Sorted averaging improves quality of auditory steady-state responses

Journal of Neuroscience Methods 216 (2013) 28–32

Contents lists available at SciVerse ScienceDirect

Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth

Clinical Neuroscience

Sorted averaging improves quality of auditory steady-state responses Torsten Rahne a,∗ , Jesko L. Verhey b , Roland Mühler b a b

Department of Otorhinolaryngology and Halle Hearing and Implant Center, Martin-Luther-University Halle-Wittenberg, Halle (Saale), Germany Department of Experimental Audiology, Otto von Guericke University Magdeburg, Leipziger Str. 44, 39120 Magdeburg, Germany

h i g h l i g h t s • Increasing signal-to-noise ratio (SNR) is essential for the recording of auditory evoked potentials. • Sorted-averaging protocol averages EEG epochs after sorting the epochs according to their estimated root-mean-square (RMS) amplitude until maximum SNR.

• Sorted averaging is, for the first time, applied to auditory steady-state responses (ASSR). • Higher SNR of the average was achieved by sorted sorted-averaging protocol as compared to adaptive-rejection-level and weighted-averaging protocols. • Sorted averaging may be a powerful tool to increase the quality of ASSR.

a r t i c l e

i n f o

Article history: Received 19 February 2013 Received in revised form 9 April 2013 Accepted 10 April 2013 Keywords: Sorted averaging Artifact rejection Weighted averaging Signal-to-noise ratio (SNR) Auditory state-state responses (ASSR) Auditory brainstem responses (ABR)

a b s t r a c t Increasing signal-to-noise ratio (SNR) is essential for the recording of auditory evoked potentials with electroencephalography (EEG). Several protocols have been proposed to increase the SNR, starting with an averaging of EEG epochs which decreases noise level. Since artifacts decrease the SNR by increasing the noise level, artifact detection and reduction protocols are other important tools to reduce the noise level. The current study focuses on the sorted averaging protocol where the epochs are sort according to their estimated root-mean-square (RMS) amplitude. Calculating an estimated SNR by averaging of the sorted epochs, this process of averaging can be interrupted at the maximum SNR, i.e., at an optimal number of epochs. In contrast to the often used protocol that weighs every epoch by its inverse average rootmean-square amplitude, sorted averaging is a linear operation, i.e., it does not change signal amplitudes. In this study, the sorted averaging is, for the first time, applied to auditory steady-state responses (ASSR) which are evoked by amplitude modulated tones or trains of transient acoustic stimuli. In contrast to other evoked potentials, the ASSR is analyzed in the frequency domain, using the property of auditory system to retain the modulation frequency (or the repetition rate) of the stimulus. ASSR were recorded in 11 subjects with normal hearing. Results of four artifact processing protocols (1) fixed rejection level, (2) adaptive rejection level, (3) weighted averaging and (4) sorted averaging were compared. The results showed a higher normalized SNR with a sorted averaging protocol than with adaptive rejection level and weighted averaging protocols. An advantage of the sorted averaging protocol is that, compared to a fixed-rejection threshold, the ASSR amplitudes were unchanged when the sorted averaging protocol was used, whereas they were significantly reduced by the weighted averaging protocol. The residual noise was also significantly lower for the sorted averaging protocol than for the weighted averaging and adaptive rejection protocols. Thus, the sorted averaging may be a powerful tool to increase the quality of ASSR. © 2013 Elsevier B.V. All rights reserved.

