Conceived and designed the experiments: NN SH PA JG. Performed the experiments: NN SH PA. Analyzed the data: NN JG. Contributed reagents/materials/analysis tools: NN. Wrote the paper: NN JG.
The authors have declared that no competing interests exist.
General anesthesia is a reversible state of unconsciousness and depression of reflexes to afferent stimuli induced by administration of a “cocktail” of chemical agents. The multi-component nature of general anesthesia complicates the identification of the precise mechanisms by which anesthetics disrupt consciousness. Devices that monitor the depth of anesthesia are an important aide for the anesthetist. This paper investigates the use of effective connectivity measures from human electrical brain activity as a means of discriminating between ‘awake’ and ‘anesthetized’ state during induction and recovery of consciousness under general anesthesia.
Granger Causality (GC), a linear measure of effective connectivity, is utilized in automated classification of ‘awake’ versus ‘anesthetized’ state using Linear Discriminant Analysis and Support Vector Machines (with linear and non-linear kernel). Based on our investigations, the most characteristic change of GC observed between the two states is the sharp increase of GC from frontal to posterior regions when the subject was anesthetized, and reversal at recovery of consciousness. Features derived from the GC estimates resulted in classification of ‘awake’ and ‘anesthetized’ states in 21 patients with maximum average accuracies of 0.98 and 0.95, during loss and recovery of consciousness respectively. The differences in linear and non-linear classification are not statistically significant, implying that GC features are linearly separable, eliminating the need for a complex and computationally expensive non-linear classifier. In addition, the observed GC patterns are particularly interesting in terms of a physiological interpretation of the disruption of consciousness by anesthetics. Bidirectional interaction or strong unidirectional interaction in the presence of a common input as captured by GC are most likely related to mechanisms of information flow in cortical circuits.
GC-based features could be utilized effectively in a device for monitoring depth of anesthesia during surgery.
General anesthesia is a drug-induced reversible state of unconsciousness and depression of reflexes to afferent stimuli
An insight into the complex process of general anesthesia can be obtained through studying how the administration of this chemical “cocktail” affects the observed brain activity. The action of the anesthetic agents causes measurable effects on the brain activity, which can be observed through methods such as the electroencephalogram (EEG). The use of EEG monitors during anesthesia has allowed the identification of some characteristics that are related to the administration of anesthetic agents. For example, anesthesia causes characteristic changes in the spectral content of the EEG: as the depth of anesthesia increases, the faster α (8–12 Hz) and β (12.5–30 Hz) brain rhythms are replaced by slower δ (1.5–3.5 Hz) and θ (3.5–7.5 Hz) activity. In very deep anesthesia the EEG may develop a peculiar pattern of activity known as burst suppression, during which alternating periods of normal to high activity and low voltage (or even isoelectricity) are observed
In addition to studying the mechanisms of general anesthesia and the effects of different anesthetic agents, the use of brain activity has an additional and more direct clinical application: it provides a means of monitoring the depth of anesthesia (DOA) during surgery. The combination of agents and the doses at which these are administered are very much dependent on patient characteristics and surgery requirements, therefore each case is unique. As a result, there are no direct instructions that the anesthetist can follow, but only rough guidelines. Thus, DOA monitors provide an objective method of assessing the state of hypnosis of the patient and provide a useful and welcome aid for the anesthetist. The main concerns of the anesthetist are over- and under-dose of anesthetic agents. Both could have serious implications for the patient. Over long periods over-administration can be costly in terms of agent usage and because of increased patient recovery time. In the worst case, overdosage can lead to death. Underdosage can lead to regaining of consciousness during surgery, which is extremely traumatic. Costs involved with underdosage are related to post-traumatic stress therapy and compensation claims. Intra-operative awareness has been confirmed in a number of cases, with incidence ranging from 0.11–0.8%
Currently EEG-based DOA monitors are being introduced for routine patient monitoring during surgery. The most commonly used commercially available devices include the BIS® monitor (Aspect Medical Systems, Natick, MA)
