1 Department of Anaesthesiology, Universität Erlangen-Nürnberg, Germany. 2 Department of Anaesthesiology, Technische Universität München, Germany. 3 Department of Anaesthesiology, Universität München, Germany. 4 Department of Information Technology, Universität Duisburg-Essen, Germany. 5 Department of Medical Psychology, Universität München, Germany
* Corresponding author. Klinik für Anästhesiologie, Krankenhausstraße 12, D-91054 Erlangen, Germany. E-mail: schwilden{at}kfa.imed.uni-erlangen.de
Accepted for publication December 22, 2004.
![]() |
Abstract |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
Methods. A total of 81 variables (31 EEG, 22 SSEP, 28 AEP) were simultaneously recorded in 48 surgical patients during anaesthesia. A total of 307 cases of the 81 variables in stable anaesthetic states were recorded. A factor analysis was performed for this data set.
Results. Sixteen variables were excluded because of multicollinearity. We extracted 13 factors with eigenvalues >1, representing 78.3% of the total variance, from the remaining 65 x 307 matrix. The first three factors represented 12%, 11% and 10% of the total variance. Factor 1 had only significant loadings from EEG variables, factor 2 only significant loadings from AEP variables and factor 3 only significant loadings from SSEP variables.
Conclusion. EEG, AEP and SSEP measure different aspects of neural processing during anaesthesia. This gives rise to the hypothesis that simultaneous monitoring of these quantities may give additional information compared with the monitoring of each quantity alone.
Keywords: factor analysis ; monitoring, auditory evoked potentials ; monitoring, EEG ; monitoring, somatosensory evoked potentials
![]() |
Introduction |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
During the past decade it has become evident that anaesthesia can be caused by different mechanisms4 5 and that the anaesthetic state itself should be described in a multi- dimensional rather than one-dimensional state-space.6 Monitoring the anaesthetic effect by a number between 0 and 100 has the attraction of being simple but it may be too simplistic. Within the framework of electrophysiological approaches, electrical brain activity such as spontaneous EEG, auditory evoked potentials (AEP)7 8 or somatosensory evoked potentials (SSEP)9 10 have been advocated for the monitoring of anaesthesia. Monitoring derivations of the spontaneous EEG have the longest tradition.11 12 This poses the question of whether the addition of descriptors of AEP and SSEP to this monitoring procedure adds substantial information to the monitoring of the anaesthetic effect by the spontaneous EEG alone.
The aim of this study was to perform a factor analysis for a number of derivations taken from the spontaneous EEG, AEP and SSEP under surgical conditions in anaesthetized patients and to determine whether, and to what degree, AEP and SSEP variables make a relevant contribution to the loadings of the extracted factors.
![]() |
Methods |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
Anaesthetic procedure
The anaesthetic procedure was left to the discretion of the attending anaesthetist and performed according to the daily clinical practice of the institution. In all cases, the following procedures were undertaken. An intravenous cannula was placed and infusion of lactated Ringer's solution started prior to administration of anaesthesia. After preoxygenation via a facemask, opioid was administered and anaesthesia was induced with either propofol or thiopental; muscle relaxation was achieved with atracurium. The trachea was intubated and the lungs were ventilated mechanically to maintain normocapnia with a mixture of oxygen in air appropriate to maintain normoxaemia. Drug dosing was maintained according to clinical observation of the patient. Repetitive doses of opioid were administered whenever inadequate analgesia was assumed. The administration of propofol or volatile anaesthetics was increased in the case of undesired patient movements. No further doses of muscle relaxants were administered after induction. At the end of surgery, anaesthetic agents were discontinued. When the patient regained apparent consciousness, the endotracheal tube was removed, the monitors were disconnected and the patient was transferred to the postanaesthesia care unit.