1. Introduction Auditory steady-state responses (ASSR) are used in auditory neuroscience and clinical routine for an objective estimation of the hearing threshold. Here the response of the auditory system

∗ Corresponding author at: Universitäts-HNO-Klinik, Ernst-Grube-Str. 40, 06120 Halle (Saale), Germany. Tel.: +49 345 557 5362; fax: +49 345 557 1859. E-mail address: [email protected] (T. Rahne). 0165-0270/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jneumeth.2013.04.004

to amplitude modulated tones or trains of transient acoustic stimuli as clicks or chirps are measured using electroencephalography (EEG) or magnetoencephalography (MEG). The ASSR uses the characteristics of the auditory system at the level of the brainstem and the primary auditory cortex to retain the modulation frequency or the repetition rate of transient sound stimuli, resulting in a peak in the spectrum at the respective frequency (Picton et al., 2003). In general, the power of this peak is correlated with the stimulation sound pressure level (SPL), increasing from threshold to super-threshold intensity levels. Although the ASSR audiograms are

T. Rahne et al. / Journal of Neuroscience Methods 216 (2013) 28–32

strongly correlated with behaviorally measured audiograms, they cannot be directly linked to behavioral pure-tone thresholds, since the relation between ASSR and behavioral thresholds depends on the frequency content of the stimulus and the hearing loss (Ahn et al., 2007; Aoyagi et al., 1999; Cone-Wesson et al., 2002). In general, the magnitude of the ASSR is much smaller than the background bioelectrical activity, the latter commonly referred to as noise. A reliable determination of ASSR responses with stimulation SPL close to the threshold is thus problematic. To improve the signal-to-noise ratio (SNR), the response is measured for several epochs, i.e., by increasing the stimulus presentation and recording time. Assuming an invariance of the ASSR and a Gaussian distribution of the mean amplitudes of epoched EEG, the averaging over several epochs reduces the residual noise and improves the SNR in proportion to the square root of the number of averaged epochs. Thus, to detect small ASSR responses, long recording times are necessary (D’Haenens et al., 2008). A further possibility to improve SNR of the ASSR is to reduce the noise level of the recorded EEG epochs by using artifact reduction algorithms. Note that the assumption of a Gaussian distribution of mean amplitudes of epoched EEG is violated by occurrence of these artifacts which are mainly caused by muscle or eye movement. Therefore, in addition to averaging, an effective artifact detection (or correction) algorithm is needed to calculate reliable ASSR spectra. Artifact rejection with a fixed threshold for all subjects is not efficient enough since the background noise can vary significantly between subjects and occasionally even during a single recording session (Riedel et al., 2001). For auditory brainstem responses (ABR), several algorithms were developed to reduce background noise. For example, weighting every epoch by the inverse variance of the epoch reduces the influence of noisy epochs on the average waveform (Hoke et al., 1984; Lütkenhöner et al., 1985). This method has also been applied to ASSR, increasing the SNR significantly (John et al., 2001). In addition, an individual artifact threshold can be applied to further optimize artifact detection, taking into account the variability of the EEG amplitude across subjects (Riedel et al., 2001). The present study focuses on a sorted averaging algorithm as another possibility to increase the SNR. This sorted averaging algorithm was introduced by Mühler and von Specht (1999) for transient evoked potentials and allows for an averaging time that is optimized with respect to the SNR. Here, the noise was estimated for each epoch by the root-mean-square (RMS) amplitude of all recorded samples within an epoch. Sorted averaging relies on the assumption of interchangeability of epochs within the ensemble, and is applied by sorting all epochs according to their estimated RMS amplitude and successively averaging of this sorted ensemble. Starting with the epochs with a low noise level the SNR increases and after reaching a maximum, SNR starts to decrease because more and more noisy epochs are included in the ensemble average. This maximum is used to stop the process of averaging at an optimal number of epochs. In contrast to weighted averaging, sorted averaging is a linear operation which does not change signal amplitudes. The sorted averaging algorithm has already been successfully applied in auditory brainstem responses (ABR) (Mühler and von Specht, 1999) and cortical responses (Rahne et al., 2008), showing for both responses a significantly improved SNR. In this study, sorted averaging was applied to ASSR. In contrast to ABR, the ASSR are evaluated in the frequency domain. The hypothesis was that a sorted averaging protocol would increase the SNR by decreasing the noise level. An implementation of this averaging method in clinical routine as well as in research would thus improve the quality of ASSR and would allow to optimize the recording time.