This is due to the fact that the operation of current monitors is based on features that are characteristic of the observed changes in the EEG activity, which may not be a direct reflection of the actual physiological process underlying general anesthesia and which are not unique to anesthetic-induced LOC. However, the measures utilized must be based on ‘neurobiologic phenomena that represent the
In this work Granger Causality (GC), a measure quantifying causal interactions between two time series, is utilized as a feature for discriminating awake from anesthetized state. The main focus of the study was the use of GC as a discrimination feature to capture reversible changes with loss and recovery of consciousness, regardless of the anesthetic protocol used. Our previous investigations showed that GC captures such reversible anesthetic-induced changes in brain activity
The dataset used in this study was collected from 21 male patients (mean age 37.6±19.1) who underwent routine general surgery at the Nicosia General Hospital, Cyprus. The administration of general anesthesia was not confined to a particular anesthetic regime. The study was approved by the National Bioethics Committee of Cyprus and the patients gave written informed consent for their participation. Participants were not previously taking any medication that influences the central nervous system and were of normal weight. One patient was diagnosed with multiple sclerosis (very early stage). However, the data of this patient were not excluded from the study as the findings were similar with other patients. General anesthesia was induced by the on duty anesthetist using the regular procedures of the hospital. Standard monitoring devices, including pulse oximetry, electrocardiogram, and non-invasive blood pressure, were utilized. All patients were preoxygenated via a face mask prior to anesthesia induction with a Diprivan (propofol 1%, 10 mg/ml) bolus. The induction dose varied from 2 mg/kg to 4 mg/kg depending on patient characteristics. During induction some patients also received boluses of neuromuscular blocking agents (cisatracurium, rocuronium, or atracurium) and analgesic drugs. Depending on patient characteristics and surgery requirements maintenance of anesthesia was achieved with an intravenous administration of propofol at concentrations ranging between 20–50 ml/h (200–500 mg/h). For 2 patients (S12 and S15) maintenance was performed with an inhalational administration of sevoflurane (1–2%). In most patients this was titrated with an intravenous administration of remifentanil hydrochloride (Ultiva®; 2 mg, dissolved in 40 ml) throughout surgery at a rate ranging between 2–15 ml/h (0.1–0.75 mg/h). Following induction of anesthesia the patients' trachea was intubated and surgery commenced. Lungs were ventilated with an air-oxygen or air-oxygen-N2O mixture. During surgery boluses of neuromuscular blocking agents and other drugs, such as antibiotics, were administered as required and depending on surgery requirements.
EEG data were collected using the TruScan32 system (Deymed Diagnostic) at a sampling rate of 256 Hz. Electrodes were placed at positions Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1 and O2, according to the International 10/20 system, and were recorded with an FCz reference. No filtering was performed during or after data collection; this is to ensure that the timing relations on which GC depends on are not disrupted by the introduction of causal artifacts from filtering
The main function of a depth of anesthesia (DOA) monitor is to alert the anesthetist when a subject becomes aware during surgery. Therefore, a minimal requirement for a DOA monitor is the ability to distinguish between the two states ‘Awake’ and ‘Anesthetized’. The ability to classify these two states using GC as a feature was investigated following the methodology described below.
Data from 21 subjects were available for analysis (S1–S21). Using the dataset described above, segments of a few minutes duration corresponding to the two classes were extracted from the continuous EEG recordings. Such data is available both at initial loss of consciousness at induction, and recovery of consciousness at the end of surgery. The segments were extracted based on the manual markers inserted in the EEG record during surgery, indicating anesthetic induction and recovery of consciousness. Loss of consciousness after anesthetic induction is patient-dependent and occurs 10–30 s after administration of the anesthetic bolus. In the following analysis we did not use the first 5 minutes of data after the marker for anesthetic induction; this ensured that the data used corresponded to the patient being fully unconscious, and did not contain any artifacts caused from tracheal intubation.
