- Email: [email protected]
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA WITH ISOFLURANE
Br. J. Anaesth. (1989), 63, 36-43
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA WITH ISOFLURANE C. E. THOMSEN, K. N. CHRISTENSEN AND A. ROSENFALCK
CARSTEN ECKHART THOMSEN, M.SC.E.E.; ANNELISE ROSENFALCK, M.SC.E.E.; Department of Medical Informatics and
SUMMARY Changes in brain activity were studied at different depths of isoflurane anaesthesia. Ten healthy women (ASA group I) were investigated during non-critical surgery. Two channels of the EEG were stored on tape simultaneously with alveolar concentration of carbon dioxide, inspired oxygen concentration, mean arterial pressure, ECG and temperature. Signal processing was made offline. Spectral information from 2-s EEG segments was extracted using autoregressive modelling. Repetitive hierarchical clustering was used to define a common learning set of basic patterns. With this learning set, the EEG was classified, and the results presented in a class probability histogram. The basic patterns were related to the clinical depth of anaesthesia in all patients and assigned specific colours. Using this colour code, the class probability histogram showed a high degree of simplicity. Decreasing or increasing the isoflurane concentration caused the same trend in the class profile in all patients. This indicates that the EEG pattern might be a sensitive tool for decision making during administration of general anaesthetics.
to, EEG has not been useful clinically for monitoring depth of anaesthesia because the changes were agent specific and dependent on the critical state of the patient [13]. To overcome this problem, only patients belonging to ASA group I were selected and the same anaesthetic agent was used in all patients.
Image Analysis, Aalborg University, Badehusvej 23, DK-9000 Aalborg, Denmark. KURT NGRREGAARD CHRISTENSEN, M.D.,
MATERIALS AND METHODS
Department of Anaesthesia, Aalborg Hospital, Reberbansgade, DK-9000 Alborg, Denmark. Accepted for Publication: January 3, 1989. Correspondence to K.N.C.
We studied 10 healthy women (aged 25-45 yr, ASA I, body weight not exceeding 10% of ideal) undergoing simple hysterectomy. There was no
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
Monitoring of brain activity during surgery has been used mainly to assess cerebral activity during carotid endarterectomy and extracorporal circulation [1-5]. Another objective has been to correlate changes in brain activity with depth of anaesthesia. These investigations include both volatile and i.v. anaesthetic agents [4,6,7]. The aim of the present study was to investigate the changes in the electroencephalogram (EEG) during changes in depth of anaesthesia. A computer-based system was used off-line for signal analysis of the EEG and for visualizing and recording the EEG pattern and physiological variables monitored traditionally during anaesthesia. Special attention was payed to the design of the screen display, as pattern recognition of simultaneous changes in EEG and other physiological variables may, in future on-line use, serve to help the anaesthetist to make decisions. Visual inspection of the EEG strip chart recording is time consuming and requires the presence of an experienced neurophysiologist. Computer-based signal processing methods are useful [1—9] and have been used during surgery when brain metabolism was at risk; it was observed that extensive monitoring improved the neurological outcome for the patient [5,7]. The automatic methods use spectral analysis by means of Fast Fourier Transform (FFT) and period/amplitude analysis. The amplitude spectra obtained by the FFT have been displayed as compressed spectral array (CSA) plots [10,11], or as density spectral array (DSA) plots [12]. Hither-
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA
After operation all anaesthetics were discontinued and the neuromoscular block was antagonized with neostigmine. SIGNAL PROCESSING
The signals were analysed off-line in a laboratory using an Intel 310 computer system. The computer performs multi-tasking real-time programs, and is equipped with analog-to-digital converter (ADC: 12-bit resolution in eight channels) and a high-resolution colour graphic display system (1024x512 pixels and 256 programmable colours). Hard-copies of the displayed parameters were made on a Tektronix 4696 colour-graphic jet-ink printer. Clinical events were marked on the magnetic tape recorder during surgery and transferred to the computer system. The EEG signals were filtered with a 25-Hz, fourth order anti-aliasing filter, sampled at 100 Hz and pre-emphasized digitally by a first order high-pass filter. The EEG was divided into 2-s segments. From each segment, 11 parameters representing the amplitude and spectral information of the EEG were extracted using auto regressive (AR) modelling [14-16]. The pattern recognition was performed using a probabilistic classifier (Bayes' decision theory [17]). The classifier used a learning set of patterns defined by hierarchical clustering analysis. This type of learning is often referred to as "unsupervised learning". For this analysis we selected a set of approximately 1000 EEG segments representing typical EEG patterns from different stages of anaesthesia. During hierarchical clustering [17] the algorithm merged the two most likely segments into one, step by step, starting with the initial number of segments and ending with one. As in pattern recognition, the segments were characterized by 11 parameters derived by AR-modelling. The clustering analysis was represented graphically in a dendrogram (fig. 1 A). By interactive inspection of nodes in the dendrogram, where the dissimilarity measure between patterns increased rapidly, a representative set of 10-15 classes could be defined (fig. 1 B). These classes were used as learning set for the classification of EEG activity during on-line analysis. During analysis the results were displayed on the graphic screen. The last 1 h of anaesthesia was shown, but the whole period was stored for later inspection and production of hard copies for the