29

2. Materials and methods 2.1. Subjects ASSR were recorded in 11 subjects with normal hearing. For the purposes of this study, normal hearing was defined as puretone thresholds of 20 dB HL or better for all audiometric frequencies between 500 and 4000 Hz. The age range of the subjects (6 female and 5 male) was 20–38 years. The protocol used in this study was in accordance with the Declaration of Helsinki. It was approved by the Ethics Revisory Board of the Otto-von-Guericke-University Magdeburg and all subjects provided written informed consent. 2.2. Stimuli Auditory steady-state responses were recorded with a multiple stimulus paradigm, as described by John et al. (1998), known as MASTER (multiple auditory steady-state responses). Stimuli were sinusoidal amplitude-modulated pure tones that were generated digitally and converted to analog signals by a sound card with 24 bit resolution (RME HDSP-9632, Haimhausen, Germany): the sounds were attenuated using a programmable attenuator and headphone buffer (g.PATH, g.tec, Graz, Austria). Air-conducted stimuli were presented to the right ear through a HDA280 headphone (Sennheiser, Wennebostel, Germany). Stimuli for ASSR for multiple carriers (multiple ASSR) consisted of sine waves at carrier frequencies of 0.5, 1, 2, and 4 kHz, 100% amplitude-modulated at frequencies of 80, 88, 104, and 112 Hz, respectively. The four stimulus components were calibrated with an ear simulator according to IEC 603168-1 (International Electrotechnical Commission (IEC), 2009), using the reference equivalent threshold sound pressure levels (RETSPL) delivered by Sennheiser. ASSR were recorded for stimulus levels between 20 and 60 dB HL. For the sake of simplicity, only the data for 1 kHz were further analyzed. It is reasonable to assume that the same result is obtained for the other signal frequencies. 2.3. EEG-recording and data processing EEG data were collected with a laboratory setup running on a personal computer. Subjects were seated on a comfortable couch in a sound-insulated and electrically shielded booth and were instructed to relax but not to fall asleep. Vigilance of the patients was visually monitored by the investigator. Ag/AgCl-electrodes were placed at the vertex (+) and contralateral mastoid (−) with a ground electrode at the forehead. Impedances were kept below 5 kOhms. The EEG activity was amplified 100,000 times with an analog g.Bsamp biosignal amplifier (g.tec, Graz, Austria), bandpass filtered from 60 Hz to 140 Hz (36 dB/octave), digitized with a 12 bit resolution at a sampling rate of 1024 Hz and stored on disk. For online visual inspection of the ASSR, subsequent EEG epochs (length: 1 s) were averaged in the time domain and were analyzed in the frequency domain by a fast Fourier transform (FFT). To ensure a constant quality of all measurements, recording stopped when the residual noise, determined by averaging the noise value in the frequency bins between 75 Hz and 120 Hz, was below 15 nV. To evaluate the influence of different artifact processing protocols, data were analyzed offline with an algorithm based on Matlab software (MathWorks, Natick, MA, USA). RMS amplitudes of EEG epochs of 1 s duration were calculated and four artifact processing protocols were applied: (1) a fixed rejection level, (2) an adaptive rejection level, (3) weighted averaging and (4) sorted averaging. For the ‘fixed level’ protocol, all epochs with RMS amplitudes exceeding 6 ␮V were excluded from averaging. For the ‘adaptive level’ protocol, the rejection level was derived from the mean RMS amplitude of the entire recording. All epochs with RMS amplitudes exceeding

30

T. Rahne et al. / Journal of Neuroscience Methods 216 (2013) 28–32

4 estimated SNR (a.u.)