The original data space is 19-dimensional (number of electrodes). In order to reduce this, five brain areas were defined as the average activity of specified electrode grids. The five brain areas defined were: left frontal (LF: electrodes 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 midline (Z: Fz, Cz, Pz). The rationale behind these groupings was that fronto-posterior interactions appear to play an important role in (un)consciousness, thus we performed grouping of activity from frontal and posterior areas in order to investigate such fronto-posterior interactions. Electrode impedance is measured automatically by the EEG hardware. Electrodes with high impedance resulting from bad contact or no contact were subsequently excluded from estimation of the average activity.
The great interest in investigations of causal relationships, particularly when dealing with neurophysiological data, has motivated the development of measures that capture such relationships. One such measure is Granger Causality (GC). GC has been developed explicitly to allow inferences about causality between two time series to be made
In the univariate case,
Similarly, for the bivariate AR model:
Let us denote the variance of the prediction errors as
Other considerations during feature extraction included:
Stationarity: Granger Causality was estimated over 4-second EEG windows (window sliding by 1-s). Segments with such short duration were chosen for two reasons. Firstly, to identify cases of impending awareness as quickly as possible; and, secondly, this is common practice in EEG analysis to ensure the stationarity of the EEG segments analyzed
AR Modeling: The Durbin-Watson test was used to assess the residual variance
Artifacts: The main sources of artifacts during anesthesia are artifacts during tracheal intubation at anesthetic induction, and diathermy noise during surgery. We removed intubation artifacts by excluding the first 5 minutes following anesthetic induction; this also served the purpose of ensuring that the patient was fully unconscious. Segments which were contaminated with diathermy were excluded from further analysis. We also investigated the application of a 50-Hz notch filter for removing line noise (using the Matlab® function ‘
To estimate the GC values for each 4-s EEG segment,
Based on our preliminary investigations, the most characteristic change of the GC index was the significant increase of GC from frontal to posterior regions when the subject was anesthetized
An important consideration in EEG analysis is volume conduction. Even though the Laplacian transform offers a solution to this problem, this acts as a bandpass spatial filter, which ‘may remove genuine source activity associated with very low spatial frequencies’
Classification performance was obtained for each subject over B = 200 bootstrap repetitions (sampling with replacement). For each patient the number of samples (windows) available for the ‘awake’ and ‘anesthetized’ classes were
An additional consideration is that a patient awaking from surgery does not regain full alertness until some time afterwards; this time frame is very much dependent on the rate at which each person is able to metabolize the administered drugs. This implies that the awareness state of a patient at ROC could be more similar to the awareness state of the patient in the case that awareness is experienced during surgery. In order to investigate whether wakefulness prior to anesthetic administration and wakefulness at the end of surgery differed, we performed two separate investigations utilizing data extracted around (a) the marker for anesthetic administration; and (b) the marker for recovery of consciousness. Therefore, we performed separate classifications with different classifiers for these two cases.
Maroon line: posterior→frontal direction. Blue line: frontal→posterior direction. GC between (a) left frontal – left posterior, (b) right frontal – left posterior, (c) left frontal – right posterior; (d) right frontal – right posterior. Shaded areas: mean GC ± standard deviation. An increase in fronto→posterior GC after anesthesia induction is observed. Vertical lines indicate anesthetic induction (AI) and recovery of consciousness (ROC). As expected, the fronto→posterior GC returns to baseline at recovery of consciousness. Subjects with no GC over the right posterior area due to bad electrode contact were excluded from the average (2 patients). X-axis in arbitrary samples.
50-second segments of ‘Awake’ (pre-LOC, post-ROC) and ‘Anesthetized’ (mean GC for post-LOC and pre-ROC) states. (a) GCLF→LP, (b) GCRF→LP, (c) GCLF→RP, and (d) GCRF→RP. The differences in GC between ‘Awake’ and ‘Anesthetized’ states are statistically significant (ANOVA F-test, α = 0.05, p = 0).