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
history of previous neurological disorder and tranquillizers and analgesics were not administered routinely for several weeks before operation. The day before operation, informed consent was obtained. The investigation was approved by the local Ethics Committee. Anaesthesia was induced with thiopentone 5 mg/kg body weight and, before tracheal intubation, a priming dose of gallamine, and suxamethonium 100 mg were administered. After intubation, anaesthesia was maintained with 1.0-2.5% isofiurane and 66% nitrous oxide in oxygen. Relaxation was maintained with gallamine and the lungs were ventilated mechanically by a Dameca MC 801 anaesthetic ventilator connected to a circle system with carbon dioxide absorption. The duration of anaesthesia was 50-150 min and the temperature was constant (±0.7 °C). Monitoring of EEG from both hemispheres (P3-Fp! and P4-Fp2) was started before induction of anaesthesia. In addition, the traditional physiological variables, ECG, heart rate and temperature, were recorded. End-tidal (alveolar) carbon dioxide concentration (FA CO) ) and inspiratory oxygen concentration (FiOi) were measured breath to breath with a Datex Normocap CO2 monitor. Mean arterial pressure (MAP) was recorded noninvasively by a Criticon Dinamap Vital Signs Monitor. All information was stored on a eightchannel Racal Store 8 frequency modulated magnetic tape recorder. Further analysis of signals was performed off-line. According to the procedure, the initial isofiurane vaporizer setting was selected by the same experienced anaesthetist in all patients, corresponding to clinical signs indicating depth of anaesthesia adequate for surgical needs. The loading and maintenance concentration was guided individually by the clinical condition of the patient. From induction of anaesthesia the mechanical ventilator was adjusted to FACC,2 5.0-5.5 vol%, corresponding to normoventilation. Twenty-five to 30 min after start of surgery, when the clinical condition was considered stable, the isofiurane concentration was increased by approximately 50 % to study the EEG at a greater depth of anaesthesia. After 20 min maintenance concentration of isofiurane was re-established. Thirty minutes later, slight hyperventilation was induced, corresponding to Fkc^ 4.0-4.5 vol%; this was discontinued after 20 min.
37
38
BRITISH JOURNAL OF ANAESTHESIA o 10 20
§ 30 ~ 40
I 50 (ft
| 60
I 70. 80 100 Figure 1 A
Figure 1 B
FIG. 1. A : Example of a dendrogram—a graphic description of the result from hierarchical clustering. The bottom level (level 1) in the dendrogram shows all segments as singleton clusters. At any level, after level 1, two clusters are merged to form a new cluster, and they stay together at all subsequent levels. The dendrogram is drawn to a similarity scale (ordinate) to show the similarity between the two clusters that are merged. This particular dendrogram is based on data from all patients in the group anaesthetized with isoflurane. The small amplitude spectra represent mean values for the marked subtrees, B: Enlarged set of mean spectra for the nodes marked in A. Abscissae: 0-50 Hz; ordinates arbitrary amplitude scale (spectral density) (range 0-1). All spectra are drawn to the same scale, comparable with respect to amplitude. A parametric description of this set of basic patterns acts as a common learning set for classification of EEG activity during surgery—common for all patients anaesthetized with isoflurane.
patient record. The parameters from the EEG analysis were displayed as a function of time: (a) as Colour Density Spectral Array plots, where each plot (line) represents the amplitude spectrum for the most frequently occurring EEG pattern during a 10-s period; (b) as a class probability histogram, where each plot (line) represents the relative occurrence of any class in percent—the class profile; (c) as the symmetry between hemispheres.
In addition, the trend curves for FiOt, FACO>, MAP and heart rate were displayed. In this way the correlation between EEG and the traditionally monitored physiological variables could be studied by pattern recognition. A colour-graphic hard copy could be printed on-line during signal processing or off-line after the analysis had been performed. Changes in depth of anaesthesia and surgical procedure were marked on the print.
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
90
39
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA Depth of anaesthesia Drowsy (dark blue) a
so
3O
so
< HS './; r\ f: Very light (light blue)
- / • •V
: : : : :Vj : : o
20
so
o
20
30
so
30
so
o
so
30
so
30
so
o
so
3a
Light (green) o
Normal (yellow) so
so
o
so
3a
so
. CL.HS9 .
Deep (magenta) O
SO
3O
O
SO
3O
SO
O
SO
3O
SO
Very deep (red) 5O
FIG. 2. The set of basic EEG patterns correlated to depth of anaesthesia and colour coded in order to give the anaesthetist support in his decision making.