A

B

3 2 1 0

single epoch rms (µV)

C

D

2.5 2 1.5 1 0.5 0

0

50

100 150 epoch index

200 0

50 100 150 epoch index sorted

200

Fig. 1. Principle of sorted averaging: the estimated SNR (panels A and B) and the single epoch RMS (panels C and D) of a typical ASSR-recording are plotted as a function of epoch index. SNR was estimated by calculating the inverse residual noise of the averaged epochs. Panel A shows the application of a fixed artifact level on an EEG time course with a steep increase of EEG amplitude (cf. panel C). Panel B shows the SNR for the same epoch ensemble as plotted in panel A, but now based on the epochs that were sorted by the single epoch RMS (cf. panel D). The arrow indicates the maximum SNR.

the mean plus one standard deviation were excluded from averaging. ‘Weighted averaging’ was applied by dividing all samples of an epoch by the variance of that epoch. The epochs were summed together and the summed epoch was divided by the sum of the

estimated SNR (a.u.)

weighted 130% sorted 142%

weighted 104% sorted 101%

4

weighting factors. ‘Sorted averaging’ was implemented by sorting all epochs of a recording according to their RMS amplitude. The sorted ensemble was successively averaged starting with low noise epochs until the maximum of the estimated SNR was reached.

weighted 99% sorted 156%

weighted 97% sorted 120%

3 2 1

single epoch rms (µV)

0

A

B

D

C

4 3 2 1 0

0

100 200 epoch index

300 0

100

200 300 400 epoch index

0

100 200 300 400 epoch index

0

100 200 300 epoch index

Fig. 2. Typical examples of the time course of the EEG RMS amplitudes and the SNR of four subjects. RMS amplitudes are plotted as a function of epoch index in the bottom row. The top row shows the corresponding SNR functions for the fixed artifact level (thin line), weighted averaging (dashed line) and sorted averaging (thick line). The normalized signal to-noise ratios for weighted and sorted averaging are shown in the top left corner of each panel. For the (noisy) examples shown in panels B–D, the sorted averaging protocol is superior to the weighted averaging protocol.

T. Rahne et al. / Journal of Neuroscience Methods 216 (2013) 28–32

3. Results Fig. 1 visualizes the principle and efficiency of sorted averaging for ASSR recordings. With an adaptive artifact threshold the estimated SNR increases in proportion to the square root of the number of averaged epochs (panel A). Epochs with high RMS amplitudes (see panel C) produce sudden drops of this function. After sorting the epochs by their RMS amplitude (see panel D), this SNR function shows a faster increase for the first few epochs than for the unsorted epochs and reaches a distinct maximum (panel B). Increasing the number of epochs further results in a decrease in SNR. Finally, for all recorded epochs, the estimated SNR is the same as for the unsorted average. The sorted averaging algorithm stops the averaging process at the maximum SNR. Fig. 2 shows the time course of RMS amplitude of an EEG response recorded in four subjects, revealing different artifact distributions. If only a few artifacts occur, as in panel A, all artifact processing protocols yield to similar SNR. However, if more artifacts occur (panels B–D), sorted averaging is superior to the other algorithms. An ANOVA revealed that artifact processing protocol affected SNR (F(2.1, 111.3) = 14.0, p < 0.001), noise level (F(1.3, 72.6) = 34.5, p < 0.001), and ASSR amplitude (F(1.9, 104.3) = 7.9, p < 0.01). Post hoc tests showed that the SNR with the weighted averaging protocol was larger than that obtained with the adaptive rejection threshold protocol which in turn was higher than as the SNR with the fixed rejection level protocol. Sorted averaging resulted in the largest SNR when compared to all other artifact rejection protocols. Residual noise level is highest for the fixed rejection level protocol, followed by the adaptive rejection level and the weighted averaging protocols. The lowest residual noise was obtained with the sorted averaging protocol. Response amplitudes were smaller for the weighted averaging protocol than for the other protocols (all Bonferroni corrected ps < 0.05). No interactive effect of artifact processing algorithm and stimulus level was found. Fig. 3 shows the residual noise, SNR, and response amplitudes of ASSR recordings normalized to the results obtained with the fixed artifact rejection level protocol. An ANOVA of normalized results revealed a main effect of artifact processing protocol (F(1.3, 69.4) = 4.4, p < 0.05), noise level (F(1.1, 58.6) = 20.6, p < 0.001), and ASSR amplitude (F(1.7, 90.6) = 5.8, p < 0.01) on SNR. Post hoc comparisons showed significant larger normalized SNRs for the

