Elsevier

Clinical Neurophysiology

Volume 117, Issue 8, August 2006, Pages 1772-1782
Clinical Neurophysiology

Minimization of cochlear implant stimulus artifact in cortical auditory evoked potentials

https://doi.org/10.1016/j.clinph.2006.04.018Get rights and content

Abstract

Objective

To compare two methods of minimizing cochlear implant artifact in cortical auditory evoked potential (CAEP) recordings.

Methods

Two experiments were conducted. In the first, we assessed the use of independent component analysis (ICA) as a pre-processing filter. In the second, we explored the use of an optimized differential reference (ODR) for minimizing artifacts.

Results

Both ICA and the ODR can minimize the artifact and allow measurement of CAEP responses.

Conclusions

When using a large number of recording electrodes ICA can be used to minimize the implant artifact. When using a single electrode montage an optimized differential reference is adequate to minimize the artifact.

Significance

The use of an optimized differential reference could allow cortical evoked potentials to be used in routine clinical assessment of auditory pathway development in children and adults fit with cochlear implants.

Introduction

The latency and morphology of the cortical, auditory-evoked potential (CAEP) can provide information about the maturation of central auditory pathways (Ceponiene et al., 2002, Eggermont, 1988, Ponton et al., 1996, Ponton et al., 2000, Sharma et al., 1997, Sharma et al., 2002b, Sharma et al., in press). In a series of papers, we have documented the use of the CAEP as a measure of central auditory maturation in children who have received cochlear implants (CIs) and have suggested use of this measure in monitoring auditory pathway development in hearing-impaired patients (Sharma et al., 2002a, Sharma et al., 2002b, Sharma et al., 2002c, Sharma et al., in press). In normal hearing children, the latency of the P1 CAEP decreases systematically as age increases (Ponton et al., 2002, Sharma et al., 2002b). For newborn infants the latency of the P1 can be as long as 300–400 ms. The latency can be as short as 50 ms in adults. Thus, the region of interest for the P1 is between 50 and 300 ms. Recording the CAEP in CI patients poses a unique problem in that implant devices, during stimulation, create electrical artifacts on the scalp, which interfere with identification of the CAEP (Sharma et al., 2002b, Singh et al., 2004). Several aspects of these electrical artifacts can be visualized by comparing the averaged CAEP records in Fig. 1 obtained using free-field acoustic stimulation with a normal-hearing child and a child fit with a CI. For the normal-hearing child the CAEP is dominated by the P1 neural response in a post-stimulus latency window of 50–300 ms. In the case of the CI user the record of the P1 response is obscured by a large magnitude (50 μV) pedestal beginning slightly after the stimulus onset and ending slightly after the stimulus offset. This pedestal is followed immediately by a larger negative-going overshoot and subsequent ringing of the recording amplifier filters. In addition, there is a low magnitude noise floor introduced throughout the entire record due to the ongoing background stimulation by the speech processor even during periods of low sound levels.

To a large extent, distribution of the artifact on the scalp is influenced by the type of cochlear implant device, its mode of stimulation (e.g. monopolar or bipolar coupling), and the surgical placement of the remote return electrode. The placement of the remote return electrode for monopolar stimulation varies widely across devices, being integrated into the stimulator package in the Clarion device but implemented as a separate electrode in the Nucleus and Med El devices. In our earlier work (e.g. Sharma et al., 2002b) we reported an artifact in about 12% of cases when using a contralateral mastoid as the reference electrode. In that earlier work, data were collected in large part from patients fit with the Nucleus 22 device, which uses bipolar stimulation. Devices running with bipolar electrodes in the cochlea produce substantially smaller artifact on the scalp as compared to the now commonly used monopolar-coupled electrodes. In fact, recent data collected in our laboratory have revealed a much greater incidence of the artifact problem from patients with devices using monoplar stimulation as described above. In general, the monopolar configuration will be the most common configuration used in the future, so the future incidence of artifact issues is likely to be very large in children receiving cochlear implant devices. Because the presence of the CI stimulation artifact diminishes the utility of the CAEP in children fit with implants, it would be useful to explore and understand the nature of the artifact in order to create methods of minimizing the artifact in CAEP recordings.