A common learning set of EEG patterns representing isoflurane anaesthesia was obtained using "linked" clustering analysis: from the clustering sequences of the 10 individual patients we selected a step, where 50 different clusters were represented. By linking the 50 clusters from the 10 patients, we gathered 500 EEG patterns representing a typical common EEG sequence. After a new hierarchical clustering analysis, we obtained a common set of classes (fig. 1 B), which could be used for all patients anaesthetized with isoflurane. The dominant feature in the common learning set was the peak frequency and the configuration of the spectrum. RESULTS
In all patients the class probability histogram generated from the common learning set was related to the clinical events during anaesthesia. The classes found during light, medium and deep
levels of anaesthesia were tabulated. From this table a set of classes of frequency and amplitude patterns related monotonically to the depth of anaesthesia were extracted and assigned colours (fig. 2). A peak frequency of 10-12 Hz was found during light anaesthesia (blue and green). At medium anaesthesia the peak frequency shifted to 5-7 Hz (yellow), and a small peak was detected in the 15-20 Hz range. During deep anaesthesia the peak was found in the very low frequency band, 1-5 Hz (magenta and red). Figure 3 shows a black-and-white representation of a typical hard copy from the isoflurane series. Major events were marked to the left. Events of special interest are: induction with thiopentone (T): 20 min hyperventilation (between | and —); 20 min increased anaesthetic concentration (from 1.5 to 2.5 vol%) (between > and < ) ; decreasing anaesthetic concentration
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
30
BRITISH JOURNAL OF ANAESTHESIA
40
was considerable. It is important to notice that, although the basic depth of anaesthesia was variable, the shift in the class profile was always related to the vaporizer adjustments. For example, the reaction from one patient receiving 1 MAC could be green (light) changing to yellow (medium) when receiving 1.25 MAC; alternatively another patient's basic depth of anaesthesia could be magenta (deep) with 1 MAC and red (very deep) when changing to 1.25 MAC. This shows that the EEG is a more sensitive test than the clinical evaluation, during medium anaesthesia. Comparing clinical evaluation with the EEG analysis, it was obvious that, in some instances, the anaesthetist adjusted the isoflurane concentration, without knowing the result of signal processing. In these cases the EEG analysis "agreed" in the treatment of the patient. The computer program was not designed to detect burst suppression. A visual inspection of the EEG strip chart in periods with isoflurane concentrations exceeding 1.5 MAC showed no burst suppression. In all patients slight hyperventilation was induced according to the procedure. This caused no change in the class profile in any of the patients. As expected in a non-critical surgical procedure, the asymmetry parameter was unchanged in all patients. DISCUSSION
Intraoperative monitoring of the EEG activity has been used mainly during anaesthetic procedures in which cerebral perfusion and metabolism are at risk. Several types of monitoring system have shown to be of help as warning systems [18]. In contrast, the use of EEG monitoring as a guide to depth of anaesthesia is controversial and un-
FIG. 3. Black and white version of a hard copy record from a patient monitored during 2 h 45 min of isoflurane anaesthesia. Vertical axis is time. Data from the EEG are displayed in the three columns to the left, the traditional physiological parameters in the two columns to the right. Events during anaesthesia are marked along the time axis at the left: T = thiopentone; x, X = start and stop of anaesthesia; o = start of surgery; •(•, — = start and stop of hyperventilation; >, < = increase and decrease in isoflurane concentration. The left column shows a DSA plot for one of the EEG recordings (horizontal axis is frequency in Hz). The second column shows symmetry between the two hemispheres (red on colour record; left of added white line here) and spectral similarity between the actual EEG sequence and the most likely class in the learning set (green on colour record; right of added white line here). The column in the middle shows the class histogram of the classification of the EEG activity. On the original colour record, the classes are colour-coded according to the depth of anaesthesia as defined in figure 2; here, white areas correspond to yellow on the colour record, and labels A-E have been added to identify areas of dark blue, light blue plus green (A, E) and of magenta plus red (B-D). The fourth
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
from 1.5 to 1.0 vol% (second <) and again from 1.0 to 0.5 vol% (third <). Induction with thiopentone (T) caused a short presence of very low frequency, high amplitude activity, hardly noticeable in the hard copy because of low time resolution. Isoflurane was administered at x. The depth of anaesthesia increased gradually, and after 1 h the class probability histogram (central record) was dominated by the yellow classes (white areas in figure 3) as assigned to medium anaesthesia. At > , the isoflurane vaporizer setting was increased by 50%. This change was immediately reflected in the class histogram, which was now dominated by the magenta-plus-red classes (black in figure 3: area B) assigned to deep levels of anaesthesia. When maintenance concentration was re-established (<), the yellow class reappeared. The response time for EEG changes to these steps in isoflurane concentrations was approximately 10 min, which is fast compared with the normal interval between manual notes in the anaesthesia log. The shift to a red period (black in figure 3; areas C, D) after approximately 2 h of recording may indicate that the patient reached deeper anaesthesia when the surgical procedure was less painful. The effect of the decreases in isoflurane concentration at the second and third < is clearly reflected by the reappearance of the yellow dominance (white in figure 3) in the class histogram. Towards the end of the recording, the histogram shifted to blue and green classes (black area E in figure 3), indicating transition from medium to light, very light and drowsy state, similar to the shift noted at start of surgery (event o in figure 3; black area A). In the majority of patients the yellow, medium depth dominated, but the interpatient variability
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA
41
TIME • DRTE • O i STHRTl r»T. I D I F»T _ 6 , X S O F L U R A H E
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
column shows end-tidal carbon dioxide concentration (CO2/black on colour record: 0-10°,,) and inspiratory oxygen concentration (O2/red on colour record; more continuous line here: 0-50",,). The column to the right shows heart rate (HR/green on colour record, 0-150 beat min~'; the width of this line represents the beat to beat variation in heart rate) and mean arterial pressure (MAP/blue on colour record; thicker, continuous vertical lines here: 0-150 mm Hg). The screen display during on-line analysis includes the last 1 h of the recording. The traces at the top are the raw signals of EEG and ECG displayed in 2-s segments for quality control during recording. The frame to the right is patient identification. The numbers below the columns are instantaneous displays of the variables.
