Monitoring the Awake and Anesthetized Unconscious States Using Bispectral Index and Electroencephalographic Connectivity Measures

Abstract

Objective. Our objective was to compare three electroencephalography (EEG)-based methods with anesthesiologist clinical judgment of the awake and anesthetized unconscious states. Methods. EEG recorded from 25 channels and from four channel bilateral Bispectral index (BIS) electrodes were collected from 20 patients undergoing surgery with general anesthesia. To measure connectivity we applied Directed Transfer Function (DTF) in eight channels of the EEG, and extracted data from BIS over the same time segments. Shannon's entropy was applied to assess the complexity of the EEG signal. Discriminant analysis was used to evaluate the data in relation to clinical judgment. Results. Assessing anesthetic state relative clinical judgment, the bilateral BIS gave the highest accuracy (ACC) (95.4%) and lowest false positive discovery rate (FDR) (0.5%) . Equivalent DTF gave 94.5% for ACC and 2.6% for FDR. Combining all methods gave ACC = 94.9% and FDR = 1%. Generally, entropy scored lower on ACC and higher on FDR than the other methods (ACC 90.87% and FDR 4.6%). BIS showed at least a one minute delay in 18 of the 20 patients. Conclusions. Our results show that BIS and DTF both have a high ACC and low FDR. Because of time delays in BIS values, we recommend combining the two methods.

Introduction

Observation of clinical signs has been the traditional way of monitoring depth of anesthesia (DoA) to ensure appropriate dosage of anesthetics and analgesics, prevent awareness during surgery, and ensure comfortable awakening after surgery.^1–4 However, differentiating between the awake and anesthetized states and assessment of DoA based entirely on clinical signs are unreliable,^5,6 and have moderate sensitivity and specificity to identify intraoperative awareness. ⁷ Indeed, hemodynamic variables can be stable even when patients are awake during anesthesia. ¹

Bispectral index (BIS) is an electroencephalogram (EEG)-based technology, composed of a combination of time domain, frequency domain, and a weighted sum of high order EEG spectral parameters, ⁸ that is used during surgery to monitor DoA, thereby aimed at preventing undesired awareness and overdose of anesthetics.^6,9–11 BIS has been crafted and tuned to replicate clinical consciousness assessment.^12,13

BIS displays a single number, an index that can be used as a proxy to arousal level of the patients during anesthesia. The index simplifies interpretation of the EEG signals ¹⁴ ; it ranges from zero (isoelectricity) to 100 (awake), where values below 60 are associated with unconsciousness. ⁷ During surgery values between 40 and 60 are regarded as optimal. ¹⁵

The interpretation of BIS values can be confounded by changes in electromyogram (EMG) activity, ¹⁶ dementia and other neurological disturbances affecting EEG, as well as increased values with the use of nitrous oxide (N₂O) and ketamine. ⁸ As a processed EEG monitor, it is claimed that BIS has limitations of calibration range and interindividual variability in dose-response curves, ¹⁷ as well as differences between the predicted DoA and the actual DoA. ¹⁸

Effective connectivity between EEG electrodes can be used to distinguish between conscious and unconscious states in humans.^19–22 In this study we have used raw EEG measures as the complexity measure ‘Shannon's Entropy’ and the effective connectivity measure ²³ called ‘Directed Transfer Function’ (DTF) (Granger causality group).^{19,24–28,20,29}

DTF determine the directional influence in any given pair of channels in a multivariate data set, estimating the information flow between EEG sources. ³⁰ Opposed to BIS and entropy, DTF is not influenced by volume conducted signals in different electrodes. Hence, eg volume conducted muscle artefacts are not expected to influence the assessment. DTF is measured by means of phase information in the frequency domain, ³¹ as implemented and used here, it is normalized on inflow, not on outflow.

Entropy is a complexity measure also conveying information of local connectivity, ³² and was included to assess yielded information besides what was returned by BIS and DTF. ²⁹ Entropy was chosen due to simple, fast calculations and transperancy in the assessment. Opposed to the other methods it was calculated for each single channel giving a local measure.

The objective of this study was to assess ways to increase yield in arousal level estimations, either by alternative methods to BIS or to find methods to increase the quality of BIS by methods adding information. Consciousness is a loosely defined entity, and we have used the consciousness as assessed by the anesthesiologist as reference in the evaluation of the different methods.

