Free
Clinical Science  |   August 2004
Signal Verification of Middle Latency Auditory Evoked Potentials by Automated Detection of the Brainstem Response
Author Affiliations & Notes
  • Gerhard Schneider, M.D.
    *
  • Eberhard F. Kochs, M.D.
  • Henry Arenbeck
  • Michael Gallinat
  • Gudrun Stockmanns, Ph.D.
    §
  • *Assistant Professor, †Professor, Director, and Chair, Department of Anesthesiology, Klinikum rechts der Isar, Technische Universität München; ‡Research Fellow, §Assistant professor, Institute of Information Technology, Faculty of Engineering Sciences, University Duisburg-Essen.
Article Information
Clinical Science
Clinical Science   |   August 2004
Signal Verification of Middle Latency Auditory Evoked Potentials by Automated Detection of the Brainstem Response
Anesthesiology 8 2004, Vol.101, 321-326. doi:
Anesthesiology 8 2004, Vol.101, 321-326. doi:
SEVERAL studies suggest that auditory evoked potentials (AEPs) can be used as a monitor of anesthetic depth. Increasing depth of anesthesia is reflected by increased latencies and decreased amplitudes of the midlatency peaks of the AEP (MLAEP).1–4 However, deep anesthesia may not be the only reason for depression or loss of the MLAEP. Identical depression of MLAEP occurs when headphones are disconnected or broken, i.e.  , when the auditory stimulus is not applied. Furthermore, invalid AEPs with suppression of MLAEPs are generated with undiagnosed hearing loss, i.e.  , when the auditory trigger is not transmitted via  the auditory pathway. Under these circumstances, MLAEP suppression may be falsely interpreted as deep anesthesia even when the patient is conscious. On the other hand, waveforms may be generated which resemble MLAEP but are not based on stimulus related responses. Thus, before MLAEP analysis is performed, it is necessary to verify that the analyzed signal is a valid AEP. It is a shortcoming of most MLAEP studies in anesthesia that they give no clue whether the analyzed signal has been validated as a true MLAEP. For this purpose, the brainstem AEP (BAEP) can be used, as it reflects the function of the auditory pathway and is almost unchanged by sedation and anesthesia. In the current study, an automated procedure was developed that is based on wavelet transform of the AEP and verifies that an analyzed signal is evoked by auditory stimuli (“valid AEP”). It can easily be integrated into an online MLAEP monitor.
Materials and Methods
We analyzed AEP signals from a database that contained perioperatively recorded electroencephalogram and AEP data.5 Data were from a study that was approved from the Human Investigations Committee of the University (Technische Universität) Munich, Germany. Selected signals were from the time interval between induction of anesthesia with remifentanil and extubation at the end of anesthesia. Electroencephalogram was recorded with 1 kHz sampling rate from Fpz-A2 with a bandpass of 0.5–500 Hz. Binaural rarefaction clicks were applied at 70 dB above hearing threshold using insert earphones (AW 180; Oticon, Strandvejen, DK). Stimulus frequency was 8.3291 Hz with a 10% variation of the interstimulus interval. The exact position of each click was marked in the electroencephalogram data file, allowing offline averaging of AEPs. Based on the 1 kHz sampling rate, electroencephalogram was digitally filtered with a 400 Hz low-pass and a 25 Hz high-pass filter. Sampling rate and filter settings are not only sufficient to analyze MLAEPs but also to extract peak V of the BAEP. Identification of peak V is routinely used in the AEP laboratory or during visual AEP monitoring and is the target of the current analysis.