42
bral hypoxia. In the present work we found no information in these features related to depth of anaesthesia and we conclude that the sensitivity is too low inside the therapeutic range of isoflurane concentration. This indicates the necessity of implementing specific common sets of basic patterns corresponding to the actual combination of anaesthetics used. In our view, the generation of this reference material will be a major task in the construction of monitoring systems in the future. ACKNOWLEDGEMENT This study was supported by the Danish Technical Research Council and the Danish Medical Research Council. REFERENCES 1. Prior PF. EEG monitoring and evoked potentials in brain ischaemia. British Journal of Anaesthesia 1985; 57: 63-81. 2. Pronk RAF, Simons AJR. Automatic recognition of abnormal EEG activity during open heart surgery. In: Buser PA, Cobb WA, Okuma T, eds. Kyoto Symposia, EEG Suppl 36. Amsterdam: Elsevier Biomedical Press, 1982; 590-602. 3. Chiappa KH, Burke SR, Young RR. Results and electroencephalo-graphic monitoring during 367 carotid endarterectomies. Stroke 1979; 10: 381-388. 4. Blackshear WM, Vincenzo DC, Seifert KB, Connar RG. Advantages of continuous electroencephalographic monitoring during carotid artery surgery. Journal of Cardiovascular Surgery 1986; 27: 146-153. 5. Trojaborg W, Boysen G. Relation between EEG, regional cerebral blood flow and internal carotid artery pressure during carotid endarterectomy. Electroencephalography and Clinical Neurophysiology 1973; 34: 61-69. 6. Klein FF, Davis DA. The use of time domain analyzed EEG in conjunction with cardiovascular parameters for monitoring anesthetic levels. IEEE Transactions on Biomedical Engineering 1981; 28: 36-40. 7. Pronk RAF, Simons AJR, Ackerstaff RGA, Boezeman EHJF. Intra-operative EEG monitoring. Proceedings of the Ninth Annual Conference of the IEEE Engineering in Medicine and Biology Society 1987; 3: 1250-1251. 8. Schwilden H, Stoeckel H. Quantitative EEG analysis during anaesthesia with isoflurane in nitrous oxide at 1.3 and 1.5 MAC. British Journal of Anaesthesia 1987; 59: 738-745. 9. Pronk RAF. EEG Processing in Cardiac Surgery (Thesis). Amsterdam: Free University, 1982. 10. Bickford RG, Flemming NI, Billinger TW. Compression of EEG data by isometric power spectral plots. Electroencephalography and Clinical Neurophysiology 1971; 31: 632. 11. Bickford RG, Brimm J, Berger L, Aung M. In: Application of compressed spectral array in clinical EEG. Kellaway B, Petersen I, eds. Automation of Clinical Electroencephalography. New York: Raven Press, 1973; 55-64. 12. Flemming RA, Smith NT. Density modulation: A technique for the display of three-variable data in patient monitoring. Aneslhesiology 1979; 50: 543-546.
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
reliable [13]. The reason for this could be that most investigations have focused on recordings in which critical physiological variables show a considerable degree of instability. In the present work, our intention was to investigate the changes in brain activity related to the depth of anaesthesia in healthy subjects. Major efforts were made to keep basic physiological variables constant during monitoring. Many investigators have reported disappointing results, with no correlation between EEG frequency, amplitude and depth of anaesthesia. In particular, the changes have been shown to be specific for the anaesthetic agent used [7]. Nevertheless, many attempts have been made to visualize the EEG signal in a compressed and simplified form, such as CSA-plot or DSA-plot [10,11], in order to give the anaesthetist the opportunity to compare the clinical evaluation of anaesthetic depth with an on-line signal interpretation. Using computerized signal processing, a great number of variables and features can be extracted, which represent information undetectable by simple visual inspection of the EEG signal. The present investigation of the EEG involves autoregressive modelling and clustering analysis in order to define a set of basic classes of activity, representing different depths of anaesthesia. The set of basic classes was used as learning set for classification of EEG activity. The presentation of the result of classification in a class probability histogram demonstrated a high degree of simplicity. At the different depths of anaesthesia more than two classes were seldom represented. The learning set was common and derived specifically from the isoflurane series. By this technique, the class histogram shows a significant profile of class occurrence related to the depth of anaesthesia. The difference seen in some patients between the profile of the class histogram and the clinical level of anaesthesia reflects the fact that the features displayed are more sensitive than the clinical evaluation. Even though some patients had a different class profile during the same clinical level of anaesthesia, a change to lighter or deeper clinical levels caused a stereotyped shift in class histogram to profiles representing the same trend. In previous investigations it has been found that spectral edge frequency, centre frequency and mean energy content in the conventional frequency bands are valuable indicators of cere-
BRITISH JOURNAL OF ANAESTHESIA
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA
43
13. Editorial. Lancet 1986; 1: 553-554. mation of short segment spectra for computerized EEG 14. Makhoul J. Linear prediction: A tutorial review. Proanalysis. IEEE Transactions on Biomedical Engineering ceedings IEEE 1975; 63: 561-580. 1981; 28: 630-638. 15. Jansen BH. Analysis of biomedical signals by means of 17. Duda RO, Hart PE. Pattern Classification and Scene linear modelling. CRC Critical Reviews in Biomedical Analysis. New York: John Wiley & Sons Ltd, 1973. Engineering 1985; 12: 343-392. 18. Sebel P. Monitoring the central nervous system. Current 16. Jansen BH, Bourne JR, Ward JW. Autoregressive estiMedical Literature 1987; 1: 33-37.