Signal acquisition and data selection
A specially adapted four-channel electrophysiological recording system (Neuroscreen, Viasys Healthcare, Hoechberg, Germany; EEG Infinity POD, Siemens, Erlangen, Germany) was used for signal acquisition. Before placement of Ag/AgCl EEG electrodes at F1, F2, C3', M2, Fz (reference) and Fp1 (ground) (electrode positions according to the international 1020 system), the skin was prepared with alcohol to keep impedances <5 k. The raw signal was digitized at a rate of 4 kHz per channel and stored on hard disk. SSEP stimuli were applied to the right median nerve via an electrode on the right wrist. The appropriate stimulus site and threshold were determined before induction of anaesthesia. Stimulus intensity was chosen 20% above threshold for motor twitch, and stimuli of duration 100 µs were applied with a frequency between 3 and 3.098 Hz. AEP stimulation used rarefraction clicks with a duration of 100 µs, 90 dB (SPL) and a repetition rate of 9.3 Hz, which were presented binaurally via shielded earphones (Viasys Healthcare, Hoechberg, Germany). Each AEP epoch consisted of 1000 stimuli, and each SSEP epoch consisted of 306 stimuli. EEG (i.e. no stimulus), AEP and SSEP stimuli were sequentially applied and evenly distributed. Each modality was measured for 1.7 min. In each patient, the order of stimulus modalities was randomly chosen for nine intervals, and these intervals were constantly repeated during anaesthesia.
We extracted 31 variables from the spontaneous EEG, 22 from the SSEP and 28 from the AEP for factor analysis. A detailed description of all variables is given in the Appendix. Ideally, the spontaneous EEG, SSEP and AEP should be measured simultaneously. As this is not possible, we defined parameter triplets as tightened values of variables from the spontaneous EEG, SSEP and AEP if the following criteria were satisfied: (i) the epoch proved to be artifact free on application of an artificial neuronal net for EEG artifact recognition, (ii) the epochs were recorded in close temporal succession and (iii) the clinical state of anaesthetic depth was constant within the epochs. Therefore no bolus doses of anaesthetic were given for up to 2.5 min before the period, constant anaesthetic infusions or concentrations of volatile anaesthetics were maintained for at least 2.5 min before the period, no prominent surgical stimuli were applied and patients were haemodynamically stable. Thus a parameter triplet is a collection of 81 parameters recorded during such a period (stable anaesthetic period) consisting of 31 parameters from the EEG, 22 from the SSEP and 28 from the AEP.
Factor analysis
Before extraction of the principal factors, the data matrix was screened for the presence of outliers, absence of multicollinearity and factorability of the correlation matrices. The criterion for multivariate outliers was a Mahalanobis distance at P<0.001, evaluated as 2 with the number of degrees of freedom equal to the number of variables,13 i.e. 81 for this investigation. Multicollinearity and singularity in the data matrix were identified by computing the squared multiple correlations (SMCs) for the variables, where each variable serves as the dependent variable, with the remaining variables as independent variables, in a multiple correlation. Variables with SMCs >0.99 were excluded from further analysis. Factorability of the correlation matrix was considered if there were several correlations >0.30.
![]() |
Results |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
|
Detailed factor analysis results
A total of 377 parameter triplets during stable anaesthesia periods were recorded for 48 of the 59 patients enrolled in the study. Seventy of the 377 parameter triplets were identified as multivariate outliers at a cut-off level of =0.001. These cases were deleted before extraction of the principal factors. Sixteen of the 81 variables were identified as multicollinear at a tolerance level of 0.01 (1-SMC) and were excluded from further analysis (Table 1).
|
|
|
|
|
![]() |
Discussion |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
As shown in Table 2, the extracted principal factors are mainly determined by subsets of variables calculated from different EEG recordings: six factors by spontaneous EEG variables, four factors by AEP variables and three factors by SSEP variables. None of the variables included in the factor interpretation overlapped on factors determined by variables of different sets. The first four extracted principal factors are determined by unique subsets of variables calculated from spontaneous or evoked EEGs and account for 42.4% of the variance in the data set in almost equal proportions (approximately 10% each). Taking into consideration the fact that the extracted principal factors are linearly independent of each other (orthogonality for vectors of factor loadings) and assuming the absence of a non-linear relationship between variables loading on at least the first four principal factors, we can assume that information extracted in each case from spontaneous and evoked EEGs represents different underlying processes.