For the current study, a subset of 40 patients was selected receiving general anesthesia with either propofol/remifentanil (n = 20) or sevoflurane/remifentanil (n = 20). The patient age was between 18 and 68 yr (median, 44 yr), height was between 158 and 191 cm (median, 172 cm), and weight was between 45 and 98 kg (median, 77 kg). Patients with hearing deficits were not included in the study. One hundred and ninety AEPs were randomly chosen. For this purpose, the beginning of the AEP was randomly selected in the electroencephalogram files. Starting from these time points, AEPs of 120 ms duration were averaged from 500 single sweeps using normal ensemble averaging. At an average, 4.75 BAEPs of each patient were used (minimum, 1; maximum, 7). Ninety-three AEPs were from patients receiving sevoflurane/remifentanil, and 97 AEPs were from patients receiving propofol/remifentanil anesthesia. In addition to AEP extraction, a nontrigger synchronized averaging procedure was performed on the same data subset to produce 190 “non-AEP” signals. “Non-AEP” signals were generated by the same algorithm used to generate AEP signals, but the trigger information was altered. Time values Δtiwere randomly chosen out of an interval from −30 to + 30 ms and added to the original trigger information ttrg,I. This produces a variable shift of the original trigger information, i.e.  , an artificial trigger information that is not in a constant relation to the auditory stimulus (fig. 1). Averaging of sweeps synchronized to the artificial trigger information produces “non-AEP” signals from the nearly identical electroencephalogram intervals. The 190 AEP and 190 “non-AEP” signals were used to develop a system of brainstem signal detection. First, the averaged waveforms were transformed by discrete wavelet decomposition up to level 6 using the biorthogonal 2.2 mother-wavelet (Matlab V5.3.0.10183, Wavelet Toolbox V1.2; Mathworks, Natick, MA) (fig. 2). Based on the structure and appearance of peak V of the brainstem response, a wavelet coefficient was selected that bears the relevant information. The retransform of a specific coefficient calculates the according signal component. This component is located in a specific time window and a specific frequency band. In AEP signals, retransform of the selected coefficient produced a signal showing a peak that resembled peak V of the brainstem response (fig. 3). The area under the receiver operating characteristic curve6 was calculated to assess the ability of the selected coefficient to separate between AEP and “non-AEP” signals. Results of this analysis are in the interval from 0–1. “0” and “1” indicate a complete separation, a value of 0.5 is the result that would be obtained by chance (e.g.  , flipping a coin). A threshold value was determined to separate AEP from “non-AEP” signals. Selection of this value was based on receiver operating characteristic curve analysis; sensitivity and specificity was calculated for each possible threshold value. The threshold value that produced the greatest sum of sensitivity and specificity for BAEP detection was selected.
Fig. 1. Generation of “not auditory evoked potential” (“non-AEP”) signals; 190 “non-AEP” signals were averaged from the identical electroencephalogram segment as the according AEP signals. For this purpose, the original trigger information in the electroencephalogram data stream (ttrg i) was shifted within the time interval [−30ms; 30ms] (t’trg i= ttrg i+Δti). Based on these new trigger marks (t’trg i), a “non-AEP” was averaged from 500 sweeps. 
Fig. 1. Generation of “not auditory evoked potential” (“non-AEP”) signals; 190 “non-AEP” signals were averaged from the identical electroencephalogram segment as the according AEP signals. For this purpose, the original trigger information in the electroencephalogram data stream (ttrg i) was shifted within the time interval [−30ms; 30ms] (t’trg i= ttrg i+Δti). Based on these new trigger marks (t’trg i), a “non-AEP” was averaged from 500 sweeps. 
Fig. 1. Generation of “not auditory evoked potential” (“non-AEP”) signals; 190 “non-AEP” signals were averaged from the identical electroencephalogram segment as the according AEP signals. For this purpose, the original trigger information in the electroencephalogram data stream (ttrg i) was shifted within the time interval [−30ms; 30ms] (t’trg i= ttrg i+Δti). Based on these new trigger marks (t’trg i), a “non-AEP” was averaged from 500 sweeps. 
×
Fig. 2. The figure shows wavelet decomposition of a characteristic auditory evoked potential (  left  ) and the separate retransform of detail levels 1 to 5 (  right  ). The retransform of detail level 3 shows a signal that resembles peak V of the brainstem response. 
Fig. 2. The figure shows wavelet decomposition of a characteristic auditory evoked potential (  left  ) and the separate retransform of detail levels 1 to 5 (  right  ). The retransform of detail level 3 shows a signal that resembles peak V of the brainstem response. 
