Predictors of Low Risk for Delirium during Anesthesia Emergence

We derived our results from retrospective post hoc analyses of a previously published data set from our patients with the goal of identifying EEG signatures that correlate with delirium, which was conducted between April 2018 and November 2019.⁸  The study was approved by the ethics committee of Klinikum rechts der Isar, Technical University of Munich (Munich, Germany; Clinical Trial NCT03287401, approved on May 24, 2017). We conducted a retrospective analysis and included all patients from the original study (n = 192). In our reporting, we tried to follow the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) and Standards for Reporting of Diagnostic Accuracy Studies (STARD) guidelines.

Included patients who underwent elective surgery in general anesthesia were older than 18 yr of age and gave written and informed consent to participate in the study. Patients who underwent surgery in the 30 days prior, who had emergency interventions, or who suffered from psychiatric disorders or substance abuse were excluded from the study. All patients were preoperatively screened for delirium via Confusion Assessment Method for Intensive Care Units (ICU), a verified multistep procedure to identify patients with delirium.⁹  It is easily and quickly applied, has high inter-rater reliability, and shows high sensitivity and specificity.¹⁰  In total, 201 patients were included in the study. Anesthesia was maintained either with inhalational sevoflurane or desflurane or with intravenous propofol via syringe pump according to clinical standards. Paralysis, if required, was achieved either with rocuronium or (in one case) with mivacurium. Sufentanil or remifentanil were used for intraoperative pain management. Dosage was chosen by the anesthesiologist in accordance with clinical standards. Patient monitoring was conducted according to the guidelines of the German Society of Anesthesiology (Heidelberg, Germany). The figure in Supplemental Digital Content 1 (https://links.lww.com/ALN/D288) shows the resulting protocol of patient inclusion.

Trained personnel set up the 10-channel EEG recording before anesthesia induction, using noninvasive EEG electrodes applied according to the 10/20 system. The exact electrode layout is shown in the supplemental figure in Supplemental Digital Content 2 (https://links.lww.com/ALN/D289). The reference electrode was Cz. We checked and confirmed the correct positioning of electrodes each time the patient was moved, after induction, and before emergence. Signal quality was monitored throughout the whole intervention. The EEG was recorded with the NIM-Eclipse intraoperative neuromonitoring system (Medtronic, Ireland) with a 250-Hz sample rate and a 1-Hz hardware high-pass filter. The data were stored in the native.eeg format from Medtronic.

After termination of anesthetic delivery, the patients were verbally addressed at regular intervals until they responded purposefully. Addressing the patients began either when the end-tidal alveolar gas concentration reached the minimum alveolar concentration awake (0.35% for sevoflurane, 0.55% for desflurane) or 5 min after terminating propofol delivery. The patients were then addressed at 1-min intervals until they reached a score greater than 2 on the Observer’s Assessment of Alertness/Sedation (OAA/S) scale. The OAA/S is a scale used to measure the level of alertness in sedated patients and consists of the following four categories: responsiveness, speech, facial expression, and eye contact. The scale ranges from 1 (deep sleep) to 5 (alert), where a score of 3 represents a response with eyes opening only after name is called loudly.¹¹  This was defined as the end of emergence. Then, 15 and 60 min later, the patients were assessed in the recovery room to check for delirium using the Confusion Assessment Method for ICU. The patients were defined as positive for delirium in the PACU if they scored positive on the Confusion Assessment Method for ICU at either (or both) 15 or 60 min. There was no follow-up after the stay in the PACU.

We processed the EEG data with MATLAB 2020a (MathWorks Inc., USA) and the MATLAB toolbox eeglab.¹²  Data import into MATLAB was done using custom routines using information provided by Medtronic.

A single frontal channel (Fp2-Fz) was chosen for this analysis because this reflects the current layout of most commercial EEG-based monitoring devices. First, a low-pass filter was applied at 47 Hz using the eeglab function eegfilt. This was done for two reasons: first, to eliminate the 50-Hz line noise, and second, to remove high frequent signal distortions like the electromyography (EMG) that becomes dominant in higher frequencies but overlaps with the EEG spectrum.^13,14  The EEG was cleaned from artifacts in two steps. In the first step, we used an automated artifact subspace reconstruction with clean_rawdata and set the artifact subspace reconstruction parameter to 25 standard deviations, as suggested for automated protocols.¹⁵  The other options of the function were turned off. For the emergence period, density spectral arrays were created using the pwelch function with number of fast Fourier transform points equaling 512 over 10 s EEG segments with a 1 s shift. The density spectral array was calculated for the emergence phase, with t₀ starting at 90 s before the start of emergence representing the maintenance phase, and t₁ representing end of emergence (OAA/S greater than 2). An example density spectral array derived from raw EEG data is shown in figure 1A.

