Through a series of EEG experiments that manipulated reference electrode placement and application of offline frequency domain and spatial filters, we observed AEPs with a peak-component structure containing a prominent N1 at approximately 50 ms, and P2 at approximately 75 ms. Accounting for a polarity inversion attributable to electrode montage, this component structure closely resembles that reported by Hernandez et al. (2007). Application of a 50-Hz low-pass filter during EEG pre-processing slightly increased the SNRL, but choice of reference electrode location had a larger effect on the noise in the responses. For analysis in sensor space, the best SNRL for the auditory cortical response was obtained at the non-inverting electrode 20 cm posterior to the blowhole on the midline referenced to the meatus. In component space (from ICA), the greatest SNRL was obtained using the dorsal reference montage.
Recording and pre-processing considerations
We used SNRL measures to compare the fidelity of AEPs over the presumed cortical response latency window recorded from different electrode montages and pre-processing parameters. Many ERP studies estimate signal strength based on peak amplitudes, and rely on a baseline correction, or subtraction of a pre-stimulus voltage from the remaining epoch window, to effectively center the voltage to zero at the time of stimulus onset. Given that our recordings included significant artifact contamination during the pre-stimulus baseline for some conditions, it seemed prudent to use another method to assess AEP fidelity. We adopted a method commonly used in human ABR analysis (Elberling and Don 1984), which was previously applied to dolphin ABR measurements using a similar recording paradigm to that used here (Finneran et al. 2019). By the SNRL metric, the meatus-referenced montage produced the highest fidelity AEP. While not addressed here, follow-up work could make use of this SNRL metric to estimate the trade-off between number of epochs and SNRL for mid- or long-latency AEPs to help characterize what might be an optimal number of epochs for ACR experimental sessions. As previously described by Popov and Supin (1986), somewhere between 300 and 3000 epochs was appropriate for most applications. However, their experiment was performed with needle electrodes with an insertion depth of 2–3 mm, which yields recordings with potentially higher SNR than electrodes applied to the skin surface.
One of the parameters we varied was the location of the reference electrode. Animals in two prior studies were suspended in a stretcher in a shallow pool (Woods et al. 1986; Popov and Supin 1986). This allowed the use of electrode locations, such as on fins, which could present additional complications in the study of free-swimming dolphins in underwater environments with an EEG system. Since our long-term goal is to undertake such studies, we did not test these locations. Likewise, we did not consider placing electrodes on the rostrum. As we want to test awake, unrestrained animals who are periodically fed with fish throughout the recording session, attachments to the rostrum are likely to fail. These considerations guided our selection of the three reference electrode locations we tested. While the meatus position was lateralized similarly to a mastoid reference in humans, the dorsal and melon sites were at midline, which should avoid a lateral bias in the measures.
Application of a digital 50-Hz low-pass filter did not significantly affect amplitude measures of the ACR. While it nominally increased the estimated SNRL by about 1 dB, it did not seem to affect the peak to peak amplitude of the N1–P2. This supports the use of such restrictive filtering in future experimental protocols where there is the potential for contamination by higher frequency noise. It also supports the use of such a filter as part of the pre-processing for ICA.
ICA decomposition produced at least one component for each of the three reference montages that contained enough information to represent an AEP source relatively devoid of noise contamination. As ICA maximizes the independence of shared information between components (Makeig et al. 2004), it will inevitably affect the underlying variance across epochs in the component activations. This could bias the noise term in the SNRL estimates, thus inviting caution in interpretation of the relative component SNRL values. For example, the first and third components from the meatus-referenced montage have roughly equivalent SNRL values; however, the average waveform for component one contains much larger background noise than the average waveform for component three (see Fig. 4A, C). The larger noise term in component one is offset by a larger signal term, as well. Both components appear to represent a portion of the true AEP. This suggests that the decomposition could be even more successful at isolating the AEP with a larger array of electrodes. Given that we used only three non-inverting electrodes, ICA has limited utility and cannot adequately decompose all of the diverse underlying neural and noise sources in the recordings into a full-rank set of three components. However, even in this limited capacity with a small electrode array, the ability to separate an AEP from background noise shows significant potential for application to future dolphin studies where recording conditions are less than ideal. With ICA separation of the ECG component, the dorsal-referenced montage produces better results for a relatively clean AEP component. Thus, it may be that a reference location even farther away from the AEP generator tissue than the meatus or melon references could effectively serve as a more neutral reference in that it is less likely to contain AEP information.