Fig. 2. The figure shows wavelet decomposition of a characteristic auditory evoked potential (  left  ) and the separate retransform of detail levels 1 to 5 (  right  ). The retransform of detail level 3 shows a signal that resembles peak V of the brainstem response. 
×
Fig. 3. Auditory evoked potentials (AEP signals) (  dotted lines  ) and retransformed signals based on the third coefficient of the third level of wavelet transform (with the biorthogonal 2.2 mother wavelet,  solid lines  ). Part A shows the average of all 190 AEP signals (  upper part  ) and an AEP signal from a representative subject (  lower part  ) with the corresponding retransformed signals, which resemble peak V of the brainstem response. Part B shows the average of the corresponding 190 “non-AEP” signals (  upper part  ) and a typical “non-AEP” signal (  lower part  ). In the “non-AEPs” and the corresponding retransformed signals, peak V of the brainstem response is absent. 
Fig. 3. Auditory evoked potentials (AEP signals) (  dotted lines  ) and retransformed signals based on the third coefficient of the third level of wavelet transform (with the biorthogonal 2.2 mother wavelet,  solid lines  ). Part A shows the average of all 190 AEP signals (  upper part  ) and an AEP signal from a representative subject (  lower part  ) with the corresponding retransformed signals, which resemble peak V of the brainstem response. Part B shows the average of the corresponding 190 “non-AEP” signals (  upper part  ) and a typical “non-AEP” signal (  lower part  ). In the “non-AEPs” and the corresponding retransformed signals, peak V of the brainstem response is absent. 
Fig. 3. Auditory evoked potentials (AEP signals) (  dotted lines  ) and retransformed signals based on the third coefficient of the third level of wavelet transform (with the biorthogonal 2.2 mother wavelet,  solid lines  ). Part A shows the average of all 190 AEP signals (  upper part  ) and an AEP signal from a representative subject (  lower part  ) with the corresponding retransformed signals, which resemble peak V of the brainstem response. Part B shows the average of the corresponding 190 “non-AEP” signals (  upper part  ) and a typical “non-AEP” signal (  lower part  ). In the “non-AEPs” and the corresponding retransformed signals, peak V of the brainstem response is absent. 
×
To test the new detector in intermediate situations, two basic tests were performed in three additional patients: first, with ongoing AEP stimulation earplugs were inserted and combined with AEP earphones in the neck (not on the ears) to examine whether stimulus artifact imitates a peak V. Next, stimulus intensity was stepwise reduced (80, 60, 40, 20, and 0% stimulus intensity) to simulate the influence of partly dislodged earphones or partial hearing loss.
Results
The third coefficient of the third level was selected as a detector of the brainstem response. Retransform of this coefficient produces a time frequency component that shows a peak that resembles peak V of the brainstem response. Figure 3Ashows AEP signals and the retransformed parts of the signals that are based on the particular wavelet coefficient. The grand average of AEP signals (upper part) and a characteristic example for an AEP signal (lower part) are shown. Figure 3Bshows the corresponding “non-AEP” signals and the retransformed parts of the signals that are based on the particular wavelet-coefficient. The grand average of “nonclick-related” signals and a characteristic “non-AEP” are shown, whereas a corresponding peak V of the brainstem response is absent. Receiver operating characteristic curve analysis of the selected coefficient resulted in a receiver operating characteristic curve area of 0.008698. The optimum threshold value was −0.4283. This threshold allowed a correct detection of the brainstem response in 97.90% of the AEP signals (sensitivity). Because of the absence of the brainstem response, 99.48% of the “nonclick-related” signals were correctly identified as “non-AEPs” (specificity, fig. 4). The first basic test of the detector showed a correct identification as “non-AEPs” for all signals that were recorded with earphones on the neck and earplugs in the ears. In the second test, the percentage of detected brainstem components showed a stepwise decrease from 70 dB (AEPs with stimulus intensity of 100%) to 0 dB (AEPs with stimulus intensity of 0%, i.e.  , without auditory stimulus). Detailed results of these tests are given in Table 1.
