6, P4, O2), and mid-
line (Z: Fz, Cz, Pz). The average EEG activity
over each of the ve grids was then estimated.
The particular groupings were chosen such that
broad areas corresponding to frontal and poste-
rior activity were obtained, as fronto-posterior
interactions appear to play an important role in
(un)consciousness.
2. For each patient, the continuous averaged EEG
data obtained over each of the five grids were win-
dowed into 2-s non-overlapping segments.
3. Using the manual markers in the EEG record, the
windows corresponding to wakefulness (class A)
and anaesthesia (class B) were identified.
4. For each 2-s segment, pairwise fronto-posterior
GC features were estimated, resulting into the fol-
lowing 4-dimensional feature vector:
F
i
C
= [GC
i
LFLP
,GC
i
RFLP
,GC
i
LFRP
,GC
i
RFRP
]
where C {A,B} corresponds to one of the two
classes, and i = 1,...,N
C
denotes the i
th
2-s seg-
ment from all the available segments of each class
(N
C
). The order of the fitted regression models
was set to 6.
5. The ‘Bayes Net Toolbox’ for Matlab is used for
HMM modelling (Murphy, 2001). An HMM
model is trained using 40% of the available data
for each class. The HMM outputs are modelled
as continuous Gaussians. The estimated predic-
tion probabilities are then obtained for all avail-
able data.
The specific parameters, such as duration of the win-
dows and the order of the fitted regression mod-
els were based on previous investigations (Nicolaou
et al., 2012). The performance of the HMM for each
patient is then estimated via the average specificity
(4), sensitivity (5) and accuracy (6) estimated over
B = 50 bootstrap repetitions. T
ru
P(T
ru
N) is the num-
ber of true positives (negatives), and T
ot
P(T
ot
N) is the
total number of ‘ground truth’ positive (negative) ex-
amples of each class.
SP =
T
ru
P
T
ot
P
(4)
SE =
T
ru
N
T
ot
N
(5)
AC =
1
2
1
B
B
b=1
SP
b
+
1
B
B
b=1
SE
b
!
(6)
3 RESULTS AND DISCUSSION
Figures 1 and 2 show examples of the marginal proba-
bilities for wakefulness and anaesthesia obtained from
one of the 50 bootstrap repetitions and for two ran-
domly chosen patients (S20 and S2 respectively). For
visualisation purposes a moving average filter (n =
10) was applied to the marginal probabilities. The
probabilistic framework allows the anaesthetist to as-
sociate a likelihood to a particular decision. It can
be seen from the figures that the marginal probabil-
ities track the transitions between wakefulness and
anaesthesia well: as expected, the marginal probabil-
ity for wakefulness is close to 1 prior to anaesthetic in-
duction and after recovery of consciousness, while it
drops below 0.5 during surgical anaesthesia (and vice
versa for the marginal probability for anaesthesia).
This implies that the Granger Causality-based fea-
tures provide a good representation of the two states
and that the HMM is able to track the successions of
the two states successfully. The presence of artefacts,
such as the use of diathermy, may cause brief distor-
tion in state estimations. This can be seen in figure 1,
where artefacts caused by diathermy introduce spu-
rious brief changes in the estimated marginal proba-
bilities. In such cases, it is easy for the anaesthetist
NEUROTECHNIX2013-InternationalCongressonNeurotechnology,ElectronicsandInformatics
258
Figure 1: Estimated state probabilities for wakefulness (black) and anaesthesia (maroon) for patient S20. The state with a
probability greater than ‘0.5’ is the classified state. Large peaks in the marginal probabilities during anaesthesia are caused by
diathermy artefacts, while the peaks after induction are caused by tracheal intubation.
Figure 2: Estimated state probabilities for wakefulness (black) and anaesthesia (maroon) for patient S2. The state with a
probability greater than ‘0.5’ is the classified state. Large peaks in the marginal probabilities during anaesthesia are caused by
diathermy artefacts.
to assess whether such changes are true or whether
they are indeed spurious, e.g. if diathermy is being
utilised at the particular moment. However, the pres-
ence of artefacts does not necessarily induce spurious
MonitoringDepthofHypnosisunderPropofolGeneralAnaesthesia-GrangerCausalityandHiddenMarkovModels
259
state changes, as the marginal probabilities may not
go over/under the 0.5 threshold; this can also be seen
in figures 1 and 2. Moreover, changes in the marginal
probabilities that are persistent rather than transient
is another cause of alert for the anaesthetist, as such
changes could be another indication that the patient
may be regaining consciousness.
