Can Bispectral Index Monitoring Predict Recovery of Consciousness in Patients with Severe Brain Injury?

The study was prospectively and consecutively performed in 25 critically ill brain-injured adult patients who did not regain consciousness after sedation withdrawal. These patients were treated in an 8-bed surgical ICU in an 800-bed university hospital. In this ICU, approximately 90 patients with brain injury are admitted per year. All patients included in the study were unconscious and unable to respond to verbal commands, most of them were under pressure support ventilation (mechanically assisted spontaneous ventilation), and all had their trachea intubated. They had not received intravenous sedation for at least 24 h before the study day. The study started in September 1999 and finished in June 2001. All procedures were reviewed and approved by the Institutional Review Board of the Hospital Clinic of Barcelona (Barcelona, Spain).

On admission to the ICU, a number of data were recorded: sex, age, associated diseases, computed tomography scan, type of brain lesion, APACHE II severity score,13and therapeutic intervention scoring system.14Other data collected during the stay of the patients in the ICU were type of surgery, neurologic invasive and noninvasive monitoring, medical complications, presence of tracheostomy, and cause of death. Days from the injury to the day of study, days under sedation, sedative drugs used, and days without receiving sedation before BIS recording were also registered.

Bispectral Index recording was performed by means of an A-2000® monitor (Aspect Medical Systems, Inc., Newton, MA) using BIS algorithm version 3.4. A three-electrode sensor (BIS® sensor; Aspect Medical Systems, Inc.) was located on the forehead of the healthiest brain hemisphere, which was identified by computed tomography scan, with unilateral bipolar modified frontomastoid montage (Fpz-A¹or Fpz-A², International 10-20 System of electrode placement).15Electrode impedance was maintained below 5,000 Ω to ensure adequate signal quality. BIS, 95% spectral edge frequency, burst suppression ratio (BSR), spontaneous electromyographic activity in the frontal area, and signal quality index (SQI) were recorded directly from the output of the A-2000® monitor for 20 min. Although electromyographic activity is defined as the total power (in decibels) of the electroencephalographic signal within 70 to 110 Hz frequency range of the spectrum, it was quantified as a unitless value ranging from 0 to 10.16From the BIS values, for each individual recording, estimates of maximal (BIS^max), minimal (BIS^min), average (BIS^mean), and range of BIS (BIS^range) values during the 20-min recording were calculated. Maximal electromyographic value was chosen for calculations when a range was given.

Patients’ neurologic status at the moment of study, the day of BIS measurement, was assessed by the GCS (GCS^BIS).2Patients were followed up for 6 months after the injury or until they recovered consciousness or died. Consciousness recovery was evaluated by measuring the ability of the patient to respond to verbal commands, independently of the degree of disability of each particular patient. Glasgow Outcome Score (GOS)17was used to define neurologic patient status on discharge from the ICU, on discharge from the hospital, and at the end of the follow-up period.

Patients were divided into two groups according to their final outcome: those who recovered consciousness (good recovery [GOS = 1], moderate disability [GOS = 2], severe disability [GOS = 3]) and those who had a poor neurologic outcome (persistent vegetative state [GOS = 4] or death [GOS = 5]).

The values of BIS^max, BIS^min, BIS^mean, BIS^range, 95% spectral edge frequency, BSR, electromyographic activity, and SQI, as well as the clinical indicators measured on admission (GCS, APACHE II, therapeutic intervention scoring system), and on measurement day (GCS^BIS) in the group of patients who recovered consciousness (GOS = 1–3), were compared to the same variables in the group of subjects who did not recover consciousness (GOS = 4 or 5), by means of the Mann–Whitney U test. Statistical significance was considered when P  was less than 0.05.

The P^ktest was used to evaluate the predictive ability of each of the parameters that significantly differed between both groups. The goal was to assess the ability of an indicator to predict an observed response—in our case, the recovery of consciousness. The statistic P^kis a type of nonparametric correlation known as a measure of association that quantifies how well the indicator can predict the observed response. The closer P^kis to 1, the better prediction of the indicator is, whereas a P^kvalue of 0.5 demonstrates that the indicator is not better than random to predict the observed response. A detailed explanation of the statistic properties of P^kand several applications have been published elsewhere.18–20 

In an attempt to model the relation between the variables that scored the higher values of P^k, meaning that their predictive capacity was maximal, and the outcome observed (recovery of consciousness), univariate and multivariate logistic regression, using a stepwise method, was used. The objective was to establish a relation between the variables and the probability of recovery of consciousness.