The scalp recorded EEG is assumed to be a linear, instantaneous mixture of multiple neural sources plus noise, and when multiple EEG epochs are averaged in response to a common auditory stimulus the CAEP reflects neuronal activity in response to that stimulus (Makeig et al., 2004, Nunez, 1981, Scherg and Von Cramon, 1986). In practice, most of the noise recorded in the EEG is minimized by averaging multiple EEG epochs in response to repeated stimulation. Because brain activity responding to the stimulus is assumed to be represented in each EEG trial, the resulting average should reveal the appropriate average evoked potential. Electrical activity generated from the implanted electrode array is not temporally random throughout the EEG recording because bipolar electrical pulses are generated with each presentation of the auditory stimulus. Therefore, the CI stimulation artifact, as well as the biologic response, are time-locked to the stimulus and are represented in the averaged response. Assuming that the recording system maintains linear operation throughout, one possible correction for this might be to use an acoustic stimulus that alternates in polarity on subsequent stimulus presentations throughout the recording, thus canceling the averaged stimulus artifact for an even number of trials. However, this is not possible in recordings from CI patients wearing their clinical processors as the speech processors do not encode the phase of the incoming acoustic stimulus. This approach would also require temporal synchronization of the external acoustic stimulus, the pulse train delivered by the processor, and the sampling clock of the recording system (Miller et al., 2000, van den Honert and Stypulkowski, 1986). This is an achievable goal using specialized hardware and software, but is beyond the capability available in most clinics.

A CI stimulation artifact will last for at least the duration of the stimulus. Given that the amplitude of the artifact can be 5–10 times larger than the averaged evoked response, the artifact will mask a biologic response of interest that occurs within the time frame of the stimulus duration. We have used a 97 ms speech sound, /ba/, to elicit the CAEP in implant patients (Sharma et al., 2002a, Sharma et al., 2002b, Sharma et al., 2002c). It is reasonable to assume that shortening the speech stimulus would result in a shorter artifact—one that does not coincide with the region of interest for the P1. To determine if this is the case we compared CAEPs from a short duration (23 ms) vowel sound, ‘uh’, and a 97 ms speech sound, /ba/. As can be seen in Fig. 2, the duration of the artifact pedestal was shorter in the brief stimulus condition. However, the artifact did, in fact, overlap the early portion of the time region of interest because of filter ring. Limiting the amount of filter ring may be possible by changing the analog filter characteristics of the recording amplifiers (Andersen and Buchthal, 1970). However, the use of wide-band filters and higher sampling rates may also limit the feasibility of the CAEP as a clinical tool, because additional biologic artifacts at higher frequencies (e.g. muscle activity and fast ocular activity) often contaminate higher bandwidth recordings. The introduction of additional artifacts will increase the computational resources required to achieve a useful CAEP response and increase the time needed to retain useable EEG recordings. Obtaining a useful CAEP in a relatively short period of time is of great importance in a clinical setting, especially when limited recording time is an issue (e.g. recording from small children). The 23 ms signal used in this preliminary test is at the lower limit of duration for speech stimuli and still produced an artifact in the time region of interest. Thus, it would be best to look for other techniques to minimize the stimulus artifact.

Our experience has shown that, if not handled in a robust manner, the electrical artifact will either obscure the desired biological response, or worse yet, be misinterpreted as a biological response. The later outcome is common when CAEP recordings are low-pass filtered at 30 Hz—a typical procedure when processing CAEP recordings. In this paper we examine two approaches for minimizing the two problems described above.