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA WITH ISOFLURANE C. E. THOMSEN, K. N. CHRISTENSEN AND A. ROSENFALCK
CARSTEN ECKHART THOMSEN, M.SC.E.E.; ANNELISE ROSENFALCK, M.SC.E.E.; Department of Medical Informatics and
SUMMARY Changes in brain activity were studied at different depths of isoflurane anaesthesia. Ten healthy women (ASA group I) were investigated during non-critical surgery. Two channels of the EEG were stored on tape simultaneously with alveolar concentration of carbon dioxide, inspired oxygen concentration, mean arterial pressure, ECG and temperature. Signal processing was made offline. Spectral information from 2-s EEG segments was extracted using autoregressive modelling. Repetitive hierarchical clustering was used to define a common learning set of basic patterns. With this learning set, the EEG was classified, and the results presented in a class probability histogram. The basic patterns were related to the clinical depth of anaesthesia in all patients and assigned specific colours. Using this colour code, the class probability histogram showed a high degree of simplicity. Decreasing or increasing the isoflurane concentration caused the same trend in the class profile in all patients. This indicates that the EEG pattern might be a sensitive tool for decision making during administration of general anaesthetics.
to, EEG has not been useful clinically for monitoring depth of anaesthesia because the changes were agent specific and dependent on the critical state of the patient [13]. To overcome this problem, only patients belonging to ASA group I were selected and the same anaesthetic agent was used in all patients.
Image Analysis, Aalborg University, Badehusvej 23, DK-9000 Aalborg, Denmark. KURT NGRREGAARD CHRISTENSEN, M.D.,
MATERIALS AND METHODS
Department of Anaesthesia, Aalborg Hospital, Reberbansgade, DK-9000 Alborg, Denmark. Accepted for Publication: January 3, 1989. Correspondence to K.N.C.
We studied 10 healthy women (aged 25-45 yr, ASA I, body weight not exceeding 10% of ideal) undergoing simple hysterectomy. There was no
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
Monitoring of brain activity during surgery has been used mainly to assess cerebral activity during carotid endarterectomy and extracorporal circulation [1-5]. Another objective has been to correlate changes in brain activity with depth of anaesthesia. These investigations include both volatile and i.v. anaesthetic agents [4,6,7]. The aim of the present study was to investigate the changes in the electroencephalogram (EEG) during changes in depth of anaesthesia. A computer-based system was used off-line for signal analysis of the EEG and for visualizing and recording the EEG pattern and physiological variables monitored traditionally during anaesthesia. Special attention was payed to the design of the screen display, as pattern recognition of simultaneous changes in EEG and other physiological variables may, in future on-line use, serve to help the anaesthetist to make decisions. Visual inspection of the EEG strip chart recording is time consuming and requires the presence of an experienced neurophysiologist. Computer-based signal processing methods are useful [1—9] and have been used during surgery when brain metabolism was at risk; it was observed that extensive monitoring improved the neurological outcome for the patient [5,7]. The automatic methods use spectral analysis by means of Fast Fourier Transform (FFT) and period/amplitude analysis. The amplitude spectra obtained by the FFT have been displayed as compressed spectral array (CSA) plots [10,11], or as density spectral array (DSA) plots [12]. Hither-
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA
After operation all anaesthetics were discontinued and the neuromoscular block was antagonized with neostigmine. SIGNAL PROCESSING
The signals were analysed off-line in a laboratory using an Intel 310 computer system. The computer performs multi-tasking real-time programs, and is equipped with analog-to-digital converter (ADC: 12-bit resolution in eight channels) and a high-resolution colour graphic display system (1024x512 pixels and 256 programmable colours). Hard-copies of the displayed parameters were made on a Tektronix 4696 colour-graphic jet-ink printer. Clinical events were marked on the magnetic tape recorder during surgery and transferred to the computer system. The EEG signals were filtered with a 25-Hz, fourth order anti-aliasing filter, sampled at 100 Hz and pre-emphasized digitally by a first order high-pass filter. The EEG was divided into 2-s segments. From each segment, 11 parameters representing the amplitude and spectral information of the EEG were extracted using auto regressive (AR) modelling [14-16]. The pattern recognition was performed using a probabilistic classifier (Bayes' decision theory [17]). The classifier used a learning set of patterns defined by hierarchical clustering analysis. This type of learning is often referred to as "unsupervised learning". For this analysis we selected a set of approximately 1000 EEG segments representing typical EEG patterns from different stages of anaesthesia. During hierarchical clustering [17] the algorithm merged the two most likely segments into one, step by step, starting with the initial number of segments and ending with one. As in pattern recognition, the segments were characterized by 11 parameters derived by AR-modelling. The clustering analysis was represented graphically in a dendrogram (fig. 1 A). By interactive inspection of nodes in the dendrogram, where the dissimilarity measure between patterns increased rapidly, a representative set of 10-15 classes could be defined (fig. 1 B). These classes were used as learning set for the classification of EEG activity during on-line analysis. During analysis the results were displayed on the graphic screen. The last 1 h of anaesthesia was shown, but the whole period was stored for later inspection and production of hard copies for the