Table 1 shows the mean performance of the pro-
posed method for each patient and overall. Patient-
wise, a mean sensitivity, specificity and accuracy of
0.98, 0.85 and 0.92 is obtained. This is comparable to
results from other studies, which range from 64-93%
(for more details see (Nicolaou et al., 2012) and ref-
erences within). For 12 patients the overall accuracy
is more than 90%. The lower specificity (performance
for anaesthesia) compared to sensitivity (performance
for wakefulness) is expected if we take into account
that no artefact removal has been performed, thus
some misclassification due to artefacts during surgery
is expected. For 3 patients the mean specificity is be-
tween 0.63-0.66. This could be mainly attributed to
the small number of training features resulting from
the small number of available features for wakeful-
ness for the particular patients. Thus, this has a neg-
ative effect on the generalisation ability of the HMM
classifier for the anaesthesia state, given that the small
Table 1: Mean performance of GC-based HMM classifica-
tion for 20 patients (S1,...S20). SE: Sensitivity, SP: Speci-
ficity, AC: Accuracy. The maximum (best) classification is
indicated by ’1’ (corresponding to 100%).
PATIENT SE SP AC
S1 1.00 0.98 0.99
S2 1.00 0.94 0.97
S3 0.89 0.93 0.91
S4 1.00 0.93 0.97
S5 1.00 0.93 0.97
S6 1.00 0.66 0.83
S7 1.00 0.96 0.98
S8 1.00 0.87 0.94
S9 1.00 0.81 0.91
S10 0.95 0.88 0.92
S11 0.95 0.79 0.87
S12 0.94 0.65 0.80
S13 0.98 0.94 0.96
S14 0.95 0.77 0.86
S15 1.00 0.63 0.82
S16 0.96 0.78 0.87
S17 1.00 0.90 0.95
S18 1.00 0.78 0.89
S19 0.91 0.87 0.89
S20 0.99 0.92 0.96
TOTAL 0.98 0.85 0.92
number of training features cannot be expected to cap-
ture all feature attributes associated with anaesthesia.
An important advantage of the proposed method-
ology is its clinical applicability. Individual patient
variability is taken into account through calibration
of the BCI system for each patient, as opposed to
commercially available systems that employ a univer-
sal 0-100 scale without system calibration to patient
specifics. This calibration is also possible due to the
ability of HMMs to learn incrementally, thus facili-
tating a more higher-level learning through the addi-
tion of new information. The probabilistic framework
adds credibility by associating a probability likeli-
hood to each decision. This strengthens the anaes-
thetist’s decisions regarding the assessment of the un-
derlying patient state of hypnosis by allowing a given
degree of certainty to their actions.
The proposed methodology has some limitations.
These are mainly related to the particular features
utilised: Granger Causality (GC) has been the recip-
ient of some criticism regarding the effects of sta-
tionarity, volume conduction, filtering and regression
model order to the estimated GC values (Bressler and
Seth, 2010; Florin et al., 2010). However, the na-
ture of causality itself implies that causality-based
measures likely capture more general mechanisms of
anaesthetic action. If these specific issues are taken
into account causality can, thus, constitute a robust
and reliable feature of anaesthetic-induced changes
in spontaneous brain activity. A more extensive dis-
cussion on GC, general limitations posed by GC and
potential solutions can be found in (Nicolaou et al.,
2012).
In addition, as previously mentioned, no artefact
removal was performed in this study. Prior to anaes-
thetic induction the EEG signals may contain artefacts
originating both from the patient, such as body move-
ment, eye blinks and speech, and external sources,
such as placement of monitoring equipment by the
hospital staff. During induction the EEG is usually
contaminated by artefacts due to tracheal intubation.
During anaesthesia the administered anaesthetics and
muscle relaxants ensure that there is no muscle, eye
or movement activity from the patient. Artefacts
during anaesthesia are mainly caused by the use of
diathermy equipment, but some artefacts due to other
surgical stimuli could also be present. The latter is not
likely as we excluded surgical procedures performed
in close proximity to the EEG sensors, such as ear-
nose-throat surgeries. Despite having performed no
artefact removal, it is not likely that discrimination
between wakefulness and anaesthesia is performed
solely due to the presence of these artefacts. On the
contrary, the performance of the proposed method
NEUROTECHNIX2013-InternationalCongressonNeurotechnology,ElectronicsandInformatics
260
could be improved by removal of artefacts prior to
feature estimation, as the differences between the two
states would be more prominent.
