A comparison of electrophysiological measures for characterizing the cochlear-implant electrode-neuron interface

Abstract

The health of the auditory nerve is a key factor for hearing outcomes in cochlear-implant patients. Many electrophysiological markers of auditory-neural health have been proposed, with varying predictive power and independence from non-neural factors. The “Failure Index” (FI), the ratio between stimulation current level and response magnitude of the electrically-evoked compound action potential (eCAP) was compared to the Panoramic eCAP method. Both methods predicted localized areas of reduced neural responsiveness in a group of human cochlear-implant users, but the FI showed a greater dependence on a non-neural factor, namely, the distances between the cochlear-implant electrodes and the auditory nerves.

1. Introduction

Cochlear implants (CIs) are neuro-prosthetic devices that provide a sense of sound by directly stimulating the auditory nerve using electrodes inserted into the cochlea (Mudry and Mills, 2013). While many CI users achieve good speech perception, there is a lot of variability between users, and many struggle to communicate with their implant, especially in noisy conditions (Blamey et al., 2013). Variation in the health of the auditory nerve has been proposed as one factor contributing to this variation in speech perception between patients (Pfingst et al., 2015). As the health of the auditory nerve cannot be characterized directly in humans, many behavioral (Bierer, 2010; Carlyon et al., 2018; Goldwyn et al., 2010; Kalkman et al., 2022; Peng et al., 2025; Zhou and Pfingst, 2014) and electrophysiological (Brochier et al., 2021; de Vos et al., 2018; DeVries et al., 2016; Garcia et al., 2021; Kim et al., 2017; McKay et al., 2013; Ramekers et al., 2014) markers have been proposed. Many of the electrophysiological markers are derived from the electrically-evoked compound action potential or “eCAP,” a measure of the synchronized response of the auditory nerve to electrical stimulation that can be recorded using the electrodes of the CI itself that is used in clinical environments (Abbas et al., 1999). While one goal of these techniques is to predict speech-perception outcomes across different CI patients, such measures could also inform strategies to optimize CI stimulation by focusing current or deactivating electrodes for individual patients, although previous approaches have had varying degrees of success in improving performance on speech-perception tasks (Carlyon and Goehring, 2021).

An eCAP-derived metric, the “Failure Index” (FI), has been proposed for predicting spiral ganglion nerve (SGN) survival, a key aspect of neural health that has been quantified in animal studies (Prado-Guitierrez et al., 2006). The FI is constructed from the ratio of the stimulation current level required to evoke the maximum eCAP amplitude divided by the corresponding eCAP amplitude (Konerding et al., 2025). The eCAP amplitudes are calculated from recorded electrical potential waveforms by subtracting the first negative peak from the subsequent positive peak. The maximum eCAP response amplitude is determined by recording an amplitude growth function (AGF), whereby stimulation current levels range from below the minimum level required to elicit an eCAP (threshold) up to the point of response amplitude saturation, and fitting a sigmoid curve to characterize amplitude growth and saturation amplitude (as per Ramekers et al., 2014).

Konerding et al. showed that the FI could reveal localized lesions of the auditory nerve in deafened guinea pigs implanted with 6-electrode CIs. The FI identified the location of the lesion within approximately one electrode. The authors further showed that the magnitude of the FI was positively correlated with the size of the lesion across guinea pigs and suggested translational potential to characterize neural health in human CI patients. Importantly, they also suggested that the normalization by input current makes the FI less susceptible to non-neural factors (such as the electrode-modiolus distance, EMD) than eCAP-derived metrics that rely solely on response amplitude.

One potential barrier for the FI to translation into the clinic is the necessity to reach the upper saturation point of the AGF. While it is possible to record eCAPs at sufficiently high stimulation levels to achieve eCAP saturation in guinea pigs, it is rarely done so in humans, as this often requires stimulation levels above the maximum comfortable loudness level for the patient. Konerding et al. show that in the animal model, the value of the FI is preserved with at least 90% accuracy provided the AGF is sampled at up to 70% of its dynamic range. For humans, a possible recording point on the AGF is at the maximum comfortable level (MCL). This would not guarantee reaching 70% of the eCAP's dynamic range, but may still be highly reflective of the FI calculated at the saturation point. Throughout this study, we use the FI measured at MCL, as this is the closest approximation to the saturation point that is feasible with awake human CI users.