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
history of previous neurological disorder and tranquillizers and analgesics were not administered routinely for several weeks before operation. The day before operation, informed consent was obtained. The investigation was approved by the local Ethics Committee. Anaesthesia was induced with thiopentone 5 mg/kg body weight and, before tracheal intubation, a priming dose of gallamine, and suxamethonium 100 mg were administered. After intubation, anaesthesia was maintained with 1.0-2.5% isofiurane and 66% nitrous oxide in oxygen. Relaxation was maintained with gallamine and the lungs were ventilated mechanically by a Dameca MC 801 anaesthetic ventilator connected to a circle system with carbon dioxide absorption. The duration of anaesthesia was 50-150 min and the temperature was constant (±0.7 °C). Monitoring of EEG from both hemispheres (P3-Fp! and P4-Fp2) was started before induction of anaesthesia. In addition, the traditional physiological variables, ECG, heart rate and temperature, were recorded. End-tidal (alveolar) carbon dioxide concentration (FA CO) ) and inspiratory oxygen concentration (FiOi) were measured breath to breath with a Datex Normocap CO2 monitor. Mean arterial pressure (MAP) was recorded noninvasively by a Criticon Dinamap Vital Signs Monitor. All information was stored on a eightchannel Racal Store 8 frequency modulated magnetic tape recorder. Further analysis of signals was performed off-line. According to the procedure, the initial isofiurane vaporizer setting was selected by the same experienced anaesthetist in all patients, corresponding to clinical signs indicating depth of anaesthesia adequate for surgical needs. The loading and maintenance concentration was guided individually by the clinical condition of the patient. From induction of anaesthesia the mechanical ventilator was adjusted to FACC,2 5.0-5.5 vol%, corresponding to normoventilation. Twenty-five to 30 min after start of surgery, when the clinical condition was considered stable, the isofiurane concentration was increased by approximately 50 % to study the EEG at a greater depth of anaesthesia. After 20 min maintenance concentration of isofiurane was re-established. Thirty minutes later, slight hyperventilation was induced, corresponding to Fkc^ 4.0-4.5 vol%; this was discontinued after 20 min.
37
38
BRITISH JOURNAL OF ANAESTHESIA o 10 20
§ 30 ~ 40
I 50 (ft
| 60
I 70. 80 100 Figure 1 A
Figure 1 B
FIG. 1. A : Example of a dendrogram—a graphic description of the result from hierarchical clustering. The bottom level (level 1) in the dendrogram shows all segments as singleton clusters. At any level, after level 1, two clusters are merged to form a new cluster, and they stay together at all subsequent levels. The dendrogram is drawn to a similarity scale (ordinate) to show the similarity between the two clusters that are merged. This particular dendrogram is based on data from all patients in the group anaesthetized with isoflurane. The small amplitude spectra represent mean values for the marked subtrees, B: Enlarged set of mean spectra for the nodes marked in A. Abscissae: 0-50 Hz; ordinates arbitrary amplitude scale (spectral density) (range 0-1). All spectra are drawn to the same scale, comparable with respect to amplitude. A parametric description of this set of basic patterns acts as a common learning set for classification of EEG activity during surgery—common for all patients anaesthetized with isoflurane.
patient record. The parameters from the EEG analysis were displayed as a function of time: (a) as Colour Density Spectral Array plots, where each plot (line) represents the amplitude spectrum for the most frequently occurring EEG pattern during a 10-s period; (b) as a class probability histogram, where each plot (line) represents the relative occurrence of any class in percent—the class profile; (c) as the symmetry between hemispheres.
In addition, the trend curves for FiOt, FACO>, MAP and heart rate were displayed. In this way the correlation between EEG and the traditionally monitored physiological variables could be studied by pattern recognition. A colour-graphic hard copy could be printed on-line during signal processing or off-line after the analysis had been performed. Changes in depth of anaesthesia and surgical procedure were marked on the print.
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
90
39
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA Depth of anaesthesia Drowsy (dark blue) a
so
3O
so
< HS './; r\ f: Very light (light blue)
- / • •V
: : : : :Vj : : o
20
so
o
20
30
so
30
so
o
so
30
so
30
so
o
so
3a
Light (green) o
Normal (yellow) so
so
o
so
3a
so
. CL.HS9 .
Deep (magenta) O
SO
3O
O
SO
3O
SO
O
SO
3O
SO
Very deep (red) 5O
FIG. 2. The set of basic EEG patterns correlated to depth of anaesthesia and colour coded in order to give the anaesthetist support in his decision making.