It can be difficult to interpret ICA components, and they can yield ambiguous results, such as the polarity of activations and their resulting ERPs. As visible in the melon-referenced trace in Fig. 4C, the polarity of a component activation can flip relative to the information it represents from sensor space. As ICA may converge on a solution that has two possible polarity values for the weight matrix, the sign of the sensor-space data is represented by the product of both the component activation and its inverse weight matrix for back projection (Onton et al. 2006).
ICA solutions are not found from a closed form computation, but instead are estimated from a stochastic process. Therefore, the same (or very similar) data can produce different solutions. To deal with this uncertainty, we employed a reliability algorithm using split-half comparisons (Groppe et al. 2009). While there is no objective criterion for determining the reliability of an ICA solution, if disjoint data sets produce similar solutions, it bolsters faith that the decompositions are capturing key aspects of the data. This approach is strengthened if solution similarity, captured as the cosine distance and residual variance measures, is greater for data sets that should contain the same signal and noise components (e.g., both representing statistically identical AEP or ECG activity) than for data sets that should differ (e.g., comparing components derived from AEPs and ECGs). If the former are smaller than the latter, it supports the view that the decomposition is identifying functionally relevant and reliable components of the measurements.
Interpretation of auditory-evoked potentials
Our ACR results resemble those reported by Hernandez et al. (2007), but with opposite polarity. This difference is likely explained by differences in the locations of the non-inverting electrodes on the head surface in the two studies. Supin et al. (2001) described a shift in polarity of the ACR moving along the anterior–posterior axis, with a polarity flip approximately 5–10 cm posterior to the caudal lip of the blowhole. Hernandez et al. (2007) reported their non-inverting electrode was 4 cm posterior to the blowhole; in contrast, we used montages with electrodes 10, 20, and 30 cm posterior to the blowhole, which supports this anterior–posterior axis change in polarity.
Our results found the greatest SNRL in sensor space for the 20-cm non-inverting electrode, consistent with prior reports that the largest magnitude ACRs are recorded 20 cm posterior to the blowhole (Popov and Supin 1986). However, differences in electrode depth and inverting electrode placement can also influence SNRL and ACR magnitude measures. To address this directly, one could perform a follow-up multi-electrode study using non-inverting recording sites that span the range over which the polarity is expected to flip, including the locations we tested here and those used by Hernandez et al. (2007). The results could help to pinpoint the location of auditory cortices, or at least the orientation of a prospective AEP dipole projecting from the cortical surface to the skin surface (described as “tilted to the rostroventral-dorsocaudal direction” by Supin et al. 2001).
The latency of a potential originating from a cortical source to anywhere on the skin surface is very similar, given conductance times for electrical signals. However, the way in which potentials from different generators in the cortex sum at the skin surface can alter peak values as well as peak latencies in the summed electrical activity; moreover, these effects differ, depending on electrode location (Luck 2014). In the present data, the temporal window that Hernandez et al. (2007) used to quantify P50 and N75 magnitudes resulted in a lower mean magnitude response at the 20 cm non-inverting electrode site than a temporal window of equal length that was shifted earlier in time. Electrode location is likely an important factor in explaining this difference; however, latency differences at the same electrode site can also occur due to idiosyncratic differences between dolphins (such as differences in brain geometry, or attentiveness). Determining the optimal temporal window over which to extract component peaks requires additional research using a greater number of non-inverting electrode locations and testing additional dolphins.