The graphical representation of the relation between BIS^maxand the probability of recovery of consciousness was plotted, and estimates of BIS^50%(the value of BIS^maxassociated with a probability of 0.5 of recovery of consciousness) and γ (the parameter defining the slope of the sigmoid curve) were obtained.

To assess the ability of the model in predicting recovery of consciousness and in the absence of a prospective sample of individuals, the method of cross-validation using the “leave-one-out” approach was used.21,22For that purpose, the parameters defining 25 submodels of 24 individuals each were estimated using a reparametrization of the logistic regression model as follows:

where Pⁱis the probability of recovering consciousness in the i  th patient, BIS^maxis the value of BIS^maxin the i  th patient, BIS^50%iis the estimate of the value of BIS^maxassociated with a probability equal to 0.5 of recovering consciousness, and γⁱis the estimate of the slope of the sigmoid relation, being BIS⁵⁰ⁱand γⁱthe parameters to be estimated for each submodel. Using this approach, 25 submodels were fitted, according to the maximum likelihood criteria; hence, 25 values of BIS^50%and 25 values of γ were estimated, by fitting each submodel to the data of 24 patients each time, using a pooled approach.

The predictions of each submodel were compared to the observed data of the remaining individual. Cross-validation is a convenient method to evaluate the predictive ability of a proposed model when sample size is small and a prospective evaluation unfeasible, and as such, it has been used by different authors.21,23This approach also allows us to identify any individual that could have been strongly influential in the model.

Evaluation of the performance was done by comparing the observed probability of recovering consciousness and the predicted probability estimated by each submodel. The observed probability was calculated for each patient who had been left out of the submodel building for validation purposes, by dividing the number of patients who recovered consciousness by the total number of patients in each of three BIS intervals arbitrarily defined: 0–40, 41–70, and 71–100. The predicted probability was estimated by the respective submodel according to the BIS^maxvalue of the patient. Performance error was calculated as follows:

where P^obsis the value of the observed probability for each patient, and P^CDis the value estimated by each one of the 25 “n−1” submodels. From the 25 estimations of performance error, the median value, median cross-validation performance error (MDCV), was calculated. Also, the absolute values of performance error were computed and the median, defined as the median absolute cross-validation performance error (MDACV), was calculated.21,24 

The MDCV and MDACV can be considered as measures of bias and performance, respectively, of the 25 submodels with respect to the observed probabilities. Confidence intervals were calculated for MDCV and MDACV as described in Campbell and Gardner.25The distribution of the 25 estimates of BIS^50%i, as well as the 25 estimates for γⁱ, was evaluated graphically.

Seventeen men and 8 women were consecutively included in the study. Table 1summarizes the main demographic characteristics of each patient, including diagnosis, age, GCS at hospital admission, and scores of disease severity, such as APACHE II and therapeutic intervention scoring system. Six patients were chronically hypertensive, 1 had chronic bronchopneumopathy, 1 had a previous liver transplantation, and another was diabetic. Brain injury was due to isolated brain trauma in 8 patients, multiple trauma in 7 patients, subarachnoidal hemorrhage in 8 patients, and stroke in 2 patients.

Five patients had diffuse axonal injury on their computed tomography scan. Eleven patients underwent surgery for evacuating a cerebral hematoma, 2 patients were underwent surgery to clip on an aneurysm, and another 2 patients had an embolization of their intracerebral aneurysms. In 7 of these patients, an external ventricular drainage was placed. All but 3 patients had intracranial pressure monitoring on their intensive care unit admission, and 17 had a continuous monitoring of jugular venous blood saturation.

The most frequent medical complications during the stay in the ICU were infections: 10 patients acquired a purulent tracheobronchitis, nosocomial pneumonia developed in 7 patients, and 4 patients had an episode of bacteriemia. Acute respiratory distress syndrome developed in 5 patients, and 16 patients required vasoactive drugs at some point during their stay. Only four patients received phenytoin treatment during intensive care unit admission. All patients except 4 needed a tracheostomy to accelerate their ventilatory weaning.

Auditory evoked potentials were recorded in 7 patients, and somatosensory evoked potentials (median nerve) were recorded in 8 patients. An electroencephalogram was performed in 8 patients within 1 week of BIS measurements. Results are summarized for each patient in table 1.

Midazolam was used as a primary sedative in 23 patients, and 22 of them also received morphine chloride. Fifteen patients received a continuous infusion of clonidine. Table 2shows the days of sedation, days without sedation before measurement, and days from admission to BIS recording.