Here, we present a re-analysis of electrophysiological recordings from previous studies to address two questions: (1) can the FI calculated at MCL (FI_MCL) be used to identify simulated “dead” regions of the auditory nerve in human CI users? (2) Is the FI_MCL more independent from non-neural factors compared to another metric proposed to predict neural health, namely, the neural-responsiveness estimate from the panoramic electrically-evoked compound action potential (PECAP) method described by Garcia et al. (2021)? The eCAP amplitude at MCL (eCAP_MCL; Scheperle, 2017) is also included as an exploratory factor.

In humans, while we of course cannot replicate the methods used in animal models and create localized lesions, we have shown that one can simulate different conditions in neural health. Garcia et al. simulated localized neural “dead” regions by presenting two pre-pulses on a selected target electrode prior to any eCAP recordings. This lowers the responsiveness of a locally constrained neural population during the eCAP recordings by putting them into a refractory state. We refer to this manipulation as temporarily creating a locally “adapted” region of auditory nerves. PECAP was used to estimate variation in neural responsiveness along the length of the cochlea, both in the presence and absence of pre-pulse stimulation. Briefly, PECAP requires eCAP amplitudes recorded at MCL using all pairwise combinations of electrodes in a masker-probe stimulation sequence (employing the forward-masking artefact-reduction method as described by Abbas et al., 1999). Based on the assumption that eCAPs reflect the response of neurons commonly activated by both masker and probe electrodes, an algorithm estimates the respective contributions of neural responsiveness and current spread to the underlying neural-activation patterns centered at each electrode (see Garcia et al., 2021). PECAP predicted reduced neural responsiveness near the target electrode in the adapted condition compared to a control condition with no pre-pulses, and importantly, the estimate of current spread remained unchanged. The data from this study were re-analyzed to calculate the FI_MCL for each electrode in both the adapted and control conditions to determine the metric's ability to detect the simulated regions of reduced neural responsiveness, addressing our first question.

In another study, a large PECAP dataset was collected from users of cochlear devices in which EMDs were calculated from computerized tomography (CT) scans (Garcia and Carlyon, 2025). The data from this second study was re-analyzed to calculate the FI_MCL and assess its independence from EMD, especially in comparison to the neural-responsiveness estimate of PECAP (as well as to the eCAP_MCL), addressing our second question.

We hypothesized that all three metrics would identify locally “adapted” regions of auditory neurons, but that they would vary in their independence from EMDs.

2. Materials and methods

2.1. Experiment 1: Detecting localized areas of reduced neural responsiveness

Experiment 1 consists of a re-analysis of published data from Garcia et al. (2021). This study involved recording eCAPs in response to the highest stimulation current level reported as MCL for each electrode activated in the study participants' clinical programming strategy. Although additional eCAPs were recorded using the forward-masking artefact cancellation technique for every combination of masker and probe electrodes, only the eCAPs recorded with masker and probe pulses on the same electrodes were required to calculate the FI. The study included seven users of Cochlear Ltd. devices and consisted of two conditions: (A) a control condition and (B) a condition simulating a neural “dead” region achieved by presenting two pulses on a fixed target electrode prior to delivering the masker and probe pulses on each electrode along the array (referred to hereafter as the “adapted” condition). Methodological details can be found in Garcia et al. (2021). PECAP successfully characterized the manipulation as a reduction in neural responsiveness localized at the area surrounding the target electrode in all seven participants (Garcia et al., 2021).