(a) GCLF↔LP for patient S11 at LOC. (b) GCRF↔LP for patient S1 at ROC. (c) GCLF↔RP for patient S13 at LOC. (d) GCRF↔RP for patient S17 at ROC. (e) GCRF↔LP for patient S8 at LOC. (f) GCLF↔RP for patient S21 at ROC. Vertical line indicates anesthetic administration ((a), (c), (e)), and recovery of consciousness ((b), (d), (f)). Outlier GC values due to the presence of artifacts in the raw EEG signal are also visible in (a) and (b).
Specificity | Sensitivity | Accuracy | |||||||
Subject | SVMNL | SVML | LDA | SVMNL | SVML | LDA | SVMNL | SVML | LDA |
|
1.000 | 0.999 | 1.000 | 0.989 | 0.992 | 0.980 | 0.995 | 0.996 | 0.990 |
|
0.948 | 0.890 | 0.849 | 0.979 | 0.923 | 0.951 | 0.964 | 0.907 | 0.900 |
|
0.902 | 0.947 | 0.962 | 0.928 | 0.835 | 0.827 | 0.915 | 0.891 | 0.896 |
|
1.000 | 1.000 | 1.000 | 0.999 | 0.999 | 0.983 | 1.000 | 1.000 | 0.992 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.997 | 1.000 | 1.000 | 0.998 |
|
0.891 | 0.884 | 0.868 | 0.953 | 0.925 | 0.930 | 0.922 | 0.905 | 0.899 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 |
|
0.968 | 0.949 | 0.953 | 0.985 | 0.977 | 0.973 | 0.977 | 0.963 | 0.963 |
|
1.000 | 1.000 | 1.000 | 0.992 | 0.971 | 0.864 | 0.996 | 0.986 | 0.932 |
|
1.000 | 1.000 | 1.000 | 0.986 | 0.980 | 0.977 | 0.993 | 0.990 | 0.989 |
|
1.000 | 1.000 | 1.000 | 0.999 | 1.000 | 0.966 | 0.999 | 1.000 | 0.983 |
|
0.997 | 0.835 | 0.845 | 0.975 | 0.774 | 0.764 | 0.986 | 0.805 | 0.805 |
|
1.000 | 1.000 | 1.000 | 0.978 | 0.980 | 0.970 | 0.989 | 0.990 | 0.985 |
|
0.968 | 0.974 | 0.973 | 0.985 | 0.975 | 0.990 | 0.977 | 0.975 | 0.981 |
|
0.997 | 0.996 | 0.985 | 0.981 | 0.981 | 0.965 | 0.992 | 0.989 | 0.975 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.988 | 1.000 | 1.000 | 0.994 |
|
0.925 | 0.925 | 0.838 | 0.978 | 0.978 | 0.989 | 0.951 | 0.951 | 0.913 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.957 | 1.000 | 1.000 | 0.979 |
|
0.980 | 0.940 | 0.940 | 0.987 | 0.962 | 0.941 | 0.984 | 0.951 | 0.941 |
|
0.976 | 0.978 | 0.988 | 0.968 | 0.962 | 0.933 | 0.972 | 0.970 | 0.960 |
|
|
|
|
|
|
|
|
|
|
: Patient administered a very small quantity of neuromuscular blocking agent (<4 mg) at induction only to facilitate tracheal intubation.
: Maintenance with sevoflurane.
Performance estimated with nonlinear and linear Support Vector Machine (SVMNL and SVML respectively), and Linear Discriminant Analysis (LDA). ‘TOTAL’ indicates the average performance over all patients.