We observed that a 50-Hz low-pass filter yielded a nominal increase in SNRL in our measures (\(^\sim 1\) dB). While often useful as a pre-processing stage for ICA, low-pass filtering may not be appropriate for all ACR studies, as AEPs were detectable even without filtering. If filtering is not used and SNRLs are smaller, it is likely to be more important to summarize AEP magnitudes by calculating the mean over a window centered around the expected peak time, rather than a local max or minima; specifically, noise at frequencies higher than the dominant frequencies of an ERP component can have a greater impact on peak values than on means (Luck 2014). This further underscores the need to better understand latency response variability for components such as the N1 and P2, and how this may vary across individual dolphins.
We observed a stimulus offset response, approximately 25 ms after the end of the stimulus, that follows the N1–P2 peak structure described for human offset AEPs (Hillyard and Picton 1978). The offset response was present in the sensor-space recordings from the meatus reference, as well as in the third ICA component from the dorsal and meatus references. This response was previously observed in dolphins with skull surface electrodes in response to a 500-ms duration pure tone and FM sweep (Ridgway 1980). The offset response was not reported by Hernandez et al. (2007), which may reflect differences in how strong the offset response is for different electrode montages. It may also be explained by differences in stimulus duration, as Hernandez et al. (2007) used 100-ms duration stimuli; in humans, the offset response strength increases as stimulus duration increases (Hillyard and Picton 1978) . The ACR offset response in humans is a form of an acoustic change complex (ACC), or response to a change in ongoing sound, such as a transition from a tone to white noise (Martin and Boothroyd 1999). This kind of ACC could be useful in probing how/where specific sound sequences or transitions are processed by the dolphin cortex.
Conclusions
By comparing responses measured in different reference electrode montages, we find that a meatus electrode served as the best reference for sensor-space AEP measures, whereas a dorsal electrode served as the best reference for component-space measures, in terms of largest SNRL and peak-to-peak measures. Given that the dorsal reference montage introduced significant heartbeat contamination, complicating monitoring of the recording progress in sensor space, the utility of a dorsal reference might be limited in practice. Both sensor- and component-space AEP measures and SNRL estimates nominally benefited from a 50-Hz low-pass filter in the pre-processing stage. The use of a meatus reference is consistent with many previous ABR recordings, and may provide advantages in terms of building on prior literature to examine the time course of an AEP from brainstem to cortical potentials. These are important considerations, as the choice of the reference electrode has not previously been discussed thoroughly for dolphin AEPs. The selection is important to avoid or mitigate artifacts from cardiac or other myogenic sources. It should also be comparable to previously published literature (for further discussion, see Luck 2014). To effectively build upon this and prior work, further experiments should be conducted with the same electrode montages. Future multi-electrode studies should include a non-inverting sensor placed just posterior to the blowhole, as reported for mid- and long-latency AEPs (Hernandez et al. 2007) and short-latency brainstem responses (Finneran et al. 2019), in addition to one placed 20–30 cm posterior to the blowhole, as recommended from the current results, and supported by previous findings (Supin et al. 2001; Popov and Supin 1986).
All measures reported here were from a dolphin in-air and lying on a foam mat. AEPs collected in dolphins submerged in seawater cannot be unequivocally compared to the current results. Underwater recording of ACRs presents unique challenges, including increased muscle artifacts from movement of the dolphin to stabilize itself against ocean waves, and maintenance of a consistent distance and orientation between its head and a sound projector. The saltwater also acts as an electrical volume conductor that shunts current away from the dolphin’s head (Supin et al. 2001). Furthermore, ambient noise while dolphins are underwater might confound the interpretation of ACRs, at least in testing environments like San Diego Bay. All of these challenges will need to be addressed to move toward the study of dolphin auditory cognition in a more naturalistic situation.