After visually inspecting the 193 density spectral array after artifact subspace reconstruction, 83 data sets had to undergo a second step of artifact rejection because of clearly identifiable artifacts. Visual inspection was focused on excessive blue or red coloring in the density spectral array plots, with red indicating a very high power caused by artifacts, and blue representing very low power caused by zero lines. z scores were calculated for total power of every column of the density spectral array, and those with a score greater than 3 (greater than 2 in few cases) were excluded from the array. The extra step of lowering the z score to greater than 2 was done if the higher boundary did not exclude all remaining artifacts as they were still visible in the resulting density array. This resulted in a data exclusion not exceeding 5%. An exemplary cleaned density spectral array is shown in figure 1B. The remaining 110 data sets did not undergo the second stage of artifact rejection. The cleaned density spectral arrays were then again visually inspected and compared to the original density spectral array for errors in the routine and remaining artifacts. Of the 83 data sets, a total of 24 data sets were excluded as they either were too contaminated by remaining artifacts or more than 10% of data was missing in the investigated interval because of technical issues. The exclusion was done by two different investigators who were blinded to the delirium scores. The final data set consisted of 169 patients eligible for further analysis.

From the cleaned density spectral arrays, we calculated the EEG band power for the delta band (1 to 4 Hz), the theta band (4 to 8 Hz), the alpha band (8 to 15 Hz), and the beta band (15 to 47 Hz). The respective band power was calculated by numerical integration using the chained trapezoidal rule (trapz function). We defined the total EEG power as the cumulative sum of the power in all frequency bands. The course of band power is exemplarily shown in figure 1C.

The linear regression of the change in total power and absolute band power with time during emergence was calculated using the fitlm function, using the total power and band power values for each second. The possible results fell into three categories: either a significantly rising or falling power, represented by an either positive or negative slope and a P value < 0.05; or no significant change in power and a P value > 0.05 (fig. 1D). An example for a patient with delirium is shown in Supplemental Digital Content 3 (https://links.lww.com/ALN/D290). Tests for autocorrelation were evaluated with the Durbin–Watson test using the dwtest function. Quality of fit was evaluated by reporting the R² values. For the exemplary test for no delirium, the results from the linear regression were discretized, and the patients were then classified according to the sign of the slope in total power and the different bands as positive (+), negative (−) or not significant (n.sig) for the corresponding bands. Therefore, absolute values were not further considered as only the signs of the value were assessed.

We conducted a complete case analysis. Corrupted EEG data sets were excluded from the study. Severe artifacts during emergence were either cleaned to a satisfactory level or excluded by two independent and blinded investigators. The excluded data sets with artifacts were included in a sensitivity analysis but were excluded for further multiband analysis. There was no missing epidemiologic data.

The patients were classified according to the combinations of signs of the slope in the different bands, specifically alpha and beta. Three options for alpha and beta each: positive, negative, and not significant. This offered nine possible classification options; e.g., A−/B− representing a negative slope in both the alpha and the beta band, and A+/B+ representing a positive slope in both the alpha and beta band.

All (statistical) analyses were conducted with MATLAB 2020a. Group comparisons were done using the nonparametric two-sided Wilcoxon rank-sum test, as normal distribution could not always be assumed. When comparing contingency data, Fisher’s exact test was used. Fisher’s test was chosen as certain groups were small. The results are either given as the median and the first and third quartile or as the number and percentage of patients with or without a positive Confusion Assessment Method for ICU test. In addition, the calculated P values and the reference group for the Fisher’s test are noted. We support the P values by effect sizes; i.e., the risk ratios with 95% CI or the area under the receiver operating curve (AUC). AUC calculation with 10-k bootstrapped 95% CI was conducted using the MATLAB-based measures of effect size toolbox.¹⁶  The statistical measures and predictive values for the preliminary test for no delirium was also 10-k bootstrapped as means for internal validation. Sensitivity analysis was done by including the data from all patients before exclusion by the two different independent investigators and comparing it to the data set after exclusion.

Of 201 patients included in the original study, 169 were eligible for this analysis and were included. Of the 201 original patients, 7 were corrupted and could not be imported, 1 did not receive general anesthesia, and 24 still contained artifacts in the relevant interval after the artifact removal process and were excluded from further analysis.