This paper compares BIS, DTF and entropy, and combinations of these methods, to map their applicability in brain monitoring during anesthesia.

State of Field

The set of elements to evaluate consciousness are under exploration and there is still no single method seen as the best to assess consciousness.^13,33 Many methods have been suggested, quite few of them have been tested in patients ³⁴ and shown diverse results, partly due to chosen parameters and to diverse patient groups.

There are available systems in the marked such as BIS, Cerebral State Monitor, SEDLine, AEP-monitor, Neurotrend, Neurosense, SNAPII and Narcotrend.^35,36 BIS is the most commonly used method, both in anesthesia, disorder of consciousness (DOC) and intensive care monitoring. It is easy to handle, and with just a few electrodes it seems to return enough clinical valuable information. However, in diseases in the brain with focal affection not including the prefrontal region, you will not expect credible results from BIS or other methods using just a few electrodes frontally. On the other side, recording with just a couple of channels in the forehead will usually be enough to safely monitor anesthesia depth.

Using EEG for monitoring is difficult; the EEG is very dynamic, volume conductions blurs the information and reduces the yield. Granger causality methods eliminates the volume conduction as it only analyses delays. These methods are implemented both in time domain and frequency domain, but there are no clear best method yet.

Stronger computers, and testing of more complex methods which will eventually be implemented, or the combinations adapted with the help of machine learning, will pave way for better methods.

Methods

We assessed values from BIS, DTF, and entropy in patients undergoing surgery with general anesthesia. The study was approved by the Regional Committee for Research Ethics (case number 2012/2015). After written and oral information about the content and purpose of the study, all patients signed a written consent to participate.

Study Population and Inclusion Criteria

Twenty-five healthy adults hospitalized for anterior cervical decompressive discectomy spine and fusion surgery were included in the study. As the results would be applied to single patients we opted for kohort of 25 patients for significant results.

To be included the patients had to be classified from I-III according to the American Society of Anesthesia (ASA) system (ASA Physical Status Classification System, American Society of Anesthesiologists; https://www.asahq.org/resources/clinical-information/asa-physical-status-classification-system), be between (18 and 75 years old), and otherwise healthy based on a complete health examination. No other distinct exclusion criteria's were applied.

EEG with 25 channels and bilateral BIS recording during surgery were mandatory to be included in the study. Of the 25 patients, two patients were excluded because of technical problems, and three patients were excluded because unilateral BIS electrodes had been used. The mean age of the patients was 52 years (range 33 to 74 years) and there were 15 women and 5 men (Table 1). The enrollment of the participants was carried out from February to June 2019.

Table 1.

Patient Demographics and Characteristics of the Operative Period.

	n = 20
Age (years)	52 ± 9.6 (SD)
Sex (W/M)	15/5
Duration of Anesthesia (min)	144 ± 36 (80-239)
Duration of Surgery (min)	96 ± 60.5 (54-174)

Data are presented as mean ± standard deviation (SD), median (range), or numbers.

Abbreviations: W/M, female/male sex; n, number; min, minutes.

Anesthetic Management

The premedication consisted of oral paracetamol (Paracet ®, Weifa, Oslo, Norway) 1 g and oxycodone sustained release tablet (opioid analgesic; OxyContin®, Dublin, Ireland) 10 mg. General anesthesia was induced and maintained using propofol 10 mg/mL (Propolipid®, Fresenius Kabi, Uppsala, Sweden) and remifentanil 50 µg/mL (Ultiva®, GlaxoSmithKline, Parma, Italy). This was administered using infusion pumps (Braun Perfusor Space®, Melsungen, Germany) programmed to achieve targeted brain concentration of these anesthetic medications to render unconsciousness and state of surgical anesthesia. The target protocol used for achieving unconsciousness by propofol was the Schnider model ³⁷ for effect-site concentrations of 2–6 ng/mL on induction and during surgery. Start of induction and surgery were noted for each patient. For remifentanil the Minto model was chosen, ³⁸ for effect-site concentrations of 3–8 ng/mL. The highest concentration was used during induction of anesthesia, and at start of surgery. The range of anesthetic doses administered during anesthesia was 2.5–7.2 mg/kg/h for propofol and 0.09–0.24 µg/kg/min for remifentanil.