Fig. 4. The figure illustrates the distribution of the values of the particular wavelet coefficient (bior2.2_d3_3). The threshold value (  solid line  ) was selected to reach the maximum of (sensitivity + specificity) for detection of peak V of the brainstem response. 
Fig. 4. The figure illustrates the distribution of the values of the particular wavelet coefficient (bior2.2_d3_3). The threshold value (  solid line  ) was selected to reach the maximum of (sensitivity + specificity) for detection of peak V of the brainstem response. 
Fig. 4. The figure illustrates the distribution of the values of the particular wavelet coefficient (bior2.2_d3_3). The threshold value (  solid line  ) was selected to reach the maximum of (sensitivity + specificity) for detection of peak V of the brainstem response. 
×
Table 1. Test Results 
Image not available
Table 1. Test Results 
×
Discussion
The presented detector of the brainstem response has high sensitivity and specificity and can be used to verify that the signal is evoked by auditory stimuli, i.e.  , that it is an AEP signal rather than an artifact. We favor the use of the brainstem response, as it is almost unchanged by anesthetics. It demonstrates that the acoustic signal is transmitted via  the auditory pathway through the brainstem. The presence of peak V of the brainstem response excludes malfunction of the acoustic stimulator and dislocation or disconnection of earphones as a reason for depressed MLAEP. In addition, it confirms the functional integrity of the auditory pathway and detects patients with hearing loss. This is also shown by the results of the first tests. Signals that were recorded with earplugs in the ears and headphones in the neck were classified as “non-AEPs.” This illustrates that there is no stimulus artifact that mimics peak V of the brainstem response. Different stimulus levels were applied to simulate clinical situations with partial hearing loss (e.g.  , from middle ear pressure changes secondary to nitrous oxide or Eustachian tube blockade) or stimulus earphones that are partly dislodged. These tests showed that reduction of the stimulus intensity decreases probability for detection of peak V of the brainstem. For situations with reduced stimulus intensity, further studies are required that should also examine the relationship between BAEP and MLAEP during reduced stimulus intensity.
An alternative approach for AEP signal verification is applied in the commercially available Danmeter AAI AEP monitor. This method is based on an analysis of the signal-to-noise ratio of the MLAEP.7 In contrast to the brainstem response, MLAEPs are strongly depressed by anesthesia. This leads to the alarm “low AEP/no AEP,” suggesting disconnection of headphones or technical problems, whenever anesthesia is deep. This disadvantage has been overcome by our detector based on peak V of the brainstem response, which remains almost unchanged by anesthetics. We used wavelet transform to decompose the AEP and identify the brainstem response. For AEP analysis, wavelet transform has several advantages8 as compared with Fourier transform. Fourier transform eliminates the time information and may deteriorate the frequency resolution of transient signals by application of a fixed time window. The principles of wavelet transform are similar to Fourier transform, but instead of sine waves a mother wavelet, i.e.  , a signal of finite length, is used as the fundamental waveform for the decomposition. To calculate the signal components, different (dilated and shifted) versions of the mother wavelet are used as basis functions for the transform. Dilation of the mother wavelet (broadening or narrowing of the wavelet along the time axis) is used to extract information about the underlying frequencies. Shifting of the mother wavelet along the time axis is used to extract time information. Thus, wavelet analysis characterizes a signal in a time frequency domain. In the current study wavelet transform was performed up to level 6. This means that six differently dilated versions of the mother wavelet were used. The results of the wavelet transform are several coefficients that represent different signal components. These are related to the applied versions (dilated and shifted) of the mother wavelet. Specific characteristics (i.e.  , waveform, structural details, and frequency contents related to time) can be evaluated by this set of coefficients. In the current study we chose the biorthogonal 2.2 mother-wavelet, as it allowed a good approximation of the signal characteristics of peak V of the BAEP (i.e.  , time-frequency component). The choice of the biorthogonal 2.2 mother wavelet was based on experience. The aim was to obtain a minimal number of coefficients that were able to represent peak V of the BAEP. This was achieved with the biorthogonal 2.2 mother wavelet. The use of only one coefficient (d3_3) of the wavelet transform allows an adequate representation of the main characteristics of peak V of the brainstem component. The third coefficient of the third level of wavelet decomposition, combined with a simple threshold function, allowed an automated detection of the brainstem response. The detector does not require in-depth experience with neurophysiological signals and may be used as an automated method of AEP signal verification that can easily be integrated into a monitor. This guarantees that such a monitor detects “non-AEPs” that may be attributable to hearing loss or technical problems (failure of the acoustic stimulator or disconnection of earphones) and can exclude those signals from further analysis. In the current study, the method has been developed for AEPs from 500 sweeps. As the number of sweeps as well as the sample rate and filter characteristics of the electroencephalogram amplifier may influence the shape of peak V of the BAEP, the developed detector must be thoroughly tested before it is applied to different settings.
