Evaluating and Comparing Behavioural and Electrophysiological Estimates of Neural Health in Cochlear Implant Users

Cochlear implants (CI) restore a sense of hearing to people with sensorineural hearing loss by stimulating the spiral ganglion neurons (SGN) of the auditory nerve. The ability of a CI to transmit speech information partly depends on the condition of the stimulated neural population, which will be referred to as neural health throughout this article. Estimates of neural health patterns along the length of the cochlea may help to guide the optimisation of sound processing strategies, by identifying electrodes in healthy neural regions that might benefit from focused stimulation (Bierer 2010) or electrodes in unhealthy neural regions that could be candidates for deactivation (Garadat et al. 2013; Zhou 2017; Goehring et al. 2019). Several different methods to estimate neural health patterns along the cochlea have been proposed. The present study explores the extent to which three different measures provide converging and consistent estimates of neural health. We consider a combination of both behavioural and electrophysiological measures of neural health.

The behavioural neural health estimates that we evaluated were multipulse integration (MPI) and the polarity effect (PE). MPI is measured either as the slope of the function relating psychophysical detection thresholds to stimulation pulse rate or as the difference between psychophysical detection thresholds at different stimulation pulse rates. As pulse rate is increased, detection thresholds decrease, with the slope being steeper at pulse rates greater than 1000 pps in both animal studies (Kang et al. 2010; Pfingst et al. 2011; Zhou et al. 2015; Pfingst et al. 2017) and studies in CI listeners (Shannon 1985, 1989; McKay and McDermott 1998; Kreft et al. 2004; Zhou et al. 2012, 2015). This characteristic rate-threshold function is consistent with a temporal integration mechanism (Stypulkowski and van den Honert 1984; Shannon 1989; McKay and McDermott 1998; McKay et al. 2003; Carlyon et al. 2005), whereby the neural excitation by consecutive pulses is summed in a sliding integration window. The magnitude of the MPI slope for pulse rates below 1000 pps has been positively correlated with SGN density in guinea pigs (Kang et al. 2010; Pfingst et al. 2011; Zhou et al. 2015). In those studies, guinea pigs were behaviourally trained to complete a detection threshold task, and then split into two groups: one that was implanted after being deafened with neomycin (leading to no residual hearing and low SGN density) and another that was implanted without being deafened (leading to some residual hearing and high SGN density). MPI slopes were significantly larger for the group of guinea pigs that had been implanted without being deafened. For bilateral CI users, between-ear differences in MPI were correlated with between-ear differences in speech perception (Zhou and Pfingst 2014). These results have led to the use of MPI as a behavioural estimate of neural health in humans (Hughes et al. 2014; Zhou and Pfingst 2014, 2016a, 2016b; McKay and Smale 2017; Zhou and Dong 2017; Zhou et al. 2018).

PE is measured as the difference in psychophysical detection thresholds between asymmetric pulse shapes in which the portion where charge that is most concentrated in time is anodic versus where it is cathodic (Carlyon et al. 2013; Macherey et al. 2017). For simplicity, we will describe these charge-balanced asymmetric pulses as ‘anodic’ and ‘cathodic’ pulses, respectively. Modelling studies show that cathodic pulses preferentially stimulate the peripheral processes of an SGN fibre, while anodic pulses preferentially stimulate the central axon (Rattay et al. 2001; Joshi et al. 2017; Resnick et al. 2018; Potrusil et al. 2020). In general, SGN degeneration begins with the peripheral processes and progresses towards the central axon (Spoendlin 1975,Spoendlin 1984; Leake and Hradek 1988; Otte et al. 1978; Hinojosa and Marion 1983; Nadol 1990; Wise et al. 2017). Theoretically, a population of SGN fibres with peripheral degeneration and healthy central axons would have higher threshold for cathodic than for anodic pulses, whereas a population of SGN fibres with healthy peripheral processes and central axons would show a smaller threshold difference between pulse shapes (low or negative PE). At low levels, PE varies across the electrode array within subjects (Macherey et al. 2017; Carlyon et al. 2018; Goehring et al. 2019; Jahn and Arenberg 2019), suggesting that PE at threshold might be a localised indicator of peripheral neural health. Both Jahn and Arenberg (2019) and Mesnildrey et al. (2020) provide psychophysical evidence for PE as an estimate for neural health. Both studies found that PE was correlated with thresholds for symmetric pulses obtained with focused stimulation, which have been shown to be estimates of combined properties of the electrode-neural interface. The electrode-neural interface refers to any factor that affects the transmission of information between the implanted electrodes and the neural population, including both neural health and non-neural factors such as electrode position, electrode orientation and impedance. Importantly, this correlation between PE and focused thresholds was significant even when the effects of electrode-modiolar distance (‘EMD’, an estimate of the distance of the electrode from the auditory nerve) were removed. It should also be noted that the model proposed by Rattay et al. (2001) predicts that PE may also be affected by EMD, although we know of no evidence that it correlates with the PE independently of focused thresholds.