To facilitate intubation, a single dose of cisatracurium 10 mg/mL (Nimbex®, GlaxoSmithKline, Oslo, Norway) at 0.15 mg/kg was administered after the patient was considered unconscious. The patients also received a single dose of 16 mg dexamethasone (Dexavit®, Vital Pharma Nordic, Hellerup, Denmark) in order to prevent postoperative nausea and vomiting. Prior to wake-up, the patients received intravenous fentanyl 50 µg/mL (Fentanyl®, Hameln Pharma, Hameln, Germany) in the range of 0.8–3.8 µg/kg based on the clinical judgment of anesthesiologist.

Clinical Assessment of Conscious State

All patients were monitored according to Norwegian Standard for the Safe Practice of Anesthesia. ³⁹ During induction, the patients were instructed to talk and respond to questions by the anesthesiologist, and were continuously checked for loss of response, until they were unresponsive. We noted loss of verbal contact and loss of verbal response in accordance with the Modified Observer's Assessment of Alertness/Sedation Scale (MOAAS). ⁴⁰ At the end of surgery, the anesthetic infusions were stopped, and the patients were spoken to and tested with a decreasing interval from about once a minute to almost constantly until return of verbal contact and response. We noted stop of anesthetic infusions to extubation, motor response to verbal command (“show me your left index finger”), and verbal response to verbal command.

BIS and EEG

In this study we used a bihemispheric 4-channel BIS™ (Medtronic^®) anesthesia depth monitor. BIS was recorded from electrodes placed frontally over the 10/20 placement of Fp1, Fp2, F7, and F8, and the reference was placed at Fpz. ⁴¹ EEG was recorded with 25 channels with electrodes placed in accordance with the 10/20 system ⁴² with additional low row electrodes (F9, F10, T9, T10, P9, and P10). The reference and ground electrode were placed at Cp3 and Cp4. ⁴¹ EEG was recorded with a Natus C64 EEG amplifier, at 512 samples per second, and filtered between 0.5 and 70 Hz upon acquisition. During every recording the impedance was below 10 kOhm for both EEG and for BIS. Just before start of recording, the two systems were manually synchronized.

EEG and bilateral BIS electrodes were attached to the patients before entering the operation theatre. Before induction of anesthesia, EEG and BIS recording were started simultaneously. The patients were instructed to be awake without further instructions, then keep their eyes closed for two minutes, followed by keeping their eyes open for two minutes, and then be awake without further instruction before induction. EEG and BIS recordings were stopped and electrodes were removed when the anesthesiologist got verbal contact with the patients.

Data Processing

Extracted EEG and logged values for each minute from the BIS monitor were exported to Matlab 2020a, where the averaged bilateral BIS data (left and right) were analyzed together with DTF and entropy. DTF was calculated as implemented in the procedure from eConnectome ⁴³ with model order 5 and 20,^44,45,46,47 in the theta (4 Hz–7 Hz) and alpha (8 Hz–13 Hz) band.^48,49 The maximum value within each frequency band, for each 1 s non-overlapping were analyzed. One second was chosen to assess the dynamics in the EEG signal.

Entropy was calculated as a standard Shannon's Entropy in time domain with 10 bins using an in-house written Matlab script, classified into ten bins according to amplitude. Values were returned for every second and median for each minute was used in the discriminant analyses.

BIS values were normalized from 0–100, whereas neither DTF outflow nor entropy were normalized. ²⁹

The median value for each minute for DTF and entropy was calculated to match the time resolution in BIS, and applied to minimize influence of high amplitude artefacts.

The analyses were performed from eight channels in frontal, central, temporal, and parietal lobe (10/20 system), with good coverage in postcentral areas. The channels were chosen to cover the default mode network and the network in the Dynamic Causal modeling (DCM)-model described by the Friston group to detect changes in consciousness. ²³

Time stamps for each minute of EEG and BIS were synchronized with clinical judgment of the patients awake and anesthetized unconscious state, and events during surgery (awake, awake with eyes closed, awake with eyes open, start of anesthesia, anesthetized unconscious state, start of surgery, end of surgery, end of anesthesia, regained consciousness by giving verbal contact and adequate motoric response to verbal command after anesthesia). Of these states, we coded awake (0), awake with eyes closed (1), awake with eyes open (2) and anesthetized unconscious state (3) for statistical analyses.To mimic clinical settings no filtering or other cleaning procedures were implemented. Besides the use of median, no other preprocessing was made.