The authors thank Georg Schäpers, M.Sc., Research Fellow, Department of Anesthesiology, Technische Universität München, Munich, Germany, for his help.
References
Thornton C, Heneghan CP, James MF, Jones JG: Effects of halothane or enflurane with controlled ventilation on auditory evoked potentials. Br J Anaesth 1984; 56:315–23Thornton, C Heneghan, CP James, MF Jones, JG
Schwender D, Kaiser A, Klasing S, Peter K, Pöppel E: Midlatency auditory evoked potentials and explicit and implicit memory in patients undergoing cardiac surgery. Anesthesiology 1994; 80:493–501Schwender, D Kaiser, A Klasing, S Peter, K Pöppel, E
Thornton C, Sharpe RM: Evoked responses in anaesthesia. Br J Anaesth 1998; 81:771–81Thornton, C Sharpe, RM
Drummond JC: Monitoring depth of anesthesia: With emphasis on the application of the bispectral index and the middle latency auditory evoked response to the prevention of recall. Anesthesiology 2000; 93:876–82Drummond, JC
Ningler M, Schneider G, Stockmanns G, Schäpers G, Kochs E: Databank for support of comprehensive study evaluations of signals for anesthesia monitoring. Biomed Tech (Berl) 2002; 47 S1:550–3Ningler, M Schneider, G Stockmanns, G Schäpers, G Kochs, E
Metz CE: Basis principle of ROC analysis. Semin Nucl Med 1978; 8:283–98Metz, CE
Martinez P, Trillo L, Jensen E, Rodriguez B, Litvan H: Validation of an algorithm for auditory evoked potential detection quality. Eur J Anaesthesiol 2002; 19 S24:A-94, 24Martinez, P Trillo, L Jensen, E Rodriguez, B Litvan, H
Kochs E, Stockmanns G, Thornton C, Nahm W, Kalkman CJ: Wavelet analysis of middle latency auditory evoked responses: calculation of an index for detection of awareness during propofol administration. Anesthesiology 2001; 95:1141–50Kochs, E Stockmanns, G Thornton, C Nahm, W Kalkman, CJ
Fig. 1. Generation of “not auditory evoked potential” (“non-AEP”) signals; 190 “non-AEP” signals were averaged from the identical electroencephalogram segment as the according AEP signals. For this purpose, the original trigger information in the electroencephalogram data stream (ttrg i) was shifted within the time interval [−30ms; 30ms] (t’trg i= ttrg i+Δti). Based on these new trigger marks (t’trg i), a “non-AEP” was averaged from 500 sweeps. 
Fig. 1. Generation of “not auditory evoked potential” (“non-AEP”) signals; 190 “non-AEP” signals were averaged from the identical electroencephalogram segment as the according AEP signals. For this purpose, the original trigger information in the electroencephalogram data stream (ttrg i) was shifted within the time interval [−30ms; 30ms] (t’trg i= ttrg i+Δti). Based on these new trigger marks (t’trg i), a “non-AEP” was averaged from 500 sweeps. 