4 CONCLUSIONS
The spontaneous EEG activity of anaesthetised pa-
tients can be used to assess their underlying state of
hypnosis. This type of a BCI system can revolutionise
routine surgery and aid towards avoidance of intra-
operative awareness. The use of causality-based fea-
tures in a probabilistic framework adds to the reliabil-
ity and robustness of such a system. State decisions
are supported by related probabilities, thus strength-
ening the weight of each individual state assessment
by both the system and the anaesthetist. Future work
will investigate whether causality features obtained
from individual electrode pairs can provide similar
discriminatory ability in order to eliminate the need
for a full set of EEG sensors to capture anaesthetic-
induced changes in causality.
ACKNOWLEDGEMENTS
The authors would like to thank Dr. Pandelitsa
Alexandrou and Dr. Saverios Hourris (Depr. of
Anaesthesia, Nicosia General Hospital, Cyprus), the
hospital staff and the anonymous volunteers who par-
ticipated in this study. This work was co-funded by
the Republic of Cyprus and the European Regional
Development Fund through the Cyprus Research Pro-
motion Foundation (DESMI 2008). Grants: ‘New In-
frastructure Project/Strategic/0308/26’ and ‘DIDAK-
TOR/DISEK /0308/20’.
REFERENCES
Barrett, A., Murphy, M., Bruno, M.-A., Noirhomme, Q.,
Boly, M., Laureys, S., and Seth, A. (2012). Granger
causality analysis of steady-state electroencephalo-
graphic signals during propofol-induced anaesthesia.
PLoS ONE, 7(1):e29072.
Bressler, S. and Seth, A. (2010). Wienergranger causal-
ity: A well established methodology. NeuroImage,
58(2):323–329.
Bruhn, J., Myles, P., Sneyd, R., and Struys, M. (2006).
Depth of anaesthesia monitoring: what’s available,
what’s validated and what’s next? British Journal of
Anaesthesia, 97:85–94.
Florin, E., Gross, J., Pfeifer, J., Fink, G., and Timmermann,
L. (2010). The effect of filtering on granger causality
based multivariate causality measures. NeuroImage,
50:577588.
Granger, C. (1980). Testing for causality: a personal view-
point. J Econ Dyn Control, 2:329–352.
McEwen, J., Anderson, G., Low, M., and Jenkins, L.
(1975). Monitoring the level of anesthesia by au-
tomatic analysis of spontaneous eeg activity. IEEE
Trans on Biomedical Engineering, BME-22(4):299–
305.
Murphy, K. (2001). The Bayes Net Toolbox for Matlab.
Computing Science and Statistics: Proceedings of the
Interface, 33.
Nicolaou, N. and Georgiou, J. (2013). Neural networkbased
classification of anesthesia/awareness using granger
causality features. Clinical EEG and Neuroscience,
doi:10.1177/1550059413486271.
Nicolaou, N., Hourris, S., Alexandrou, P., and Georgiou,
J. (2012). EEG-based automatic classification of
‘awake’ versus ‘anesthetized’ state in general anesthe-
sia using granger causality. PLoS ONE, 7(3):e33869.
Rabiner, L. (1986). An introduction to hidden Markov mod-
els. IEEE ASSP Magazine, 3(1):4–16.
Voss, L. and Sleigh, J. (2007). Monitoring consciousness:
the current status of EEG-based depth of anaesthesia
monitors. Best Pract Res Clin Anaesthesiol, 21:313–
325.
Weiser, T., Regenbogen, S., Thompson, K., Haynes, A.,
Lipsitz, S., Berry, W., and Gawande, A. (2008). An
estimation of the global volume of surgery: a mod-
elling strategy based on available data. The Lancet,
372:139–144.
Wiener, N. (1956). The theory of prediction, volume 1.
McGraw-Hill, New York.
MonitoringDepthofHypnosisunderPropofolGeneralAnaesthesia-GrangerCausalityandHiddenMarkovModels
261