Table 1 shows the results of the univariable analysis for patients who developed delirium (n = 32 [19%]) and those who did not (n = 137 [81%]). Patients with delirium were significantly older (P = 0.013). This is consistent with previously published results.¹⁷  Furthermore, patients with delirium had a significantly higher body mass index (P = 0.021). Our analysis showed a significant positive correlation between anesthesia time and delirium, which is in line with previously published studies.⁶  American Society of Anesthesiologists Physical Status 3 was associated with a statistically significant higher risk for delirium. There was no statistically significant association between sex, anesthetic regimen, and emergence time and delirium. The group excluded from EEG analysis (n = 31 [16%]) did not show a statistically significant different incidence of delirium (n = 4 [13%]) from the included group (Fisher’s exact probability = 0.66, not significant at P > 0.05). There was no demographic data missing for the included patients. There was no follow-up after the stay in the PACU.

Figure 2 shows box plots for the emergence slope distributions for total power (fig. 2A) and the power in the different bands (fig. 2, B to E). Patients who did not develop delirium tended to show a steeper negative slope for total power and across all bands during the emergence phase than patients who developed delirium. Furthermore, in total power and in all bands, the slopes were more often negative than positive in the group without delirium. Effect sizes in the form of AUCs ranged between 0.64 (95% CI, 0.52 to 0.74) for the delta band and 0.67 (95% CI, 0.58 to 0.77) for the alpha band. All AUC curves including age are shown in Supplemental Digital Content 4 (https://links.lww.com/ALN/D291). The results of the analysis for autocorrelation in the residuals with the Durbin–Watson test are shown in the table in Supplemental Digital Content 5 (https://links.lww.com/ALN/D292).

Figure 3 shows the box plots for the distributions for total starting power (fig. 3A) and the starting power in the different bands (fig. 3, B to E). Similar to the slope values in figure 2, patients who did not develop delirium had a significantly higher starting power across all bands and in total power than those who did develop delirium. This is in accordance with our previously published results.⁸  The corresponding R² values are reported in Supplemental Digital Content 6 (https://links.lww.com/ALN/D293). For means of sensitivity analysis, Supplemental Digital Content 7 (https://links.lww.com/ALN/D294) shows the box plots similar to figure 3 with all data sets included (n = 192). There are no significant differences between the two groups.

The tables show the group sizes (Table 2) and the risk ratios (Table 3) for delirium for patients with either significantly rising total power or band power and for patients without significant change in total power or band power during emergence. Patients with increasing power either across the EEG bands or just in singular bands had an approximately two-fold risk of developing delirium when compared to patients with decreasing EEG power. There was no significant difference in risk between patients with decreasing EEG power and patients without significant change of EEG power during emergence.

Figure 4 shows a scatter plot of slope value pairs for alpha band and beta band power for each patient depending on delirium status. In the lower left quadrant where alpha and beta power are decreasing, there are mostly nondelirious patients. This is also shown in the distribution plots in figure 4A, in which the mean slope value for both alpha and beta slope is negative. A detailed view is presented in figure 4B. Supplemental Digital Content 8 (https://links.lww.com/ALN/D295) shows the distributions for alpha and beta for all patients, including those with no significant change in power in beta and alpha. Furthermore, figure 2, D and E, show a negative median value for both bands.

Table 4 shows the combinations of changes in EEG band power, the corresponding risk ratios, and Fisher’s exact probability. A falling band power in both the alpha and beta band was associated with the lowest risk for delirium and was therefore set as reference. This means that a negative slope in both the alpha and beta band is highly specific for patients who wake up without delirium. Patients who showed an increase in band power in at least one band had a significantly higher risk for delirium. A test for patients at low risk for delirium is shown in Supplemental Digital Content 9 (https://links.lww.com/ALN/D296). The calculated test shows a specificity of 90.6%, a sensitivity of 51.2%, a positive predictive value of 95.9%, a negative predictive value of 30.3%, and a P value < 0.001. The corresponding 10-k bootstrapped AUC equals 0.69 (95% CI, 0.64 to 0.74).

Figure 5 shows the cumulative normalized median spectrograms for the good trajectory represented in table 4 by A−/B− and the bad trajectory (A+/B+) groups, each for patients with and without a positive Confusion Assessment Method for ICU. Figure 5, A to C, show the bad trajectory group, with figure 5C showing spectral differences mostly in the alpha band during the first two thirds of emergence. Figure 5, D and E, show the good trajectory group, with figure 5F showing differences in the alpha band in the first half of emergence. There is significant difference across all bands between the good trajectory group and bad trajectory group in the second half of emergence for patients without delirium (fig. 5H). There are no significant differences in the group with delirium between the good and bad trajectory group during emergence (fig. 5G).