The values of the maximal, minimal, mean, and range of BIS, BSR, and 95% spectral edge frequency, as well as SQI and electromyographic activity, are shown in table 3. GCS and the neurologic examination at the time of evaluation are also shown in the same table.

Sixty days (range, 1–180 days) was the median duration of the follow-up period (table 1). The degree of neurologic impairment of patients can be observed in table 2, where neurologic status, according to GOS, at discharge from the ICU, at discharge from the hospital, and at the end of follow-up is described. The severity of the neurologic situation in this study population is supported by the fact that only 2 patients were discharged from hospital, 17 patients were transferred to secondary hospitals to continue their rehabilitation procedure, 5 patients died in the ICU, another died in the hospital general ward, and the last patient died in the referred hospital. The cause of death was brain death or directly attributable to severe brain damage in all patients but one, who experienced acute respiratory distress syndrome after regaining consciousness (therefore, the BIS recorded values in this patient were included in the group of patients who regained consciousness).

No differences were found between the group of patients who recovered consciousness and the group of those patients who evolved to a vegetative state or died, with respect to demographic characteristics, severity scores, duration of stay in the ICU, days of sedation, and number of days without sedation when recording the BIS (table 4).

As can be seen in table 4, there were statistically significant differences between the group of patients who recovered consciousness and those who did not, with respect to BIS^max, BIS^min, BIS^mean, and BIS^range, as well as electromyographic and SQI values. There were also significant differences between both groups with respect to the values of GCS^BIS. BIS^rangevalues, expressing the degree of variability of BIS during the 20-min period of measurement, showed a wide range in both groups, which was significantly larger in patients who recovered consciousness (P < 0.05). SQI was significantly higher in the poor-outcome group (P < 0.05), probably related to lower forehead muscle activity in these patients. BSR was also compared between both groups. Even though there was one individual who became brain dead with a BSR of 98%, overall no significant differences were detected between the groups.

The values of P^k(SE) were 0.96 (0.05) for BIS^max, 0.92 (0.05) for BIS^mean, 0.92 (0.06) for BIS^range, and 0.82 (0.09) for GCS^BIS. The highest predictive value, based on P^kestimation, was 0.99 (0.01) for electromyography.

Univariate logistic regression identified a relation between BIS^max, BIS^mean, BIS^range, GCS^BIS, and electromyographic activity with the probability of recovering consciousness. Univariate analysis of spontaneous electromyographic activity showed a strong relation with outcome (P < 0.0001). All patients with an electromyographic score greater than 1.5 regained consciousness. All patients who did not recover had electromyographic values of 1.5 or less. Perhaps a higher number of individuals with higher spread of the data would give more information about the relation between electromyographic activity and recovery of consciousness. However, based on our data, it can only be said that electromyographic activity is a good classifier of patients according to their final outcome.

Based on the results of the univariate logistic regression, multivariate logistic regression was applied. When using multivariate logistic regression, only BIS^maxwas significantly incorporated into the model. No other variable, either electrophysiologic or clinical, increased the ability of the model to predict the probability of recovery of consciousness. The odds ratio for BIS^maxwas 1.17 (95% confidence interval, 1.01–1.35). A graphical representation of the model for all the possible values of BIS^maxis shown in figure 1. As can be seen, the value of BIS^maxassociated with a probability of 0.5 of recovering consciousness was 52.3, and the value of γ was 8.04. A BIS^maxvalue of 69 or greater has a probability of at least 0.9 of recovering consciousness.

To validate the model relating BIS^maxand probability of recovering consciousness, cross-validation was used. Figure 1shows the graphical structure of the 25 submodels estimated from each subgroup of 24 individuals each. All submodels are close to the graphical representation of the full model.

The performance was evaluated by means of the MDACV for accuracy and the MDCV for bias. As can be seen in figure 2, the results show a high accuracy as reflected in a value of MDACV of 3.06% (95% confidence interval, 1–22.15%) and a low bias with an MDCV of 0.84% (0.56–2.12%). Three individuals in the low-BIS range had observed probabilities of recovering consciousness of 0.33, and in all of them, the respective submodels predicted very low probabilities of recovering consciousness, 0.002 for a BIS^maxof 24, 0.01 for a BIS^maxof 30, and 0.0007 for a BIS^maxof 40.

Another way of comparing submodels to the full model relating BIS^maxto probability of recovering consciousness is to graphically examine the parameters defining the model (BIS^50%and γ) versus  the values of BIS^50%and γ estimated in the full model. As can be seen in figure 3, the 25 BIS^50%values are close to the full model value of 52.25, whereas the value of γ for the submodel excluding individual 25 is out of range (23 vs.  8), reflecting a more abrupt change in the relation between BIS^maxand probability of recovery of consciousness when this patient is not included in the full model.