Specificity | Sensitivity | Accuracy | |||||||
Subject | SVMNL | SVML | LDA | SVMNL | SVML | LDA | SVMNL | SVML | LDA |
|
0.996 | 0.990 | 0.945 | 0.920 | 0.902 | 0.771 | 0.958 | 0.946 | 0.858 |
|
0.878 | 0.794 | 0.817 | 0.869 | 0.762 | 0.743 | 0.873 | 0.778 | 0.780 |
|
0.943 | 0.930 | 0.931 | 0.989 | 0.987 | 1.000 | 0.966 | 0.959 | 0.965 |
|
0.917 | 0.842 | 0.845 | 0.899 | 0.721 | 0.715 | 0.908 | 0.782 | 0.780 |
|
0.797 | 0.757 | 0.750 | 0.853 | 0.918 | 0.944 | 0.825 | 0.838 | 0.847 |
|
0.761 | 0.701 | 0.704 | 0.734 | 0.563 | 0.563 | 0.748 | 0.632 | 0.633 |
|
0.988 | 0.989 | 1.000 | 0.993 | 0.992 | 0.977 | 0.990 | 0.990 | 0.988 |
|
0.950 | 0.797 | 0.823 | 0.846 | 0.768 | 0.758 | 0.898 | 0.782 | 0.791 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.978 | 1.000 | 1.000 | 0.989 |
|
0.999 | 0.991 | 1.000 | 0.909 | 0.911 | 0.786 | 0.954 | 0.951 | 0.893 |
|
1.000 | 1.000 | 1.000 | 0.995 | 0.995 | 0.992 | 0.998 | 0.997 | 0.996 |
|
1.000 | 1.000 | 1.000 | 0.962 | 0.959 | 0.872 | 0.981 | 0.980 | 0.936 |
|
1.000 | 0.998 | 1.000 | 0.976 | 0.975 | 0.902 | 0.988 | 0.987 | 0.951 |
|
0.939 | 0.925 | 0.949 | 0.928 | 0.902 | 0.910 | 0.934 | 0.913 | 0.929 |
|
1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.989 | 1.000 | 1.000 | 0.984 |
|
1.000 | 1.000 | 1.000 | 0.948 | 0.948 | 0.962 | 0.974 | 0.974 | 0.981 |
|
0.974 | 0.973 | 0.999 | 0.923 | 0.896 | 0.877 | 0.949 | 0.935 | 0.938 |
|
1.000 | 1.000 | 1.000 | 0.994 | 0.988 | 0.981 | 0.997 | 0.994 | 0.991 |
|
0.987 | 0.998 | 1.000 | 0.975 | 0.953 | 0.977 | 0.981 | 0.976 | 0.988 |
|
0.931 | 0.874 | 0.853 | 0.948 | 0.887 | 0.906 | 0.940 | 0.880 | 0.879 |
|
0.997 | 0.998 | 1.000 | 0.998 | 0.998 | 0.997 | 0.997 | 0.998 | 0.998 |
|
|
|
|
|
|
|
|
|
|
: Patient administered a very small quantity of neuromuscular blocking agent (<4 mg) at induction only to facilitate tracheal intubation.
: Maintenance with sevoflurane.
Performance estimated with nonlinear and linear Support Vector Machine (SVMNL and SVML respectively), and Linear Discriminant Analysis (LDA). ‘TOTAL’ indicates the average performance over all patients.
Perf. | LOC | ROC | ||||
SP | SVML | LDA | SVML | LDA | ||
|
F = 0.75, p = 0.39 | F = 1.28, p = 0.26 |
|
F = 0.85, p = 0.36 | F = 0.68, p = 0.41 | |
|
F = 0.11, p = 0.74 |
|
F = 0.01, p = 0.93 |
SE | SVML | LDA | SVML | LDA | ||
|
F = 2.5, p = 0.12 | F = 6.02, p = 0.02* |
|
F = 1.1, p = 0.30 | F = 2.98, p = 0.09 | |
|
F = 0.5, p = 0.48 |
|
F = 0.35, p = 0.56 |
Acc | SVML | LDA | SVML | LDA | ||
|
F = 1.75, p = 0.19 | F = 4.39, p = 0.04* |
|
F = 1.06, p = 0.31 | F = 2.01, p = 0.16 | |
|
F = 0.36, p = 0.55 |
|
F = 0.09, p = 0.76 |
Statistical significance of differences in performance of the different classifiers at loss and recovery of consciousness (LOC and ROC respectively). Classifiers: linear (SVML) and nonlinear (SVMNL) Support Vector Machine, and Linear Discriminant Analysis (LDA). Performance (Perf.) estimated as specificity (SP), sensitivity (SE), and accuracy (Acc). Significance was estimated with one-way ANOVA F-test (α = 0.05; Fcrit(1,41) = 4.079), and significant differences are marked with *.