*

250

*

225 200 relative change (%)

The data processed by the four protocols were averaged in time domain and were transformed to frequency domain by a FFT. From the resulting ASSR spectra, the residual noise was determined by averaging the noise value in the frequency bins between 75 Hz and 120 Hz, excluding the frequency bins at the modulation frequencies (80, 88, 104 and 112 Hz) from calculation. To estimate the SNR of the recording, the amplitude of the 88-Hz frequency bin (i.e., the response for the 1-kHz carrier) was divided by the residual noise. As ASSR amplitudes and SNRs recorded for a wide range of stimulus levels differ considerably, all data were normalized by dividing them by the corresponding values for the ‘fixed level’ condition. Hereby the relative increase or decrease of residual noise, amplitude and SNR for the three remaining artifact processing protocols was calculated. Mean ASSR amplitudes, noise levels and estimated SNR were compared by a two-way analysis of variance for repeated measures (ANOVA) with factors of artifact processing protocol (fixed rejection level, adaptive rejection level, weighted and sorted) and stimulus level (10–60 dB). Greenhouse–Geisser correction for sphericity was applied. Post hoc tests were performed by t-tests for paired samples using the Bonferroni correction. Normalized mean ASSR amplitudes, residual noise levels and estimated SNR were compared by two-ways ANOVA for repeated measures testing the same factors.

31

175 150 125

*

* *

100

*

75 50 25 0

a

w

s

residual noise

a

w SNR

s

a

w

s

amplitude

Fig. 3. Residual noise, SNR and response amplitude (means and standard deviations) of ASSR recordings in normal hearing adults, calculated with an adaptive artifact rejection level (a), weighted averaging (w) and sorted averaging (s) protocol, respectively. Data are normalized relative to a fixed artifact rejection level (*Bonferroni corrected p < 0.05).

sorted averaging protocol than for the adaptive rejection level and weighted averaging protocols. Normalized amplitudes were significantly lower for the weighted averaging protocol than for the adaptive rejection level and sorted averaging protocol. The normalized noise was significantly lower for the weighted averaging than for the adaptive rejection level protocol and was even lower when the sorted averaging protocol was used (all Bonferroni corrected ps < 0.05). No interactive effect of artifact processing algorithm and stimulus level was found. 4. Discussion The results demonstrate the applicability of the sorted averaging protocol, originally developed for ABR measurements by Mühler and von Specht (1999), to ASSR recordings. They further show that this protocol is significantly better in the estimated SNR of the average spectrum than the fixed rejection level, the adaptive rejection level, and the weighted averaging protocols. As shown in Fig. 2, SNR and thus the quality of the recording vary between different averaging protocols, especially if the background noise is high. An adaptive artifact rejection level protocol rejects epochs with large RMS amplitudes, i.e., those with the episodic artifacts. Therefore, SNR as a measure of the quality of ASSR response was increased significantly compared to averaging with fixed artifact threshold. An additional problem with the fixed artifact threshold protocol is the choice of the optimal artifact rejection level which depends on the individual artifact distribution. Thus, it is crucial (but difficult) to determine the optimal artifact rejection level before recording (Don and Elberling, 1994). A significant increase of the SNR was achieved by applying the sorted averaging protocol (see Fig. 3). As in Rahne et al. (2008), the sorted averaging protocol of Mühler and von Specht (1999) was extended by using overall variance instead of the single-point variance to estimate SNR which is important for the identification of the maximum of the SNR function (Cebulla et al., 2000). Lütkenhöner et al. (1985) reported an underestimation of the signal amplitude by weighted averaging which is confirmed by the current results. However, both, the adaptive or fixed artifact rejection level protocols and the sorted averaging protocol are linear operations which do not change the ASSR amplitude. Since a smaller signal amplitude reduces the SNR, this is a major advantage compared to the weighted averaging protocol. Consequently,