The eCAP amplitudes at MCL across all active electrodes in both conditions were re-analyzed using an adjusted version of the FI, the FI_MCL. As the phase duration of electrical pulses used for eCAP stimulation varied across participants in the dataset, the FI_MCL is calculated here using the input charge level instead of the input current level. To compensate for larger between-participant differences observed in humans than in animal models, it was also calculated on a logarithmic scale. Compared to the formula used by Konerding et al. [Eq. (1)], these two precautions led to a new formula [Eq. (2)] for calculating the FI_MCL used throughout the present study,

$$FI=\frac{\mathit{\text{stimulation}}\hspace{0.17em}\mathit{\text{current}}\hspace{0.17em}\left(\mu A\right)}{\mathit{\text{maximum}}\hspace{0.17em}\mathit{ECAP}\hspace{0.17em}\mathit{\text{amplitude}}\hspace{0.17em}\left(\mu V\right)},$$ (1)

$${FI}_{\mathit{MCL}}=20*{\mathrm{\text{log}}}_{10}\left(\frac{\mathit{\text{stimulation}}\hspace{0.17em}\mathit{\text{current}}\hspace{0.17em}\left(\mu A\right)\hspace{0.17em}*\hspace{0.17em}\mathit{\text{phase}}\hspace{0.17em}\mathit{\text{duration}}\hspace{0.17em}\left(\mu s\right)}{\mathit{ECAP}\hspace{0.17em}\mathit{\text{amplitude}}\hspace{0.17em}at\hspace{0.17em}\mathit{MCL}\hspace{0.17em}\left(\mu V\right)}\right).$$ (2)

The signed difference FI_MCL(Adapted) − FI_MCL(Control) between the FI_MCL in the two conditions was calculated for each available electrode (positive values indicating identification of the locally adapted region). A paired t-test was performed on this metric for the target electrode and for the two adjacent electrodes in each direction, consistent with the definition of the adapted regions in Garcia et al. A similar process was repeated for PECAP's neural-responsiveness estimate (η) for comparison, with the difference that the signed difference was calculated as η(control) – η(adapted) to ensure that successful identification of localized reductions would consistently manifest as positive differences. Finally, the same process that was applied to PECAP's neural-responsiveness estimate was additionally applied to the eCAP amplitude at MCL (eCAP_MCL).

2.2. Experiment 2: Comparisons to a non-neural factor of the electrode-neuron interface

Experiment 2 consists of a re-analysis of published data from Garcia and Carlyon (2025). This study involved a large dataset of PECAP data from multiple centers. For 41 participants included in the study who were users of Cochlear devices from two of the participating centers, the CT scans were also available and used to calculate EMDs. These analyses were performed by EIORL Antwerp, following the approach described in Sismono et al. (2022). Briefly, the modiolar wall and electrode contacts were semi-automatically segmented from the reconstructed CT volumes, and the EMD was then measured from the center of the electrode contact surfaces to the nearest point on the Modiolar wall. A single ear was assessed from each participant. The FI_MCL was calculated for all clinically-used electrodes for each of these participants, and an across-electrode Pearson correlation was computed with the associated EMD. PECAP's neural-responsiveness estimate was also correlated to the EMD, and the two correlations were compared using a t-statistic for comparing dependent correlations (Steiger, 1980). The correlation between eCAP_MCL and EMD was also tested for significance and compared to the two other correlations for exploratory comparison. A significantly weaker correlation between the EMD and one of two metrics would provide evidence to support this metric's greater independence from non-neural attributes of the electrode-neuron interface compared to the other.

3. Results

3.1. Experiment 1: Detecting localized areas of reduced neural responsiveness

Figure 1 shows the signed differences between the control and adapted conditions for PECAP's neural-responsiveness estimate, the FI_MCL, and the eCAP_MCL relative to the target electrode for all seven participants who were in the original dataset. The FI_MCL was calculated according to Eq. (2).

Fig. 1.

Signed difference between the control and adapted conditions for PECAP's neural-responsiveness estimate (left), the eCAP amplitudes (middle), and the FI (right) for all electrodes. Each electrode is numbered regarding the location of the pre-pulses (target electrode, red vertical dashed line). Only electrodes with data from at least 5 participants are shown. Thin light gray lines represent the seven individual participants; the solid black line represents the across-participant means, and the shaded regions represent ± 1 standard deviation. Asterisks (^*) indicate electrodes for which the metric's signed difference metric differed significantly from zero across participants. (^*p < α, α = 0.05/5 = 0.01.)