We used features calculated from the electrically evoked compound action potential (ECAP) as the electrophysiological estimates of neural health. While many absolute ECAP measures, such as maximum ECAP amplitude (Shepherd and Javel 1997) and slope of the ECAP amplitude growth function (Ramekers et al. 2014; Pfingst et al. 2015), have been positively correlated with neural health in animals, these measures are prone to the influence of non-neural factors such as electrode position (Shepherd et al. 1993) and impedance (Schvartz-Leyzac and Pfingst 2016). Non-neural factors are especially influential in humans, who have larger cochleas and longer durations of deafness compared to other animals, leading to wider electrode-to-electrode variance in electrode position and impedance. To remove the influence of non-neural factors, ratios can be calculated between absolute ECAP features that are measured with different stimulus parameters (e.g. interphase gap (IPG) or phase duration). These differential ECAP measures can potentially partial out the effects of non-neural factors, which are assumed to be the same across stimulus conditions. In separate guinea pig studies, Prado-Guitierrez et al. (2006) and Ramekers et al. (2014) measured the dB difference in current level required to elicit an equal normalised ECAP amplitude at different IPGs, and found that this ‘IPG offset’ was correlated with SGN density. These results have led to measures such as the IPG offset (and similar differential ECAP measures) being commonly used to estimate local neural health along the array in humans (Kim et al. 2010; Schvartz-Leyzac and Pfingst 2016; McKay and Smale 2017; Hughes et al. 2018; Schvartz-Leyzac and Pfingst 2018). We recently evaluated different ways of defining the effect of IPG on neural responses and concluded that the IPG offset provided a robust measure that was minimally affected by non-neural factors (Brochier et al. 2020).

Here, the electrophysiological measure of IPG offset and the behavioural measures of PE and MPI were evaluated and compared in a group of 11 CI users at seven different electrode locations spanning the array. Within-subject and between-subject correlations were evaluated to determine the extent to which the different neural health estimates provide converging information. It was hypothesised that if each of these measures reflects a similar characteristic of neural health, then significant within-subject and between-subject correlations would be found between all of the measures. Further, a computational model of the electrically stimulated auditory nerve (Joshi et al. 2017) was implemented to explore what factors may have contributed to the different neural health estimates.

Run 1 and run 2 were highly correlated for both PE (R(10) = 0.75, p < 0.001) and MPI (R(9) = 0.82, p < 0.001), suggesting good test–retest reliability for the behavioural measures (Fig. 2). Test–retest reliability was not explicitly measured for the IPG offset measure, but 50–100 sweeps were averaged to obtain each ECAP AGF from which the IPG offset was calculated.

An analysis of covariance (ANCOVA) with subject as a random factor was used to evaluate the correlations between PE, MPI and the IPG effect. No significant correlations were found between any of the proposed neural health estimates. PE was not correlated with MPI (R(9) = 0.11, p = 0.385), MPI was not correlated with the IPG effect (R(7) = − 0.29, p = 0.058) and PE was not correlated with the IPG effect (R(7) = 0.02, p = 0.887). The absence of correlations shows that the within-subject pattern of one neural health estimate across the electrode array is not predictive of the within-subject pattern of another neural health estimate across the electrode array, at least in our subject population.