A common learning set of EEG patterns representing isoflurane anaesthesia was obtained using "linked" clustering analysis: from the clustering sequences of the 10 individual patients we selected a step, where 50 different clusters were represented. By linking the 50 clusters from the 10 patients, we gathered 500 EEG patterns representing a typical common EEG sequence. After a new hierarchical clustering analysis, we obtained a common set of classes (fig. 1 B), which could be used for all patients anaesthetized with isoflurane. The dominant feature in the common learning set was the peak frequency and the configuration of the spectrum. RESULTS
In all patients the class probability histogram generated from the common learning set was related to the clinical events during anaesthesia. The classes found during light, medium and deep
levels of anaesthesia were tabulated. From this table a set of classes of frequency and amplitude patterns related monotonically to the depth of anaesthesia were extracted and assigned colours (fig. 2). A peak frequency of 10-12 Hz was found during light anaesthesia (blue and green). At medium anaesthesia the peak frequency shifted to 5-7 Hz (yellow), and a small peak was detected in the 15-20 Hz range. During deep anaesthesia the peak was found in the very low frequency band, 1-5 Hz (magenta and red). Figure 3 shows a black-and-white representation of a typical hard copy from the isoflurane series. Major events were marked to the left. Events of special interest are: induction with thiopentone (T): 20 min hyperventilation (between | and —); 20 min increased anaesthetic concentration (from 1.5 to 2.5 vol%) (between > and < ) ; decreasing anaesthetic concentration
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
30
BRITISH JOURNAL OF ANAESTHESIA
40
was considerable. It is important to notice that, although the basic depth of anaesthesia was variable, the shift in the class profile was always related to the vaporizer adjustments. For example, the reaction from one patient receiving 1 MAC could be green (light) changing to yellow (medium) when receiving 1.25 MAC; alternatively another patient's basic depth of anaesthesia could be magenta (deep) with 1 MAC and red (very deep) when changing to 1.25 MAC. This shows that the EEG is a more sensitive test than the clinical evaluation, during medium anaesthesia. Comparing clinical evaluation with the EEG analysis, it was obvious that, in some instances, the anaesthetist adjusted the isoflurane concentration, without knowing the result of signal processing. In these cases the EEG analysis "agreed" in the treatment of the patient. The computer program was not designed to detect burst suppression. A visual inspection of the EEG strip chart in periods with isoflurane concentrations exceeding 1.5 MAC showed no burst suppression. In all patients slight hyperventilation was induced according to the procedure. This caused no change in the class profile in any of the patients. As expected in a non-critical surgical procedure, the asymmetry parameter was unchanged in all patients. DISCUSSION
Intraoperative monitoring of the EEG activity has been used mainly during anaesthetic procedures in which cerebral perfusion and metabolism are at risk. Several types of monitoring system have shown to be of help as warning systems [18]. In contrast, the use of EEG monitoring as a guide to depth of anaesthesia is controversial and un-
FIG. 3. Black and white version of a hard copy record from a patient monitored during 2 h 45 min of isoflurane anaesthesia. Vertical axis is time. Data from the EEG are displayed in the three columns to the left, the traditional physiological parameters in the two columns to the right. Events during anaesthesia are marked along the time axis at the left: T = thiopentone; x, X = start and stop of anaesthesia; o = start of surgery; •(•, — = start and stop of hyperventilation; >, < = increase and decrease in isoflurane concentration. The left column shows a DSA plot for one of the EEG recordings (horizontal axis is frequency in Hz). The second column shows symmetry between the two hemispheres (red on colour record; left of added white line here) and spectral similarity between the actual EEG sequence and the most likely class in the learning set (green on colour record; right of added white line here). The column in the middle shows the class histogram of the classification of the EEG activity. On the original colour record, the classes are colour-coded according to the depth of anaesthesia as defined in figure 2; here, white areas correspond to yellow on the colour record, and labels A-E have been added to identify areas of dark blue, light blue plus green (A, E) and of magenta plus red (B-D). The fourth
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
from 1.5 to 1.0 vol% (second <) and again from 1.0 to 0.5 vol% (third <). Induction with thiopentone (T) caused a short presence of very low frequency, high amplitude activity, hardly noticeable in the hard copy because of low time resolution. Isoflurane was administered at x. The depth of anaesthesia increased gradually, and after 1 h the class probability histogram (central record) was dominated by the yellow classes (white areas in figure 3) as assigned to medium anaesthesia. At > , the isoflurane vaporizer setting was increased by 50%. This change was immediately reflected in the class histogram, which was now dominated by the magenta-plus-red classes (black in figure 3: area B) assigned to deep levels of anaesthesia. When maintenance concentration was re-established (<), the yellow class reappeared. The response time for EEG changes to these steps in isoflurane concentrations was approximately 10 min, which is fast compared with the normal interval between manual notes in the anaesthesia log. The shift to a red period (black in figure 3; areas C, D) after approximately 2 h of recording may indicate that the patient reached deeper anaesthesia when the surgical procedure was less painful. The effect of the decreases in isoflurane concentration at the second and third < is clearly reflected by the reappearance of the yellow dominance (white in figure 3) in the class histogram. Towards the end of the recording, the histogram shifted to blue and green classes (black area E in figure 3), indicating transition from medium to light, very light and drowsy state, similar to the shift noted at start of surgery (event o in figure 3; black area A). In the majority of patients the yellow, medium depth dominated, but the interpatient variability
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA
41
TIME • DRTE • O i STHRTl r»T. I D I F»T _ 6 , X S O F L U R A H E
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
column shows end-tidal carbon dioxide concentration (CO2/black on colour record: 0-10°,,) and inspiratory oxygen concentration (O2/red on colour record; more continuous line here: 0-50",,). The column to the right shows heart rate (HR/green on colour record, 0-150 beat min~'; the width of this line represents the beat to beat variation in heart rate) and mean arterial pressure (MAP/blue on colour record; thicker, continuous vertical lines here: 0-150 mm Hg). The screen display during on-line analysis includes the last 1 h of the recording. The traces at the top are the raw signals of EEG and ECG displayed in 2-s segments for quality control during recording. The frame to the right is patient identification. The numbers below the columns are instantaneous displays of the variables.