Fig. 1. Generation of “not auditory evoked potential” (“non-AEP”) signals; 190 “non-AEP” signals were averaged from the identical electroencephalogram segment as the according AEP signals. For this purpose, the original trigger information in the electroencephalogram data stream (ttrg i) was shifted within the time interval [−30ms; 30ms] (t’trg i= ttrg i+Δti). Based on these new trigger marks (t’trg i), a “non-AEP” was averaged from 500 sweeps. 
×
Fig. 2. The figure shows wavelet decomposition of a characteristic auditory evoked potential (  left  ) and the separate retransform of detail levels 1 to 5 (  right  ). The retransform of detail level 3 shows a signal that resembles peak V of the brainstem response. 
Fig. 2. The figure shows wavelet decomposition of a characteristic auditory evoked potential (  left  ) and the separate retransform of detail levels 1 to 5 (  right  ). The retransform of detail level 3 shows a signal that resembles peak V of the brainstem response. 
Fig. 2. The figure shows wavelet decomposition of a characteristic auditory evoked potential (  left  ) and the separate retransform of detail levels 1 to 5 (  right  ). The retransform of detail level 3 shows a signal that resembles peak V of the brainstem response. 
×
Fig. 3. Auditory evoked potentials (AEP signals) (  dotted lines  ) and retransformed signals based on the third coefficient of the third level of wavelet transform (with the biorthogonal 2.2 mother wavelet,  solid lines  ). Part A shows the average of all 190 AEP signals (  upper part  ) and an AEP signal from a representative subject (  lower part  ) with the corresponding retransformed signals, which resemble peak V of the brainstem response. Part B shows the average of the corresponding 190 “non-AEP” signals (  upper part  ) and a typical “non-AEP” signal (  lower part  ). In the “non-AEPs” and the corresponding retransformed signals, peak V of the brainstem response is absent. 
Fig. 3. Auditory evoked potentials (AEP signals) (  dotted lines  ) and retransformed signals based on the third coefficient of the third level of wavelet transform (with the biorthogonal 2.2 mother wavelet,  solid lines  ). Part A shows the average of all 190 AEP signals (  upper part  ) and an AEP signal from a representative subject (  lower part  ) with the corresponding retransformed signals, which resemble peak V of the brainstem response. Part B shows the average of the corresponding 190 “non-AEP” signals (  upper part  ) and a typical “non-AEP” signal (  lower part  ). In the “non-AEPs” and the corresponding retransformed signals, peak V of the brainstem response is absent. 
Fig. 3. Auditory evoked potentials (AEP signals) (  dotted lines  ) and retransformed signals based on the third coefficient of the third level of wavelet transform (with the biorthogonal 2.2 mother wavelet,  solid lines  ). Part A shows the average of all 190 AEP signals (  upper part  ) and an AEP signal from a representative subject (  lower part  ) with the corresponding retransformed signals, which resemble peak V of the brainstem response. Part B shows the average of the corresponding 190 “non-AEP” signals (  upper part  ) and a typical “non-AEP” signal (  lower part  ). In the “non-AEPs” and the corresponding retransformed signals, peak V of the brainstem response is absent. 
×
Fig. 4. The figure illustrates the distribution of the values of the particular wavelet coefficient (bior2.2_d3_3). The threshold value (  solid line  ) was selected to reach the maximum of (sensitivity + specificity) for detection of peak V of the brainstem response. 
Fig. 4. The figure illustrates the distribution of the values of the particular wavelet coefficient (bior2.2_d3_3). The threshold value (  solid line  ) was selected to reach the maximum of (sensitivity + specificity) for detection of peak V of the brainstem response. 
Fig. 4. The figure illustrates the distribution of the values of the particular wavelet coefficient (bior2.2_d3_3). The threshold value (  solid line  ) was selected to reach the maximum of (sensitivity + specificity) for detection of peak V of the brainstem response. 
×
Table 1. Test Results 
Image not available
Table 1. Test Results 
×