Figure 3 shows the correlations between MPI and PE, MPI and IPG effect, and PE and IPG effect. In the upper panels, the mean value across electrodes for each subject was subtracted from the measurement at each electrode. The lower panels show the correlations between the mean values across electrodes for each subject. A Fisher R to z transform was used on the correlation coefficients for each individual to calculate the 95 % CI for the range of across-electrode correlations for each comparison of neural health estimates. The 95 % CI for the range of correlation coefficients for MPI and PE was 0.03 ± 0.4; for MPI and IPG effect, it was − 0.27 ± 0.33; and for PE and IPG effect, it was 0.04 ± 0.39. All of these ranges encompass zero, reflecting the lack of statistical significance of the correlations.

The between-subject correlations of the mean of each neural health estimate across the array were also calculated and are plotted in the bottom row of Fig. 3. No significant correlations were found between any of the estimates of neural health. Mean MPI across the array was not correlated with mean PE (R(9) = 0.35, p = 0.322) or mean IPG effect (R(7) = 0.12, p = 0.783), and mean PE was not correlated with mean IPG effect (R(7) = 0.068, p = 0.873).

As noted in the ‘Introduction’ section, a number of previous studies have shown that the PE is positively correlated with the average threshold obtained with both polarities (Carlyon et al. 2018; Jahn and Arenberg 2019) or with thresholds obtained with symmetric pulses (Mesnildrey et al. 2020). Our results replicate both of those findings: PE was positively correlated with the mean of QP-CAAC and QP-ACCA thresholds (R = 0.59, p < 0.001, Fig. 4a), and with 80-pps thresholds obtained with symmetric biphasic pulses (R = 0.40, p < 0.001). MPI was correlated with the mean of the 80-pps and 1000-pps biphasic thresholds (R = 0.35, p < 0.001, Fig. 4b).

MPI, PE and the IPG offset, all proposed measures of neural health, were evaluated and compared in 11 CI users. There were no significant correlations between the neural health estimates for either within-subject comparisons across the electrode array or between-subject comparisons of the means. The results therefore suggest that these neural health estimates do not reflect the same characteristic of neural health.

One clear contributor to the lack of within-subject correlations between the IPG offset and our behavioural measures is the different range of levels used to make these measurements. The IPG offset is measured well above threshold, at levels that are high enough to elicit ECAPs. In contrast, MPI and PE were measured at psychophysical detection thresholds. Therefore, the spread of current was likely larger and the stimulated neural population would have been broader for the IPG offset measurement than the behavioural measurements. However, the different overall levels used to measure IPG offset and our behavioural measures does not fully explain why we failed to see a between-subject correlation between those measures. Even if the stimulated neural population were broader for the IPG offset measurement, one would expect that the average IPG offset would correlate with the average behavioural measures between subjects, if they all reflected the same characteristics of neural health.

An explanation for the lack of between-subject correlations between the neural health measures, supported by both our experimental and modelling data, is that the different neural health measures reflect different characteristics of the neural population. A computational model was used to explore what factors may have contributed to the behavioural and electrophysiological neural health estimates in this experiment, and to identify distinct mechanisms that may have affected one measure independently of the other measures. Computational modelling provides a unique framework whereby different neural health factors can be assessed independently, which cannot be done easily in animal studies, or at all in human experiments. For example, the absolute number of neurons could be adjusted while maintaining all other neural properties, or the peripheral processes could be degenerated independently of the central axon. Our modelling results suggest that IPG offset may be related to central axon demyelination (particularly when the peripheral processes were disabled), PE may be related to health of the peripheral processes and MPI may be related to the SD of fibre thresholds.

At first, it may seem surprising that measures such as IPG offset and MPI were not affected by absolute neuron count in the computational model, because those measures have been shown to be correlated with SGN density in animals. However, it is has been shown that the number of nerve fibres (SGN density) covaries with neural degeneration/demyelination in animal studies. The modelling results suggest that the properties of the nerve fibres, as opposed to the absolute number of nerve fibres, influence the neural health measures. These results do not necessarily conflict with the animal studies that have shown a correlation between SGN density and IPG offset or MPI, and in fact, it may serve to expand on those data by probing the anatomical cause of the effect in more detail.