The subset of spontaneous EEG variables which load on factor 1 is mainly engaged by power spectrum and entropy measures. The bicoherence variable Bic contributes only 3% to the total variance of this factor. As can be seen in the ordered variable sequence of Table 2, the naïve approximate entropy obtained using the Pincus algorithm, the 95th and 90th power quantiles and the relative power between 20 and 32 Hz load higher on this factor than on the rest (loadings >0.80 compared with loadings <0.70). Factors 5, 6 and 10 are mainly determined by different combinations of power spectrum measures from the spontaneous EEG. Factor 11 represents primarily the proportion of variance in the data set explained by the volume under the bispectrum (V29=BispV) which correlates positively with the skewness of amplitude histogram. This is plausible as the bispectrum is in essence the spectral decomposition of the skewness of the corresponding amplitude histogram. Factor 12 is loaded by the kurtosis of amplitude histogram, the EEG suppression ratio and the approximate entropy apEnt2. As the kurtosis of the amplitude histogram is related to the trispectrum, we can speculate that certain EEG patterns might cause a non-trivial trispectrum.
Without exemption, factor analysis separated the EEG variables discussed above from those derived from AEP and SSEP. Factor 2 is exclusively loaded by variables representing relative energies within level 5 of the wavelet decomposition. With high loadings (>0.7), in particular, the high and highest frequency contents load to this factor; the parameter representing a comparable low frequency band (V54=rL5P1) has the lowest factor loading (0.61).
Unlike factor 2, factor 4, although explaining a comparable amount of variance of all signals, is only loaded by variables representing absolute energies within the different levels of AEP wavelet decomposition. With one exception (V65=L5P0), the variables representing high and highest frequency contents load to this factor. Factor 7 is loaded both by variables of relative and absolute energies, representing a medium frequency band within the AEP. Factor 9, although only accounting for 3% of overall variance, contains only variables of the last two frequency bands below the lowpass of the filtering, representing both relative and absolute energies, while those of absolute energies yield higher factor loadings. This suggests that the high frequency contents may be altered differently from the rest of the AEP components.
Surprisingly, the variables representing the lowest frequency contents reveal only low threshold (V54=rL5P1 with 0.61 onto factor 2) or even subthreshold loadings on the factors. As substantial alterations of the low frequency contents of the AEP during anaesthesia are already known,14 this might be an indicator that possible variance of the signal was grouped within this parameter, consequently not leading to an independent factor with considerable factor loadings of these variables.
The SSEP variables load onto two distinct factors. Factor 3 is the major factor of the SSEP. Variables representing the medium frequency range are loading highest onto the factor. The highest frequency components also have high loadings on the factor. As a broad frequency range is included, the level of wavelet decomposition does not play the most important role in this parameter. Variables from level 2 decomposition yield similar loadings to variables in the corresponding level 5 decomposition subtrees. Factor 8 is only loaded by variables from level 5 decomposition. The two lowest frequency bands load highest to the factor. Surprisingly, the parameter representing the highest frequency range not cut off by the filters (V75=L5P11) also loads onto this factor, but only with a low loading just above threshold.
One should be careful when discussing the results in the context of possible clinical implications. Our results do not suggest that we should use an artificial construct like factor 1 as a variable to monitor anaesthesia; its usefulness would have to be tested in an appropriate clinical setting. However, our results indicate that the search for the best unique evoked or spontaneous EEG parameter is a search for a chimera. There exist numerous studies comparing, for instance, variables of the spontaneous EEG with variables from mid-latency AEP,15 16 ranking their usefulness and suggesting the use of one or other parameter. Our data support the view that EEG, AEP and SSEP measure different aspects of neural processing during anaesthesia, giving rise to the hypothesis that the simultaneous monitoring of these quantities may provide additional information compared with the monitoring of each quantity alone.
In conclusion, factor analysis revealed that the variance of variables automatically extracted from EEG, AEP and SSEP during general anaesthesia cannot be projected onto a single dimension. For the chosen variables, a set of 13 factors best explains the observed variance up to 80% of the total variance of all variables. Interestingly, none of the new derivate factors combines information from AEP, SSEP and EEG. Therefore it might be speculated that each of the three methods represents a different aspect of the changes in electrophysiological action of the central nervous system produced by general anaesthesia.
![]() |
Appendix |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
The 31 univariate EEG descriptors, calculated from each EEG epoch after DC shift subtraction, were as follows.
In order to obtain robust electroencephalographic estimators, in the final processing step we averaged the calculated descriptors over the available, consecutive electroencephalographic epochs within defined stable anaesthesia periods.