Perf. | SVML | SVMNL | LDA |
|
F = 2.37, p = 0.13 | F = 2.01, p = 0.16 | F = 1.32, p = 0.26 |
|
F = 4.19, p = 0.05* | F = 9.74, p = 0.003* | F = 5.04, p = 0.03* |
|
F = 3.6, p = 0.06 | F = 5.55, p = 0.02* | F = 3.81, p = 0.06 |
Statistical significance of differences between loss and recovery of consciousness conditions (LOC and ROC respectively). Classifiers: linear (SVML) and nonlinear (SVMNL) Support Vector Machine, and Linear Discriminant Analysis (LDA). Significance was estimated with one-way ANOVA F-test (α = 0.05; Fcrit(1,41) = 4.079), and significant differences are marked with *.
Ref. | Accuracy | Prediction Probability | Features |
|
0.98 | Granger Causality | |
|
0.92 | Narcotrend™ monitor | |
|
0.86 | Recurrence quantification analysis | |
|
0.86 | Approximate Entropy | |
0.86 | Spectral edge frequency | ||
0.78 | Median frequency | ||
0.82 | BIS® monitor | ||
|
0.77 | Approximate Entropy | |
0.87 | Permutation Entropy | ||
0.87 | Order Recurrence Rate | ||
0.87 | Phase coupling of order patterns | ||
|
0.85 | Approximate Entropy | |
|
0.86 | Permutation Entropy | |
0.79 | Approximate Entropy | ||
|
0.84 | Hilbert-Huang state entropy | |
|
0.69 | Time Encoded Signal Processing and Recognition (TESPAR) | |
|
0.87 | BIS® monitor | |
0.89 | Datex-Ohmeda S/5 Monitor (State Entropy) | ||
0.88 | Datex-Ohmeda S/5 Monitor (Response Entropy) | ||
|
0.93 | Complexity based on Lempel-Ziv | |
0.89 | Approximate Entropy | ||
0.76 | Spectral Entropy | ||
0.64 | Median Frequency |
The ability to discriminate between ‘Awake’ and ‘Anesthetized’ state is important for depth of anesthesia monitors. Using GC as features, we were able to obtain high sensitivity, specificity and accuracy. Despite the inter-subject variability in the actual GC values for each subject, the GC patterns displayed the same trend for all subjects. Even though SVM is a powerful classifier suitable for complex high-dimensional problems, it was chosen here specifically for this simpler low-dimensional problem, as it allows us to study both linear and non-linear classification utilizing a single technique. Even though linear classification was outperformed by non-linear classification, the differences in performance are not statistically significant. This implies that the GC features utilized are linearly separable and, from a statistical perspective, it is not necessary to introduce a more complex non-linear classifier with increased computational cost. Therefore, a much simpler linear classifier, such as LDA, or even a technique based on some form of adaptive threshold estimation, could be utilized. The latter could also be more appropriate for real-time applications and remains the subject of future investigations.