EEG data was classified by a fine tree discriminant analysis as implemented in the classification learner in Matlab 2020a, with one to seven predictors (BIS right average, BIS left average, DTF alpha with order 5 and 20, DTF theta with order 5 and 20, and entropy) and the four arousal states. For each patient a receiver operating characteristic (ROC) curve and a confusion matrix were calculated, giving values for area under curve (AUC), accuracy (ACC) and values for false positive discovery rate (FDR)- for unconsciousness (Figure 1 and 2 in the supplement). Data were validated with a five-fold cross-validation, and we opted mainly for reporting ACC in this study.

Results

A total of 3277 minutes of EEG recordings were sampled from 20 patients, and 184 minutes of these were clinically classified by the anesthesiologist as awake (states 0, 1, and 2). During surgery, none of the patients showed signs of being awake, or reported afterwards that they had been awake.

The method's ability to differentiate between awake and anesthetized unconscious state depends on reported interindividual variation and the mean difference between the reported values in the different states. The mean awake BIS value for all patients (state 0, 1, and 2) was 91.5, with a range from 78.8 to 97.5 (20% of the mean value). There are no overlap between reported values in awake and unconscious group using BIS. For DTF, the mean value was 0.090 (range 0.051-0.121 [Span: 0.07 = 77% of the mean]) which gives some overlap, while for entropy the mean value was 0.721 (range 0.038-1.331 [Span: 141 = 179% of the mean]) giving a full overlap between the states.

There were no significant differences between the three awake states in each method, except for DTF order 5 in the alpha band between the awake state and the following awake with eyes closed period (p = 0.0035 uncorrected [Bonferroni: 0.05/7 = 0.007]) (Table 1 in the supplement). Mean duration of anesthesia was 144 minutes (80-239), of which 96 minutes (54-174) were time of surgery (Table 1).

To assess the value of the different parameters for consciousness using BIS, DTF and entropy, discriminant analysis were used to evaluate how well the three methods conveyed the conscious state. Generally, entropy scored lower on ACC and higher on FDR than the other methods (ACC 90.87% and FDR 4.6%). The highest average value for ACC (95.4%) and lowest value for FDR (0.5%) was found for bilateral BIS. Combining all methods gave average ACC (94.9%) and FDR (1%), slightly lower ACC and higher FDR than for BIS alone. For DTF alpha, order 5, ACC was 94.5% and FDR was 2.6%. In Figure 1 a plot of mean values with minimum and maximum value of returned ACC in all patients for each method are presented. The ACC showed variations with all methods.

Figure 1.

Plot of the mean values with minimum and maximum value of returned ACC in all patients for each method. The ACC showed variations with all methods.Entropy scored lower on ACC than the other methods (ACC 90.87%). The highest average value for ACC (95.4%) was found for bilateral BIS. Combining all methods gave average ACC (94.9%), slightly lower ACC than with BIS.

Abbreviations: ACC, Accuracy; BIS, Bispectral index; bilat, bilateral; DTF, Directed Transfer Fuunction; a, alpha; t, theta; 5, order 5; 20, order.

In Figure 2 a plot of BIS, DTF and entropy values from one patient during surgery are presented. No structured consciousness assessment was used. Entropy values are divided by 10 to match in the plot.

Figure 2.

Plot of BIS, DTF and entropy values from one patient during surgery. No structured consciousness assessment was used. Entropy values are divided by 10 to match in the plot.

Abbreviations: BIS, Bispectral index; DTF, Directed Transfer Function.

From start of induction until the patient were classified with an anesthetized unconscious state, there are in 18 cases (90%) seen delayed BIS values with an average of 3.3 minutes, compared with the classification by the anesthesiologist and DTF values (average 1.3 min). As seen in Table 2 in the supplement there is no response from the patient the minute after start of anesthesia, with an immediate change in DTF values whereas BIS numbers still shows 98, which indicate that the patient is awake. This changes to a BIS value at 72 the next minute, and to a BIS value at 48 the next minute (which indicate that the patient is not responsive).

An immediate change in DTF values related to changes in blood pressure, fading of muscle relaxation or artefacts due to use of electrical diathermia or repositioning of the head were observed, whereas changes in BIS values were not apparent. Entropy showed wide distribution of calculated values from both awake and anesthetized unconscious state, and hence, infrequently classified the states correctly.