It can be said that, in terms of accuracy and bias, all the submodels are consistent with the findings of the full model. They have been prospectively validated, each one in a different patient. Although individual 25 seems to be influential in the model, the full model retains its ability to describe the data where it has been built, and it is also good for predicting the probability of recovery of consciousness given a certain value of BIS^max.

The main result of our study is that BIS values were significantly different between patients who recovered consciousness and those with poor neurologic outcome (persistent vegetative state or death) and that a model relating BIS^maxto probability of recovery of consciousness could be built. BIS^maxand BIS^meanvalues had a high prediction probability, better than traditional clinical measures usually employed in this population, such as the GCS on admission. However, GCS^BISrecording was the third variable related to outcome, indicating that this clinical score, measured when the patient is without sedation, can also be useful in predicting neurologic outcome.

Since its introduction in clinical practice, the BIS has been widely used in the operating room as a measure of hypnotic drug effect.26,27The development of BIS was entirely focused on obtaining a measure of the hypnotic effects of anesthetic drugs (primarily volatile gases and propofol). As such, the clinical correlation of BIS values in the pharmacologically sedated patient vary with the drug used to provide sedation. Avoiding excessive sedation in critically ill patients has been of increasing relevance in the ICU setting,28,29and several authors have proposed the use of the BIS as a tool to control sedation depth.10,11,30,31 

The ability of the BIS to provide clinical insight in the absence of hypnotic drugs has not been extensively studied. BIS has been shown to correlate closely with reductions in global cerebral metabolic rate produced by anesthetics.32Several clinical reports have described how BIS monitoring was used to provide confirmation of brain activity regained during critical intraoperative events including cardiac arrest and resuscitation.33–35On the other hand, BIS was the earliest indicator of acute perioperative stroke in a recently published clinical report.36 

The BIS was not designed to be applied in neurologic patients; nevertheless, its availability in the ICU has led to a number of studies focusing on these patients. Interpretation of BIS in this situation may be complicated if the electroencephalogram also reflects changes in neurologic status caused by critical illness itself. Another limitation of the BIS in this subpopulation relates to its derivation from unilateral, frontal lobe electroencephalographic signals. Gilbert et al.  12were the first to publish a complete article testing the utility of the electroencephalogram as an index of neurologic dysfunction resulting from critical illness. BIS and other electroencephalographic spectral parameters were measured in a population of unsedated ICU patients, and the investigators concluded that better neurologic function was associated with higher values of BIS.

We are aware that a limitation of our study is the fact that we did not have the plasmatic concentrations of sedative drugs during the study day. All sedative drugs were withdrawn at least 24 h, and no patient was diagnosed of acute renal or hepatic failure when BIS values were recorded before collecting the electrophysiologic data. Although it can be argued that in ICU patients, 24 h is insufficient to allow for washout of the sedative drugs, in our case, those patients who could have been at risk of having significant concentrations of sedatives, as patients 4, 7, 19, and 21, all had increased BIS^maxscores and eventually recovered consciousness.

We performed the exploration during the second week of admission in the majority of patients; none of our patients received sedation from at least 24 h before the BIS recording, and we exclusively included patients who had a delay in awakening. The progress of our patients was followed up until they recovered consciousness or at least for 6 months. In our study, we included different etiologies of the severe brain injury (brain trauma, stroke, and subarachnoidal hemorrhage), although the factors that are known to predict outcome after traumatic brain injury are somewhat different than those that predict outcome after subarachnoidal hemorrhage. We are aware that this could be a limitation of this study, especially with such a small number of patients. In our study, BIS levels were different when patients who recovered consciousness were compared with patients who did not. There was a statistically significant difference (P < 0.05) between groups in BIS^maxand BIS^meanvalues reached. Based on our results, it can be said that those patients whose BIS^maxis higher than 52.25 have a probability of recovering consciousness higher than 0.5, and those with BIS^maxhigher than 69 have a probability higher than 0.9.

Cross-validation enabled an initial validation of the model. Based on the performance of the submodels prospectively validated, it can be said that both accuracy and bias are highly satisfactory. There were three cases of poor prediction in the lower BIS value segment, probably because of the small number of individuals in that BIS segment (between 20 and 40). In fact, the model predicted a very low probability of recovering consciousness (approximately 0.1 for BIS < 40), but there was a much higher measured probability (one of three patients with BIS < 40 recovered consciousness). If there had been a larger number of subjects in the low BIS segment and, as expected based in our study sample, a higher number of no recovery, this would have decreased the measured probability of recovery of consciousness, and the prediction of the model would have been more accurate.