Several techniques have been proposed for removal of EEG artifacts that occur from biological sources such as ocular, muscle, and cardiac activity. A technique commonly used for artifact reduction is principal components analysis (PCA) (Casarotto et al., 2004, Croft and Barry, 2002, Jung et al., 2000a, Pantev et al., 2005, Vigario et al., 2000). PCA is a statistical technique that decorrelates data into a series of factors based upon the amount of variance explained. The first principal component explains the largest amount of variance in the original dataset, the second component the second largest amount of variance, and so on. However, because PCA only identifies orthogonal components based upon the variance of the data, this technique may not completely separate the biologic artifacts from the neural responses, and some loss of EEG data may occur when using this approach (Croft and Barry, 2002, Jung et al., 2000a, Jung et al., 2000b). Ideally, a decomposition of underlying activity should maximize the independence of the sources contributing to the EEG activity and minimize the loss of EEG data that may be of interest after removal of the unwanted signals. Independent Component Analysis (ICA) has been proposed as a technique for achieving the required signal separation (Bell and Sejnowski, 1995, Delorme and Makeig, 2004, Jung et al., 2000a, Makeig et al., 1997, Makeig et al., 2004, McKeown et al., 1998, Vigario et al., 2000).

The ICA model is a generative model that maximizes information from higher-order statistics (typically using an analysis of kurtosis or negentropy) to identify factors, or components, that are uncorrelated and mutually independent. Essentially, the ICA model first decorrelates the dataset using a PCA model (second order statistics). Next, an iterative process changes the weights and directions of the vectors in a mixing matrix until maximum independence is identified from the higher order statistics and the data converge. The results of this generative model are a set of components that represent the underlying structure of the data. Theoretically, each independent component represents the activation of one contributing source to the average evoked potential. Therefore, it should be possible to linearly subtract artifactual components from the ICA mixing matrix.

The ICA model must satisfy a series of criteria about the underlying sources. First, the sources are considered to be maximally independent; that is they are statistically uncorrelated with other sources. Second, the sources must have non-Gaussian distributions. Third, the sources should be, ideally, stationary (non-stationary ICA should be considered a separate problem). In the case of the CI stimulation artifact all of these criteria are satisfied. It is important to consider that these assumptions are strictly statistical in nature and do not rely on the physiologic or biologic nature of the signals to meet these criteria. The artifactual sources from the implanted device are independent, as they are generated by the implant array, and not by other neural sources. Further, the activity from the array is generated relative to one or more common electrodes, which correlates this activity and reduces the number of underlying components that are mutually independent. Most of the clinical processors worn by the patients utilize a single, common return electrode implanted at a remote location under the scalp. Because this electrode is common to each other electrode, the recorded activity is statistically correlated and, therefore, not independent. The separation of the components is then limited to the independence of the signals recorded on the scalp, which consists of the low-frequency pedestal and the high-frequency information from the biphasic pulses along the implanted electrode array, and is separated in the PCA in the first stages of ICA decomposition. The artifact is time-locked to the auditory stimulus, occurring at the same time in each EEG trial and, thus, is stationary and non-Gaussian in distribution. Based on the properties of the CI stimulation artifact in the EEG, ICA is a plausible technique for identifying and removing unwanted components in the EEG as a pre-processing step before averaging.

Data from a large number of channels (ranging from 16 to 64) must be collected in order to implement an ICA analysis. However, most audiology clinics do not have instrumentation to accomplish this type of analysis. Thus, it is necessary to explore techniques of artifact reduction, which could be implemented using a small number of recording channels. In the present study we explored a recording technique, the optimized differential reference technique (ODR), to minimize contribution of the artifact in the evoked potential recording. Based on observations from 32 to 64 channel recordings we established that the electrical activity generated by the implanted array is broadly distributed on the scalp and generally has a dipole distribution with peak magnitude levels near the active stimulation electrode(s) located within the cochlea and a common extracochlear return electrode located remotely beneath the scalp. Recalling that our object is to record a CAEP at Cz relative to a remote reference electrode (typically placed on the contralateral mastoid), we seek to minimize the artifact measured differentially at Cz by selecting a more optimal reference electrode site (Kornfield et al., 1985, McGill et al., 1982, Nilsson et al., 1988). There may be many such reference electrode sites on the scalp, which meet the criteria of (1) being located along an isopotential electrical artifact contour passing through Cz and (2) being sufficiently far away from Cz to be electrically neutral to the CAEP events being recorded at Cz. In particular, we recorded from several reference-electrode sites around the forehead with the aim of determining the reference site that showed a null artifact (i.e. a location along the isopotential equal to that at Cz). If the spatial location of this isopotential could be estimated, then placing the reference electrode at this location should minimize the contribution of the artifact (McGill et al., 1982). In other words, we aimed to place our reference electrode at a location where differential recording would minimize the artifact.