Discussion

To our knowledge this study is original with its objective to compare different EEG-based methods in clinical settings with the main goal to combine methods for a better consciousness measure. The study is of importance in understanding the classification of the state of consciousness during surgery, and, hence, the findings from this study will likely help in reducing the rate of underestimation of depth of anesthesia in patients.

Our results showed that BIS and DTF had approximately the same clinical applicability during anesthesia, as reflected in approximate equivalent values of ACC and FDR. With mean FDR of 0.5% for BIS, the FDR are probably due to the delays in BIS values and not to misclassifications in the model as such. DTF and entropy who do not feature this delay, the reported classification errors are expected to be due to misclassifications.

The entropy as implemented here was clearly inferior to the two other methods. There was apparently no information conveyed by entropy that was not already reflected in BIS and DTF. BIS was normalized and even with interindividual variations, there was no need for calibration of the monitor between patients. The two other methods were not normalized. There are methods to normalize entropy, but these were not implemented in this paper as we found the method not transparent in our hands, and we opted for raw data in this explorative round. Because of interindividual differences a preanesthesia recording period is mandatory, which opens for DTF as a more sensitive method than BIS, as interindividual differences are corrected for by the preanesthesia recording period. However, this effect is not obvius in the study.

DTF was implemented in a straightforward way as the segment length, and assessed frequencies were the same as in previous studies. ¹⁹ The segment length of one second was chosen to be long enough for calculations in the delta band, and short enough to keep much of the non-linear dynamics of the signal. Theta activity is known to be cross frequency coupled to gamma activity, which has been postulated to be central in local brain processing. Hence, we found it natural to include theta. The alpha activity will change with eye open and closure in postcentral leads in EEG. However, there are alpha activity in all brain regions influenced by brain activity. To constrain the study we did not include delta as we assumed little added usable information in this context. Gamma was excluded due to the dispersion effect, we did not expect to see connected sources more than a few centimetres apart except when conveyed by the theta.^48,49 Analyzes of beta and gamma band could have conveyed additional important information, however it was outside of the scope of this paper to further tune DTF for better classification. DTF was calculated both with order 5 and 20. Only order 5 is reported as there were no differences between order 5 and 20.

The major difference between open and closed eyes in clinical EEG is the alpha blocking in postcentral channels, ⁴² which is not easy to detect in frontally placed electrodes. ⁵⁰ The clinical EEG readings showed clear differences in values between open and closed eyes. These differences were not detected by the BIS monitor, but by DTF within the alpha range. Theoretically, DTF in the theta range should be a better method than DTF in the alpha range due to the volatile alpha activity. In this study, the opposite seems to be the case, because of the selection of eyes closed as entity of its own. In usual clinical situation, we expect DTF theta to be more accuarate.

As highlighted in another study, ¹⁸ delays in BIS values were observed. The algorithm used is not available to be scrutinized, so the reason for this delay is only a matter of speculation. It may be due to temporal averaging or a moving average beyond the one minute update. This is supported by the surprisingly smooth BIS values despite muscle and other artefacts in the EEG. The small difference between each minute supports such a notion. Also, long periods with diathermia did not give changes to BIS-values and, hence, seemed to be processed away. This gives some uncertainty if BIS does not flag important brief changes in the brain activity due to important events during anesthesia.The delays observed with BIS are disturbing, and if a situation occur, it may take 2 minutes until BIS warns the anesthesiologists. BIS updates every minute, but there seems to be averaging over a longer period. This delay is not in DTF, but as implemented in this study, there is a one minute average increasing temporal resolution.

In this research setting, the delay has lowered the classification results. Typically, in the initial part where the patients become unconscious, BIS still reported the awake state due to the delay. Because of the delay in BIS values, brief transients may not be detected, in contrast to when DTF is used.

Conclusion

This study has shown that both BIS and DTF have a high ACC and low FDR classifying state of consciousness. Combining the two methods did not increase the ACC, but due to delay in BIS values we acknowledge EEG-derived DTF as an important adjunct to BIS. Despite the observed delays in BIS, and consequent lack in detecting changes as quick as EEG, BIS has an edge over DTF because DTF needs a preanesthesia calibration period.

This study was made in a clinical setting constraining our ability to optimize the setting, all statistics should be read within the context of this sample distribution.