One patient deserves special attention because she recovered consciousness even though her maximal BIS values were low. She was admitted in the ICU with a diagnosis of subarachnoid hemorrhage, recovered consciousness, and was able to respond to verbal commands and eat without aid, although her BIS^maxvalue during recording was 40. One possibility could be that the patient already had a low-voltage electroencephalogram on baseline.37Unfortunately, we have no information about previous electroencephalographic recording or how her electroencephalogram was after she was definitively discharged.

The raw electroencephalogram is not currently used as an outcome predictor in brain-injured patients, other than those with anoxic coma. Diffuse encephalopathy is the most common report in this population. Chen et al.  38found that certain patterns of the raw electroencephalographic signal helped to project 3-month recovery in patients with anoxic coma. All our patients were in “deep coma” when included in the study. Some of them had evoked potential or conventional electroencephalographic studies performed with inconclusive results.

Despite that Theilen et al.  39observed that the fraction of time with a very low-voltage electroencephalographic signal (BSR) soon after severe traumatic brain injury correlated with 6-month outcomes, in our experience, BSR could not help to discriminate which patients would recover consciousness. In fact, we found a high BSR value in only one patient who rapidly became brain dead. Vivien et al.  40have evaluated the accuracy of BIS monitoring for the diagnosis of brain death in severely comatose patients. In their study, the BIS value enabled appropriate scheduling of either electroencephalography or cerebral angiography to confirm brain death.

Vespa et al.  41studied the incidence of seizures after traumatic brain injury. Convulsive and nonconvulsive seizures occurred in 22% of their series. In our study, 4 patients (16%) required treatment with phenytoin for convulsions, but electroencephalography was performed in only 1 patient of this subgroup. In the study of Vespa et al. , 16 of 89 patients had an electroencephalographic diagnosis of seizures, these seizures being nonconvulsive in 12 of 16 patients. Nevertheless, in this series, outcome and discharge were not different between the seizure and nonseizure groups. In our study, no seizure activity was detected by electroencephalography in the 8 patients with electroencephalographically documentation.

Another interesting finding of our study is that spontaneous electromyographic activity had a strong prediction probability (P^k= 0.99) of consciousness recovering. We do not know the real importance of electromyographic activity in the BIS algorithm, but changes in spontaneous facial muscle activity and its repercussion on BIS values is an issue of increasing concern. Our patients were not under forced–warm air blanket therapy or any other special electrical device known to potentially increase BIS readings.42No neuromuscular blocking drugs were used in our patients. Vivien et al.  43showed that the BIS in these patients may be lower with paralysis for an equivalent degree of sedation because of high muscular activity.

Harmel et al.  44suggested that facial electromyographic activity could be used as an objective index of anesthetic effect, muscle relaxation, and level of vigilance. The exclusive motor innervations by the VIIth cranial nerve of facial mimic muscles and, in particular, the multiple connections of the voluntary, involuntary, and autonomic nervous system form the basis of this interesting finding, which has already been the subject of research.44,45Surprisingly, in a short communication, Paloheimo46observed that the return of facial muscle activity after a hypoxic brain injury was always associated with long-term survival in experimental animals. The author concluded that “the mimic facial muscles express the brain’s internal state to the outside world.” We cannot rule out the effect of electromyographic activity on BIS values because high-frequency signals, which can come out from frontal or ocular muscle activity, also contribute to the computation of BIS. On the other hand, it is logical to consider that the patients likely to recover consciousness have more facial muscular activity than those who are not likely to recover. We used BIS algorithm version 3.4 for this study, but a newer version of the BIS®, BIS XP®, is specially designed to discriminate and reject artifacts such as patient movement. In the study of Vivien et al. ,4316 patients were investigated simultaneously with the two monitors (Aspect A-2000® version 2.10 and BIS XP® monitor), and neuromuscular blockade still induced a significant decrease in BIS values with the new monitor.

To summarize, the study of BIS and other electrophysiologic and clinical variables has enabled us to construct and cross-validate a model relating BIS^maxto the probability of recovery of consciousness in patients in a coma state due to a severe brain injury, after sedation has been withdrawn. These results would encourage conducting clinical trials in greater populations to validate them.

The authors thank Steven L. Shafer, Professor of Anesthesia, Anesthesia Department, Stanford University, Stanford, California, for critically reviewing the manuscript.