Pairwise time-domain GC analysis has received some criticism, mainly regarding the interpretation of the resulting causality relationships. A main limitation is that one cannot distinguish between direct and indirect causal relationships when performing pairwise GC analysis. This is related to the issue of spurious causality that can appear between two processes when both are influenced by external sources that are not taken into account
The difference between classification performance for examples from LOC and ROC indicates that both conditions show some statistically significant variations (
The marker for ROC indicates that the subject has regained consciousness. The administration of anesthetics had been switched off a few minutes prior to this event. Should the estimated GC features have been a reflection of the metabolic decrease of the anesthetic agent, the decrease in the values of GC would be gradual and not sharp as observed. Thus, there would not have been a clear boundary between the GC features for each class, leading to lower classification accuracy. However, the high performance is an indicator that GC features reflect the points at which consciousness is lost and recovered. This provides strong support for the use of such features in a DOA monitor as a change in the patient's state of awareness would be promptly captured. It is also possible that some of the segments used in the analysis may contain data from the start/end of the surgical procedure, and it is known that surgical noxious stimuli, e.g., tends to lighten the level of hypnosis
Bidirectional interaction or strong unidirectional interaction in the presence of a common input as captured by GC are related to mechanisms of information flow in cortical circuits, in terms of the anatomical connectivity principle of reciprocity in the cortex or the collective activation of cortical regions projecting to the measured sites respectively
Regarding the choice of an appropriate order for the AR model utilized, this is non-trivial: if the order is too low the properties of the signals are not captured, however if the order is too high then any measurement noise or inaccuracies are also represented and the resulting model is not a reliable representation of the signal
Another important consideration is the presence of 50-Hz line noise, which can be removed easily using a 50-Hz notch filter. The use of filtering in GC analysis has raised some contradictive opinions (see work by Florin
A: GC between left frontal and right posterior areas for patient S17; B: GC between right frontal and left posterior areas for patient S1. GC is shown for notch filtered EEG (top panels of A and B) and unfiltered EEG (bottom panels of A and B). Vertical lines denote anesthetic induction (dashed line) and recovery of consciousness (dotted line). The effects of diathermy artifacts on the GC estimates can also be identified as sharp outliers.
An ideal DOA monitor should display 100% SE and SP. However, it is very difficult to have an ideal monitor and in the majority of cases a compromise between SE and SP must be made. But what does this compromise translate to in terms of a DOA monitor? Let us first consider what SE and SP imply for a DOA monitor. Ideal SP means that all events of awareness are captured by the monitor. This implies that an alarm is raised and, in such a case, appropriate actions, such as administration of an anesthetic bolus, would have to be taken by the anesthetist to ensure adequate anaesthesia. Ideal SE would imply that when the patient is adequately anesthetized, the DOA monitor reflects this and no further action is needed. Now let us consider the consequences of non-ideal SE and SP. In case of low SE, false alarms would be raised by the monitor, falsely indicating that the patient is awake. If the anesthetist takes action in such a case, the consequences could be disastrous. In case of low SP, the monitor would fail to raise the alarm in some cases of awareness. The anesthetist would take no action and the patient would continue being aware, with possible psychological consequences to the patient. It can be seen that in the case of a DOA monitor, both SE and SP are equally as important and no sacrifice of one should be made for the other. Using GC as a feature, even though SE and SP are not ideal, both are at a similarly high level. Thus, neither is sacrificed for the other.
The feasibility of utilizing Granger Causality, a measure quantifying linear bidirectional signal interactions, as a feature for discriminating between brain activity from awake and anesthetized subjects has been investigated. Our findings support the use of GC estimated in the direction of anterior to posterior brain areas as a feature to discriminate between the EEG of an awake and anesthetized subject. High sensitivity, specificity and average accuracy were obtained for both linear and non-linear classification. The findings suggest that GC-based features are linear, thus the use of a complex non-linear classifier is not necessary. Thus, it may even be possible to employ some form of a more sophisticated and adaptive threshold for classification purposes, which would perhaps be more appropriate for the future development of a DOA monitor. The threshold would need to be adaptive as, despite the same GC patterns observed in all subjects, there is large inter-subject variability in the actual GC values that characterize ‘Awake’ and ‘Anesthetized’ states in each patient. A monitor based on adaptive classification would be advantageous over current DOA monitors, whereby the range discriminating the two states is fixed. Future work will focus on identifying the location of a small number of electrodes that can be utilized successfully in a DOA monitor, instead of utilizing the average activity of all available electrodes.
The authors would like to thank the hospital staff and the anonymous volunteers who participated in this study.