We present a novel method based on the slope of EEG power during emergence that could help to identify patients at low risk for a perioperative neurocognitive disorder, specifically delirium, and that is easily transferable into a clinical setting.⁷  The fact that patients can show a multitude of different EEG patterns during anesthesia emergence has been reported previously.^17,18  Thereby, the classification of different EEG emergence trajectories was based on arbitrarily defined thresholds of EEG (band) power that allowed classification of the EEG into delta-dominant anesthesia, spindle-dominant anesthesia, or non–slow-wave anesthesia.^5,17  This approach helped to relate the anesthesia emergence states to sleep states.¹⁷  However, a generalization of this concept to intraoperative monitoring may be complicated, as recovery from sleep and anesthesia are different. In general, commercial EEG-based patient monitoring predominantly relies on quantitative changes in EEG band power or their ratios.^19–21  Furthermore, during emergence from general anesthesia induced by aminobutyric acid–mediated (GABAergic) agents, patients transition from alpha oscillations to beta oscillations, and the slow-wave delta oscillations should disappear in an uneventful case; this was termed a “zipper opening.”²²  The consequence of this zipper-opening behavior would then lead to the universal decrease in power that we describe as a favorable change. Hence, we decided to evaluate the changes in total EEG power and absolute EEG band power during anesthesia emergence. During the transition from responsiveness to unresponsiveness or unconsciousness, the EEG changes from a fast and low-amplitude signal to slow and high-amplitude rhythmic activity.²³  Although we know that the loss and the return of responsiveness are not mirrored processes, the EEG should in the best case return to a fast signal with low amplitude and a high frequency—as we observed.²⁴  If the EEG behaves differently, the patient is at significantly higher risk to develop delirium. Synchronized oscillations of neurons from up to down states are represented by high-amplitude delta waves and are the most recognizable feature of general anesthesia. Higher frequency waves (alpha and beta) may reflect more sophisticated communication among cortical cells. Failure of cortical information processing of complex information may be part of the mechanism of delirium given that the differences between the delirium and no-delirium groups were mostly in the higher power oscillations. The patient population with the lowest risk for delirium was the one with a significant decrease in alpha band and beta band power throughout emergence, while patients with increasing power in alpha, beta, or both had a higher risk. Calculating risk differences between the higher risk groups was not feasible, as the sample size was not large enough. In a preliminary analysis, we looked at different combinations of bands, including alpha and delta. However, the combination of the alpha and the beta band showed the most promising results.

High intraoperative EEG alpha band power seems associated with an adequate anesthetic level^25,26  and with the preoperative and perioperative cognitive state of the patient.^27,28  Also, during anesthesia emergence, an episode of alpha-dominant activity seems beneficial for the patient.^8,17  However, during a smooth emergence, the alpha power should fade and hence decrease as indicated by the negative slope. For the desired change in EEG beta band power, also a decrease, the explanation is not as straightforward. In fact, it seems counterintuitive, because strong beta band activity seems related to wakefulness, but it must be remembered that this application is in a clinical setting in patients recovering from a surgical intervention. This inevitably leads to patients that at some point will start moving. EMG activity is known to influence clinical EEG recordings as its frequencies overlap with EEG.²⁹  Still, the EMG may become more dominant in the higher frequencies, i.e., in the beta band and gamma band. The unfavorable increasing trend in beta band activity may be caused by a more agitated patient. Patients with an endotracheal tube in place showed earlier signs of EMG activation than patients with a laryngeal mask.³⁰  However, this needs to be investigated in more detail in the future. We intended to design a method that helps to identify with high precision patients without increased risk for delirium. Hence, a possible inclusion of EMG activity is acceptable, as also integrated in the algorithms of the Entropy Module (GE, Finland).³¹  There are other methods that try to classify anesthesia emergence. For instance, generic algorithm support vector machine approaches on EEG band power showed that emergence patterns are age specific.³²  A linear curve fit is the easiest, most economic approach and describes a general trend over a period of time, in this case the emergence phase. A sigmoidal or quadratic fit or a curve of higher order could provide a better fit during the interval of the emergence phase. However, these approaches could cause a higher number of resulting parameters, as higher-order curves are defined by a higher number of parameters. In this case, an analysis with vector machines could be necessary, resulting in the need for a more sophisticated approach. Last, our goal was to find an approach that is easily transferable into a clinical setting, thereby possibly offering the anesthesiologist an early sign for the neurocognitive outcome of the patient.