The idea that the IPG offset might reflect neural properties rather than the total number of neurons is supported by a recent study that measured the ECAP IPG offset in children diagnosed with cochlear nerve deficiency (CND) (Skidmore et al. 2020). CND is defined as a small or absent cochlear nerve, and CI outcomes for this pathology are generally poor. However, Skidmore et al. (2020) showed that the IPG offset was significantly larger for the CND group compared to the group of CI users without CND, indicating that the degree of neural degeneration was less in the group with CND than the non-CND group. This result is consistent with temporal bone studies, which have shown that the remaining neurons in CND patients (while sparse) are not necessarily degenerated (Ylikoski and Savolainen 1984). In another study, a cochlear microphonic was measured in 9 out of 13 ears with CND (Buchman et al. 2006), indicating the presence of innervated inner and outer hair cells. Hence, compared to typical CI users with severe sensorineural hearing loss, the CND group might have fewer but healthier neurons. The properties of those neurons, and not the number, are reflected by the significantly larger IPG effect in the CND group compared to the non-CND group.

‘Neural health’ is a broad term, composed of many potentially independent factors including SGN density, the distribution of firing thresholds, demyelination and degeneration of peripheral processes and/or the central axon, and temporal properties of the remaining SGNs. In animal studies, many of the factors have been shown to covary, but in humans, there is far more between-subject and within-subject variance in terms of aetiology and electrode-neural interface, which may interfere with the typical patterns of covariance that we observe in animals. For example, the computational model of MPI showed that it was affected by the SD of firing thresholds of the neural population. While the SD of firing thresholds would certainly be affected by the health of the neural population, it could also be affected by current spread and spatial selectivity.

MPI has been shown to be related to spatial selectivity, with higher MPI in monopolar mode compared to bipolar mode (Zhou et al. 2018), and higher MPI on electrodes that produce a wider spatial spread of excitation compared to those that produce a narrow spread (Zhou and Pfingst 2016b). In addition, in our behavioural results, MPI was positively correlated with the mean detection threshold for 80-pps and 1000-pps biphasic pulse trains, suggesting that higher current levels (higher current spread) lead to higher MPI. If the spread of excitation at a particular electrode were relatively wide compared to other electrodes, then the distribution of membrane voltages at nearby SGN fibres would also be wider, leading to a larger SD of firing thresholds. One contributing factor to the spread of excitation is the EMD. In guinea pigs, the EMD remains relatively constant across the electrode array, and variation in MPI between groups of animals deafened for different amounts of time is largely influenced by the health of the SGN fibres. In humans, EMD varies considerably along the length of the cochlea (Long et al. 2014), and the number of recruited nerve fibres (and the SD of their firing thresholds) would depend not only on the health of the nearby SGN fibres but also on the spread of excitation. However, a recent study by Schvartz-Leyzac et al. (2020) measured EMD and MPI at every electrode along the array in 11 CI users, and found that MPI was not correlated with EMD. While other factors, such as electrode impedance or fibrous tissue growth around the electrode, could still contribute to spatial selectivity and hence MPI (MPI in our study was correlated with impedance (R = 0.31, p = 0.015)), the result by Schvartz-Leyzac et al. (2020) indicates that MPI in most cases might be dominated by characteristics of the neural population.

In that same Schvartz-Leyzac et al. (2020) study, the IPG effect on ECAP AGF slope was measured for the same electrodes for which MPI was measured. The IPG effect on ECAP AGF slope differs from the IPG offset, in that its magnitude depends on absolute ECAP amplitude, and is therefore affected by both neural and non-neural factors (Brochier et al. 2020). That being said, the neural factors contributing to the IPG effect on ECAP AGF slope presumably involve similar underlying mechanisms to the IPG offset. Consistent with our experimental results, Schvartz-Leyzac et al. (2020) found no within-subject correlations between MPI and IPG effect on ECAPs. They draw similar conclusions to the present study, attributing the lack of correlation to differing underlying neural mechanisms between the effect of IPG on ECAP slope and MPI. Considering our computational model, along with the results of Schvartz-Leyzac et al. (2020), MPI is more related to variability of neural thresholds in the neural population, while the effect of IPG is reflective of the central axon characteristics, which might affect the response of the neuron to a single biphasic pulse.