Paired t-tests across participants were conducted on the signed difference metrics for five adjacent electrodes: two apical through two basal relative to the target electrode. For PECAP's neural-responsiveness estimate (η), the target electrode and each electrode on either side of the target electrode showed significantly positive signed differences after corrections for multiple (15) comparisons using the Bonferroni-Dunn method (Dunn, 1961) (α = 0.05/15 = 0.0033, p = [0.153, 0.0024^*, 0.0017^*, 0.0019^*, 0.0207], t(6) = [1.63, 5.01, 5.40, 5.27, 3.12], from apical to basal). For the FI_MCL, the target electrode and the adjacent basal electrode showed significantly positive signed differences (α = 0.05/15 = 0.0033, p = [0.402, 0.0061°0.0007^*, 0.0004^* 0.0051], t(6) = [2.61, 4.13, 6.35, 7.02, 4.29], from apical to basal). For the eCAP_MCL (on a dB scale), all five tested electrodes showed significantly positive signed differences (α = 0.05/15 = 0.0033, p = [0.0002^*, <0.0001^*, <0.0001^*, <0.0001^*, <0.0001^*], t(6) = [8.20, 13.5, 18.3, 16.5, 11.4], from apical to basal). These results demonstrate that all three metrics are sensitive to a manipulation that simulates a loss of neural responsiveness.

3.2. Experiment 2: Comparisons to a non-neural factor of the electrode-neuron interface

The FI_MCL was calculated for each electrode for each of the 41 participants for whom EMDs were also available. In some cases, no eCAP response was observed despite recording-parameter optimization. To reflect the fact that these cases still constitute a failure of stimulating current to elicit a neural response, the corresponding FI_MCL was calculated with an eCAP amplitude equal to the noise floor for the eCAPs across all electrodes of the participant in question. Patient-specific rather than manufacturer-defined noise-floor levels were used due to heterogeneity in observed noise floors between participants. This was required for 79 electrodes over 17 different participants (9.2% of all electrodes in the dataset), and the noise floors ranged from 0.34 to 18.44 μV with a median of 6.48 μV.

Figure 2 shows the across-electrode correlations between EMD and the three metrics: (1) PECAP's neural-responsiveness estimate (η), (2) the eCAP_MCL (in dB re 1 μV), and (3) the FI_MCL. Between-participant differences were removed from each metric by subtracting the participant-specific means across electrodes from each individual electrode value for that participant. This reduces the degrees of freedom (d.f.) of the correlation by N, but avoids the common statistical pitfall of pooling multiple sources of variance, and is mathematically equivalent to performing an analysis of covariance with participant as a random factor (Bland and Altman, 1995). Three correlations are described here, but as Garcia and Carlyon (2025) presented a correlation between EMD and the PECAP algorithm's current-spread estimate [r = −0.097, p(uncorrected) = 0.004], the correlations shown here are corrected for four comparisons (α = 0.05/4 = 0.0125).

Fig. 2.

Across-electrode correlations between EMD and (left) PECAP's neural-responsiveness estimate (η), (middle) eCAP Amplitude (in dB re 1 μV) at MCL, and (right) the FI at MCL (in dB). The thick black lines represent significant correlations, and each individual data point represents a single electrode.

The Pearson's correlation between EMD and PECAP's neural-responsiveness estimate was not significant [r(819) = −0.070, p = 0.04, α = 0.0125], whereas it was significant between EMD and the eCAP_MCL [r(819) = −0.166, p < 0.0001^*, α = 0.0125], and between the EMD and FI_MCL [r(819) = 0.208, p < 0.0001^*, α = 0.0125]. A t-statistic difference test for comparing dependent correlations (Lee and Preacher, 2013; Steiger, 1980) showed a significantly stronger correlation to the EMD of the eCAP_MCL compared to PECAP's η [t_diff(817) = 3.191, α = 0.05/3 = 0.0167, p = 0.0014^*], a significantly smaller magnitude of the correlation to the EMD of the eCAP_MCL compared to the FI_MCL [t_diff(817) = −5.089, α = 0.0167, p < 0.0001^*], and a significantly smaller magnitude of the correlation to the EMD of PECAP's η compared to the FI_MCL [t_diff(817) = −4.403, α = 0.0167, p < 0.0001^*]. The magnitude was used for the latter two comparisons due to the expectation that the FI_MCL would correlate positively with EMD, whereas the other two metrics would relate negatively. This analysis revealed that of the three evaluated metrics, the FI_MCL showed the strongest correlation with EMD, and PECAP's η showed the weakest (and non-significant) one.