32

T. Rahne et al. / Journal of Neuroscience Methods 216 (2013) 28–32

at least the adaptive artifact rejection level protocol should be used to achieve proper SNR and thus ASSR spectrums. Given the high computation power in the current laboratories, the sorted averaging algorithm is an optimal choice to further improve the SNR, even used online during the EEG recording. 5. Conclusions Sorted averaging protocol improves the process of removing artifacts significantly and therefore appears as a fundamental step on the approach to artifact-free ASSR spectrums and thus an optimal quality of ASSR recordings. Applied in clinical routine, the objective diagnostic of hearing loss could be further improved. In addition, it may serve as a tool to further reduce measuring time. John et al. (2001) considered this approach still as computationally very demanding if used online. Nowadays, the computational power should be sufficient, especially if appropriate sorting algorithms are chosen. References Ahn JH, Lee HS, Kim YJ, Yoon TH, Chung JW. Comparing pure-tone audiometry and auditory steady state response for the measurement of hearing loss. Otolaryngol Head Neck Surg 2007;136:966–71. Aoyagi M, Suzuki Y, Yokota M, Furuse H, Watanabe T, Ito T. Reliability of 80-Hz amplitude-modulation-following response detected by phase coherence. Audiol Neurootol 1999;4:28–37.

Cebulla M, Stürzebecher E, Wernecke KD. Untersuchung verschiedener SNRSchätzer für den Nachweis von biologischen Reizantworten im Rauschen. Z Audiol 2000;39:14–22. Cone-Wesson B, Dowell RC, Tomlin D, Rance G, Ming WJ. The auditory steady-state response: comparisons with the auditory brainstem response. J Am Acad Audiol 2002;13:173–87 [quiz 225–6]. D’Haenens W, Vinck BM, De Vel E, Maes L, Bockstael A, Keppler H, et al. Auditory steady-state responses in normal hearing adults: a test–retest reliability study. Int J Audiol 2008;47:489–98. Don M, Elberling C. Evaluating residual background noise in human auditory brain–stem responses. J Acoust Soc Am 1994;96:2746–57. Hoke M, Ross B, Wickesberg R, Lütkenhöner B. Weighted averaging – theory and application to electric response audiometry. Electroencephalogr Clin Neurophysiol 1984;57:484–9. International Electrotechnical Commission (IEC). 60318-1. Electroacoustics – simulators of human head and ear – part 1: ear simulator for the measurement of supra-aural and circumaural earphones; 2009. John MS, Dimitrijevic A, Picton TW. Weighted averaging of steady-state responses. Clin Neurophysiol 2001;112:555–62. John MS, Lins OG, Boucher BL, Picton TW. Multiple auditory steady-state responses (MASTER): stimulus and recording parameters. Audiology 1998;37: 59–82. Lütkenhöner B, Hoke M, Pantev C. Possibilities and limitations of weighted averaging. Biol Cybern 1985;52:409–16. Mühler R, von Specht H. Sorted averaging – principle and application to auditory brainstem responses. Scand Audiol 1999;28:145–9. Picton TW, John MS, Dimitrijevic A, Purcell D. Human auditory steady-state responses. Int J Audiol 2003;42:177–219. Rahne T, von Specht H, Muhler R. Sorted averaging—application to auditory event-related responses. J Neurosci Methods 2008;172: 74–8. Riedel H, Granzow M, Kollmeier B. Single-sweep-based methods to improve the quality of auditory brain stem responses. Part II. Averaging methods. Z Audiol 2001;40:62–85.