The modelling results also suggest that the IPG offset is related to overall neural excitability. As IPG offset decreases, neural excitation decreases, as observed by the shifting of the AGFs towards the right as the IPG offset reduces. Physiologically, the interaction between IPG offset and neural excitability makes sense. If the IPG offset is related to biophysical properties of the central axon, neurons with a larger IPG offset would probably also be more responsive to electrical stimulation. In humans of course, we cannot measure neural excitability directly; absolute measures such as maximum ECAP amplitude, ECAP threshold or ECAP AGF slope are impacted by both neural and non-neural factors. In animal studies, where non-neural factors such as impedance and EMD are less influential, the aforementioned absolute measures are predominantly influenced by neural excitability. In those data, we see some evidence that the IPG offset is related to neural excitability. The Prado-Guitierrez et al. (2006) data, for example, showed a relationship between the IPG offset and the mean 50 % point on the AGFs for IPGs of 8 and 58 μs (on a logarithmic input and output scale), although the correlation did not reach significance (R = − 0.64, p = 0.061, N = 9).

An alternative hypothesis is that not all of these measures reflect neural health. While multiple animal studies have linked neural health to MPI (Kang et al. 2010; Pfingst et al. 2011; Zhou et al. 2015; Pfingst et al. 2017) and IPG offset (Prado-Guitierrez et al. 2006; Ramekers et al. 2014), the ability of PE to predict neural health has only been demonstrated in modelling studies (Rattay et al. 2001; Joshi et al. 2017; Resnick et al. 2018; Potrusil et al. 2020) and postulated to underlie results in human CI listeners (Macherey et al. 2017; Carlyon et al. 2018; Goehring et al. 2019; Jahn and Arenberg 2019; Mesnildrey et al. 2020).

Animal studies that have investigated PE have not shown the expected relationship between PE and duration of deafness (Macherey and Cazals 2016) or location of micro-lesions (Konerding et al. 2020). Macherey and Cazals (2016) measured the PE in deafened guinea pigs through inferior-colliculus recordings, and found no significant effect of duration of deafness (from 1 week up to 1 year) on PE. Konerding et al. (2020) applied micro-lesions to either the peripheral processes, the soma or the central axon in guinea pigs, and measured PE. Surprisingly, a negative PE (which has been proposed to indicate better peripheral neural health) was found in the group of guinea pigs with lesions in the soma or the cell body, compared to guinea pigs without lesions. It should be noted that the present study used biphasic pulses, which have not shown significant effects of polarity in humans. In addition, in the process of applying the lesions, Konerding et al. (2020) may have opened up a current pathway for both the cathodic and anodic-leading biphasic pulses to stimulate the central axon directly, which would negate any PE related to cathodic stimulation of the peripheral processes. The lack of histological evidence for PE as a measurement of peripheral health is likely due to morphological differences between humans and guinea pigs, in both the cochlea and in the neurons themselves. The smaller cochlea of the guinea pig may lead to central axon activation for both cathodic and anodic pulses, which would reduce or completely abolish the PE. At the neural level, guinea pigs have myelinated soma, compared to the unmyelinated soma of the human, which again would cause differences in the site of activation (Rattay et al. 2001). The above differences make the PE inherently difficult to validate in animal studies, compared to IPG offset or MPI. That being said, it has been repeatedly shown in human studies that PE is heterogeneous along the electrode array, and PE was found to be significantly different in the pre-lesion and post-lesion guinea pigs in the Konerding et al. (2020) study. Therefore, there must be some factor that affects PE differently along the cochlea, and there is some evidence that this factor is neural.

The combined experimental and modelling results provide evidence that PE, MPI and IPG offset reflect different characteristics of the electrode-neural interface. No correlations were found between the different neural health estimates, for either within-subject comparisons across the electrode array, or for between-subject comparison of the means. Modelling results suggest that PE may be related to peripheral neural health, IPG offset may be related to central axon demyelination and MPI may be related to the SD of firing thresholds (which in turn may be affected by spread of excitation).