The statistical trends we present here may not be entirely based on the spectral information resulting from brain network function, as in our study we did not assess the preoperative cognitive state apart from a screening for delirium. Therefore, the results we present have at some point to be validated against such a score. First, it is known that patients with cortical atrophy are more likely to have lower total EEG power during maintenance and therefore also more likely to exhibit a flatter slope. Additionally, discontinuous EEG patterns indicative of excessive hypnotic administration (e.g., burst suppression) have low total EEG power and are not only more likely to occur in older patients; they may also present a risk factor for a delirium, although young and healthy subjects may not be affected.^2,3,33  As we describe here, a low total EEG power at end maintenance was associated with delirium and can contribute to flatter (or positive) slopes during emergence. Similarly, a longer time to emergence can contribute to a flatter slope during emergence, and even though we did not find a statistical association between time spent in emergence and delirium in our study, others have reported on this, and it may contribute to the strength of our correlation of this parameter with delirium.¹⁷  These points are important to put into the context of the demonstrated correlation of this EEG parameter with delirium, and therefore the frequency information should not be misinterpreted as a complete mechanistic explanation of the brain that is more susceptible to developing delirium. Furthermore, we did not exclude patients with intraoperative burst suppression, which also results in a lower power in the EEG. While the results presented here are very promising, we openly address the fact that given our limited data set and the nature of the secondary retrospective analysis, our ability to improve the performance of our new approach is limited by hard boundaries. This approach is not ready to be used in a clinical setting, as the reliability in predictive quality and diagnostic accuracy are still bound by the nature of the used data set and the inability to fully adhere to the TRIPOD and STARD guidelines. While we internally validated our data by means of bootstrapping, we want to explicitly emphasize that external validation is needed in the future to further clarify the results we present in this article. While we saw no differences in our sensitivity analysis, they are based partly on manually cleaned data and excluded data and could thereby impede the sensitivity of the study. To make the approach usable in a clinical setting, a robust artifact removal has to be implemented, and a real-time application has to be tested in the future. Furthermore, our patient collective is predominantly male, because the study was mostly conducted in our urological department, and follow-up studies should investigate how the sex of the patient would influence the result. To quantify risk more precisely, a larger study group is needed. Our results show a clear tendency, while still having wide CI. The use of Confusion Assessment Method for ICU may also contribute to the limitations because it was shown to be less sensitive compared to other screening methods (3-Minute Diagnostic Interview for Confusion Assessment Method-defined Delirium).³⁴  Therefore, its precision to detect delirium in the PACU may not be ideal. EEG changes induced by anesthetics are well described for propofol and fluoroether vapors. While we did not have any patients receiving a dexmedetomidine or ketamine anesthesia, there may be limitations to our method when using those, especially as the EEG patterns are vastly different.

In this study, we developed an easily applicable exploratory method to analyze a single frontal EEG channel and identify patterns highly specific for patients not at risk for delirium. This method could in the future help economize resources concerning the screening of patients for delirium in the PACU. Further analysis on this matter is needed in a multicenter prospective setting to improve the prediction accuracy and develop a predictive model reliable for clinical application. Furthermore, we think that similar approaches in the future will be able to identify changes in EEG band power and find patterns that are more specific for delirium and thereby lead the focus to patients at risk.

The authors thank Dr. Sebastian Berger, Technical University of Munich, München, Germany, for creating custom MATLAB scripts.

Supported in part by a grant from the James S. McDonnell Foundation (St. Louis, Missouri; to Dr. García).

Provisional patents have been filed on this work (U.S. Patent Application No. 63/397,667 and U.S. Patent Application No. 63/459,294). Dr. Sleigh is an editor for Anesthesiology. Drs. Kreuzer and García are co-inventors on several patents related to intraoperative EEG analysis owned by Columbia University (New York, New York) and the Technical University of Munich (Munich, Germany). Dr. García is a co-founder of a company, Lantern Laboratories, Inc. (New York, New York), that has a license to build software and hardware for intraoperative monitoring. The other authors declare no competing interests.

Supplemental Digital Content 1. Patient inclusion flow chart protocol, https://links.lww.com/ALN/D288

Supplemental Digital Content 2. Electrode layout for EEG measurement, https://links.lww.com/ALN/D289

Supplemental Digital Content 3. Example of analysis in patient with delirium, https://links.lww.com/ALN/D290

Supplemental Digital Content 4. AUCs for total/band power and age, https://links.lww.com/ALN/D291

Supplemental Digital Content 5. Results of test for autocorrelation, https://links.lww.com/ALN/D292

Supplemental Digital Content 6. R² values for linear regression, https://links.lww.com/ALN/D293

Supplemental Digital Content 7. Slope analysis for all patients, https://links.lww.com/ALN/D294

Supplemental Digital Content 8. Scatter plot for alpha and beta slope, https://links.lww.com/ALN/D295

Supplemental Digital Content 9. Possible test for no-delirium, https://links.lww.com/ALN/D296