In this study, we explored two methods to minimize the stimulation artifact during CAEP recordings. In Experiment 1 we applied ICA to the contaminated EEG from a group of CI users. In this analysis, multiple components in each recording were identified and attributed to the implanted device. After removing the unwanted signals, the EEG was recomputed and processed for evaluation of the CAEP. In Experiment 2 we explored the use of an optimized differential reference, i.e. where signals on an active electrode (Cz) were differentially recorded relative to a reference electrode located at various positions. The aim was to identify a reference location along the isopotential contour of the artifact appearing at Cz and, thus, minimize the artifact by recording differentially. Finally, we compared the CAEP obtained after removing the artifact using ICA to the CAEP recorded using the ‘optimized differential reference technique’ for artifact minimization. At issue was whether the two techniques to remove the CI stimulation artifact revealed a similar biological response.

Section snippets

Subjects

Subjects were 5 children aged 5.2–12.7 years (mean age =10.46) who had been fitted with a cochlear implant. The children were selected for testing because in a previous visit to our laboratory their CAEP records had been obscured by an artifact. The children received their implants at ages ranging from 2.6 to 10.9 years (mean implant age =6.5). Four subjects used Nucleus 24 devices with a monopolar configuration, and one subject (Subj B) used a Nucleus 22 device with a bipolar configuration.

Scalp maps of the CI stimulation artifact

Scalp maps of the averaged evoked activity from the evoked potential recordings revealed a scalp artifact distribution that concealed biologic activity in the time range of the CAEP (Fig. 3). In each case, the artifact was centered on the hemisphere of the CI device and peaked in the vicinity of the subcutaneous return electrode. Although, the artifact amplitude was much lower at the Cz electrode, which is typically used for CAEP recordings, than at sites near the implant, sufficient artifact

Discussion

The CAEP could be a useful clinical tool for inferring the maturational status of the central auditory system in CI patients if the stimulus artifact could be minimized. In the present study we examined two different techniques to minimize the contribution of the artifact to the CAEP response. In Experiment 1 we explored the use of ICA as a preprocessing filter for minimization of the artifact. ICA decomposition was performed on 66-channel recordings in five subjects. In all five cases at least

Acknowledgements

We wish to thank the participants of this study; the children and their families for their enthusiastic participation. We wish to thank the two anonymous reviewers for their insightful comments and helpful suggestions. Comments from Arnaud Delorme and Julie Onton concerning the use of ICA were very helpful in guiding this work. Funding provided by the National Institutes of Health: NIH-NIDCD R01-DC004552 and R01-DC006257.

References (31)

  • A. Sharma et al.

    Developmental changes in P1 and N1 central auditory responses elicited by consonant–vowel syllables

    Electroencephalogr Clin Neurophysiol

    (1997)
  • C. van den Honert et al.

    Characterization of the electrically evoked auditory brainstem response (ABR) in cats and humans

    Hear Res

    (1986)
  • V.O. Andersen et al.

    Low noise alternating current amplifier and compensator to reduce stimulus artefact

    Med Biol Eng

    (1970)
  • A.J. Bell et al.

    An information-maximization approach to blind separation and blind deconvolution

    Neural Comput

    (1995)
  • R. Ceponiene et al.

    Maturation of cortical sound processing as indexed by event-related potentials

    Clin Neurophysiol

    (2002)
  • Cited by (145)

    View all citing articles on Scopus
    View full text