42
bral hypoxia. In the present work we found no information in these features related to depth of anaesthesia and we conclude that the sensitivity is too low inside the therapeutic range of isoflurane concentration. This indicates the necessity of implementing specific common sets of basic patterns corresponding to the actual combination of anaesthetics used. In our view, the generation of this reference material will be a major task in the construction of monitoring systems in the future. ACKNOWLEDGEMENT This study was supported by the Danish Technical Research Council and the Danish Medical Research Council. REFERENCES 1. Prior PF. EEG monitoring and evoked potentials in brain ischaemia. British Journal of Anaesthesia 1985; 57: 63-81. 2. Pronk RAF, Simons AJR. Automatic recognition of abnormal EEG activity during open heart surgery. In: Buser PA, Cobb WA, Okuma T, eds. Kyoto Symposia, EEG Suppl 36. Amsterdam: Elsevier Biomedical Press, 1982; 590-602. 3. Chiappa KH, Burke SR, Young RR. Results and electroencephalo-graphic monitoring during 367 carotid endarterectomies. Stroke 1979; 10: 381-388. 4. Blackshear WM, Vincenzo DC, Seifert KB, Connar RG. Advantages of continuous electroencephalographic monitoring during carotid artery surgery. Journal of Cardiovascular Surgery 1986; 27: 146-153. 5. Trojaborg W, Boysen G. Relation between EEG, regional cerebral blood flow and internal carotid artery pressure during carotid endarterectomy. Electroencephalography and Clinical Neurophysiology 1973; 34: 61-69. 6. Klein FF, Davis DA. The use of time domain analyzed EEG in conjunction with cardiovascular parameters for monitoring anesthetic levels. IEEE Transactions on Biomedical Engineering 1981; 28: 36-40. 7. Pronk RAF, Simons AJR, Ackerstaff RGA, Boezeman EHJF. Intra-operative EEG monitoring. Proceedings of the Ninth Annual Conference of the IEEE Engineering in Medicine and Biology Society 1987; 3: 1250-1251. 8. Schwilden H, Stoeckel H. Quantitative EEG analysis during anaesthesia with isoflurane in nitrous oxide at 1.3 and 1.5 MAC. British Journal of Anaesthesia 1987; 59: 738-745. 9. Pronk RAF. EEG Processing in Cardiac Surgery (Thesis). Amsterdam: Free University, 1982. 10. Bickford RG, Flemming NI, Billinger TW. Compression of EEG data by isometric power spectral plots. Electroencephalography and Clinical Neurophysiology 1971; 31: 632. 11. Bickford RG, Brimm J, Berger L, Aung M. In: Application of compressed spectral array in clinical EEG. Kellaway B, Petersen I, eds. Automation of Clinical Electroencephalography. New York: Raven Press, 1973; 55-64. 12. Flemming RA, Smith NT. Density modulation: A technique for the display of three-variable data in patient monitoring. Aneslhesiology 1979; 50: 543-546.
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015
reliable [13]. The reason for this could be that most investigations have focused on recordings in which critical physiological variables show a considerable degree of instability. In the present work, our intention was to investigate the changes in brain activity related to the depth of anaesthesia in healthy subjects. Major efforts were made to keep basic physiological variables constant during monitoring. Many investigators have reported disappointing results, with no correlation between EEG frequency, amplitude and depth of anaesthesia. In particular, the changes have been shown to be specific for the anaesthetic agent used [7]. Nevertheless, many attempts have been made to visualize the EEG signal in a compressed and simplified form, such as CSA-plot or DSA-plot [10,11], in order to give the anaesthetist the opportunity to compare the clinical evaluation of anaesthetic depth with an on-line signal interpretation. Using computerized signal processing, a great number of variables and features can be extracted, which represent information undetectable by simple visual inspection of the EEG signal. The present investigation of the EEG involves autoregressive modelling and clustering analysis in order to define a set of basic classes of activity, representing different depths of anaesthesia. The set of basic classes was used as learning set for classification of EEG activity. The presentation of the result of classification in a class probability histogram demonstrated a high degree of simplicity. At the different depths of anaesthesia more than two classes were seldom represented. The learning set was common and derived specifically from the isoflurane series. By this technique, the class histogram shows a significant profile of class occurrence related to the depth of anaesthesia. The difference seen in some patients between the profile of the class histogram and the clinical level of anaesthesia reflects the fact that the features displayed are more sensitive than the clinical evaluation. Even though some patients had a different class profile during the same clinical level of anaesthesia, a change to lighter or deeper clinical levels caused a stereotyped shift in class histogram to profiles representing the same trend. In previous investigations it has been found that spectral edge frequency, centre frequency and mean energy content in the conventional frequency bands are valuable indicators of cere-
BRITISH JOURNAL OF ANAESTHESIA
COMPUTERIZED MONITORING OF DEPTH OF ANAESTHESIA
43
13. Editorial. Lancet 1986; 1: 553-554. mation of short segment spectra for computerized EEG 14. Makhoul J. Linear prediction: A tutorial review. Proanalysis. IEEE Transactions on Biomedical Engineering ceedings IEEE 1975; 63: 561-580. 1981; 28: 630-638. 15. Jansen BH. Analysis of biomedical signals by means of 17. Duda RO, Hart PE. Pattern Classification and Scene linear modelling. CRC Critical Reviews in Biomedical Analysis. New York: John Wiley & Sons Ltd, 1973. Engineering 1985; 12: 343-392. 18. Sebel P. Monitoring the central nervous system. Current 16. Jansen BH, Bourne JR, Ward JW. Autoregressive estiMedical Literature 1987; 1: 33-37.
Downloaded from http://bja.oxfordjournals.org/ at University of Michigan on July 22, 2015