Somatosensory evoked potential (preprocessing, signal extraction and artifact reduction)
One SSEP was extracted from each SSEP stimulus sequence. First, the EEG was filtered with a bandpass of 3500 Hz (FIR filter with 125 coefficients) and the moving-average technique with seven samples was used to smooth the signal. SSEPs of 100 ms duration were generated from 306 single sweeps, i.e. EEG segments following the trigger. SSEPs without peaks N20 and P25, baseline drifts or signal shifts were excluded by visual inspection. These SSEPs underwent additional artifact filtering based on visual single-sweep analysis.
Auditory evoked potential (preprocessing, signal extraction and artifact reduction)
For extraction of the auditory evoked responses, EEG intervals of 100 ms duration were selected from the raw EEG and filtered with an FIR filter (bandpass 3500 Hz), starting with the corresponding trigger information. Out of each stimulation block consisting of 1000 single sweeps, signals with amplitudes >90% of the full scale of the amplifier were rejected from further analysis. All other sweeps were averaged within the post-stimulus interval of 100 ms. Averages consisting of less than 500 single sweeps were excluded from further analysis. The resulting averaged auditory evoked responses were subjected to a visual inspection of the signal; averages containing gross artifacts were excluded. As the early auditory evoked potentials within a post-stimulus interval of up 10 ms can also be recorded during deep general anaesthesia, the major peak within these signals, labelled V according to the taxonomy of Picton et al.,25 was detectable in all resulting averages.
Generation of AEP and SSEP parameters based on wavelet analysis
The methodology used to provide parameters characterizing the SSEPs and AEPs comprised the following three steps.
|
Wavelet packet transform can be understood as a generalization of the wavelet transform, such that high frequency components are also split. The iterative procedure results in a binary tree (Fig. A1).
Discrete wavelet analysis is performed by subsequently using filters approximating the wavelet function, also known as wavelet packet decomposition. The coefficients of each separated signal component are arranged in wavelet packets and assigned to a level in the binary tree according to the number of decomposition steps. The wavelet packets are related to disjoint frequency bands, whose width depends on the sampling frequency of the signal, the time and the frequency localization of the shifted and dilated version of the wavelet used. In the case of redundant WPDs, different subsets of packets can be chosen such that they give a complete representation of the signal.
Analysis of the evoked responses (AEP and SSEP) was performed using Matlab (MathWorks Inc., Natick, MA) and the affiliated Wavelet Toolbox. The averaged signals were decomposed up to level 5 by a wavelet package analysis using the Daubechies 4 wavelet.28
Selection of WPD coefficients
Two different selection procedures were used.
The aim of grouping is to find distinct groups of coefficients at each level of decomposition (and each packet). Each distinct group contains a number of coefficients, placed in successive positions, which have only either large or small absolute values. To discriminate between these two categories of coefficients, a statistical test is applied. Therefore each packet may contain one or more groups of coefficients. Packets with more than one group of coefficients (i.e. packets containing both large and small coefficients) are considered to contain more useful timefrequency information for characterizing the analysed signals.
After applying the clustering procedure separately, a combination of packets is selected using two variants. In the present study, the following subtrees were determined:
Computation of energy-based wavelet parameters
A number of meaningful parameters are selected for each signal analysed. Each packet consists of several coefficients and can be seen as a packet vector. The Euclidean norm, which is the square root of the energy of wavelet coefficients, is computed as a parameter for each vector. These parameters, which are based on the absolute energy content of the signal, are labelled LlPj, where l is the actual level of decomposition and j is the number of the packet (e.g. L5P3).
We computed the relative energies for the frequency bands corresponding to the wavelet packet vectors 011 of level 5, which is the sum of the squared coefficients per wavelet packet vector divided by the sum of all squared coefficients of level 5 (labelled rL5Pj).
![]() |
Acknowledgments |
---|
![]() |
Footnotes |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
![]() |
References |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
2 Ghouri AF, Monk TG, White PF. Electroencephalogram spectral edge frequency, lower esophageal contractility, and autonomic responsiveness during general anesthesia. J Clin Monit 1993; 9: 17685[CrossRef][ISI][Medline]
3 O'Flaherty D. Heart rate variability and anaesthesia. Eur J Anaesthesiol 1993; 10: 41932[ISI][Medline]
4 Mashour GA. Consciousness unboundtoward a paradigm of general anesthesia. Anesthesiology 2004; 100: 42833[CrossRef][ISI][Medline]
5 Urban BW. Current assessment of target and theories of anaesthesia. Br J Anaesth 2002; 89: 16783
6 Schwilden H. Neuropharmacology of anaesthetics. Curr Opin Anaesthesiol 1994; 7: 3269
7 Adler C, Pöppel E. Auditory evoked potentials indicate the loss of neuronal oscillations during general anaesthesia. Naturwissenschaften 1987; 74: 423[CrossRef][ISI][Medline]
8 Madler C, Keller I, Schwender D et al. Sensory information processing during general anaesthesia: effect of isoflurane on auditory evoked neuronal oscillations. Br J Anaesth 1991; 66: 817[Abstract]
9 Boisseau N, Madany M, Staccini P et al. Comparison of the effects of sevoflurane and propofol on cortical somatosensory evoked potentials. Br J Anaesth 2002; 88: 7859
10 Kochs E, Treede RD, Schulte am Esch J et al. Modulation of pain-related somatosensory evoked potentials by general anesthesia. Anesth Analg 1990; 71: 22530[Abstract]
11 Faulconer AJ, Pender JW, Bickford RG. The influence of partial pressure of nitrous oxide on the depth of anesthesia and the encephalogram in man. Anesthesiology 1949; 10: 6019[ISI]
12 Keirsey DK, Bickford RG, Faulconer AJ. Electroencephalographic patterns produced by thiopental sodium during surgical operations; description and classification. Br J Anaesth 1951; 23: 14152[Medline]
13 Tabachnick BG, Fidell LS. Using Multivariate Statistics, 4th edn. Boston, MA: Allyn & Bacon, 2001
14 Keller I, Madler C, Schwender D et al. Analysis of oscillatory components in perioperative AEP-recordings: a nonparametric procedure for frequency measurement. Clin Electroencephalogr 1990; 21: 8892[ISI][Medline]
15 Kreuer S, Bruhn J, Larsen R et al. Comparison of Alaris AEP index and bispectral index during propofol-remifentanil anaesthesia. Br J Anaesth 2003; 91: 33640
16 Nishiyama T, Matsukawa T, Hanaoka K. A comparison of the clinical usefulness of three different electroencephalogram monitors: bispectral Index, processed electroencephalogram, and Alaris auditory evoked potentials. Anesth Analg 2004; 98: 13415
17 Jeleazcov C, Egner S, Bremer F et al. Automated EEG preprocessing during anaesthesia: new aspects using artificial neural networks. Biomed Tech (Berl) 2004; 49: 12531[Medline]
18 Schwilden H, Jeleazcov C. Does the EEG during isoflurane/alfentanil anesthesia differ from linear random data? J Clin Monit Comput 2002; 17: 44957[CrossRef][Medline]
19 Schurmann T, Grassberger P. Entropy estimation of symbol sequences. Chaos 1996; 6: 41427[CrossRef][ISI][Medline]
20 Bruhn J, Ropcke H, Hoeft A. Approximate entropy as an electroencephalographic measure of anesthetic drug effect during desflurane anesthesia. Anesthesiology 2000; 92: 71526[CrossRef][ISI][Medline]
21 Pincus S. Approximate entropy (ApEn) as a complexity measure. Chaos 1995; 5: 11017[CrossRef][ISI][Medline]
22 Kaplan DT, Furman MI, Pincus SM et al. Aging and the complexity of cardiovascular dynamics. Biophys J 1991; 59: 9459[Abstract]
23 Rezek IA, Roberts SJ. Stochastic complexity measures for physiological signal analysis. IEEE Trans Biomed Eng 1998; 45: 118691[CrossRef][ISI][Medline]
24 Sigl JC, Chamoun NG. An introduction to bispectral analysis for the electroencephalogram. J Clin Monit 1994; 10: 392404[ISI][Medline]
25 Picton TW, Hillyard SA, Krausz HI et al. Human auditory evoked potentials. I. Evaluation of components. Electroencephalogr Clin Neurophysiol 1974; 36: 17990[CrossRef][ISI][Medline]
26 Daubechies I. The wavelet transform, timefrequency localization and signal analysis. IEEE Trans Inform Theory 1990; 36: 9611005[CrossRef][ISI]
27 Frazier M, Jawert B. A discrete transform and decomposition of distribution spaces. J Funct Anal 1990; 93: 34170[CrossRef][ISI]
28 Misiti M, Misiti Y, Oppenheim G et al. Wavelet Toolbox User's Guide for Use with MATLAB. Natick, MA: The MathWorks Inc., 1997
|