4. Discussion

We demonstrated that the FI_MCL, an adapted version of the metric presented by Konerding et al. (2025), is sensitive to regions of reduced neural responsiveness in awake humans, as previously shown in guinea pigs. It may be possible to improve the predictive power of the FI in humans by recording AGFs intra-operatively, as higher stimulation levels can be achieved under anesthesia without causing discomfort, and the eCAP saturation is therefore more likely to be reached.

However, we do not find evidence that the FI_MCL is more independent from the EMD than the other evaluated eCAP-based metrics. Both the FI_MCL and the eCAP_MCL were correlated with the EMD, while PECAP's η was not. Further tests revealed that the FI_MCL was significantly less isolated from EMD than both PECAP's η and the eCAP_MCL. PECAP's η was furthermore more isolated from EMD than the eCAP_MCL. Although it should be noted that these results are observed at the group level and may not generalize to individual ears, they nevertheless suggest that neither the FI_MCL nor the eCAP_MCL fully isolate neural from non-neural factors, while the PECAP algorithm may do this most successfully of the three evaluated metrics due to its intrinsic mechanism of separating factors of neural responsiveness and current spread. Although we have not found positive correlations between the PECAP's current-spread estimate (σ) and EMD across electrodes (Garcia et al., 2021; Garcia and Carlyon, 2025), it does correlate positively across participants (Garcia and Carlyon, 2025), suggesting some shared variation. The PECAP's current-spread estimate may therefore absorb some of the variation in eCAP amplitudes that is dependent on EMD, reducing contamination of the neural-responsiveness estimate. While not independent from non-neural factors, the FI_MCL may still hold value as a single-point measure that collectively characterizes multiple aspects of the electrode-neuron interface. Given its simple calculation from eCAP measures already routinely recorded in many clinical centers, it has few barriers to translation and could even be calculated retrospectively for many CI patients. However, when the goal is to fully isolate neural from non-neural attributes of the electrode-neuron interface, the PECAP algorithm seems a better candidate as it is less affected by EMD than both the FI_MCL and the eCAP_MCL.

Future work related to the FI_MCL could investigate practical aspects of applying the metric to the clinic. Konerding et al. calculated a FI threshold above which the presence of a lesion in guinea pigs is likely, but this calculation is unlikely to translate from the highly controlled animal models directly to humans, especially given the large amount of variability in eCAP amplitudes observed between CI users and between different electrode locations within individual patients' electrode arrays (Eisen and Franck, 2004; He et al., 2017). We suggest that transforming the FI to a logarithmic scale, using input charge instead of input current [as described in Eq. (2)], and using the noise floor for undetectable eCAPs can all reduce the sensitivity of the metric to low/absent eCAPs often observed in humans.

Another point to consider is the mixed literature on the relevance of eCAPs for perception (de Vos et al., 2018; McKay et al., 2013), and the large variability in eCAP amplitudes observed at the same perceived loudness level for different electrodes within the same participant (Jeon et al., 2010). It is possible that an FI-like metric considering the input stimulation level and the output response strength would be more clinically applicable if it captured a more central response such as a brainstem [i.e., electrically evoked auditory brainstem responses (eABRs)] or a cortical response [i.e., cortical auditory evoked responses (CAEPs)], although this of course is speculative and should also be investigated. CAEPs in particular have been shown to correlate strongly across users to behavioral measures of loudness (Mao et al., 2019), in contrast to the mixed findings within the eCAP literature.

5. Conclusion

Like the PECAP method, an adapted version of the FI_MCL could identify local regions of reduced neural responsiveness in human CI users. However, PECAP's estimate of neural responsiveness was less influenced by the non-neural factor of the electrode-modiolar distance calculated from CT scans than either the FI_MCL or the eCAP amplitude at MCL. The FI_MCL is a simple metric calculated from commonly-recorded eCAP measures in clinical settings that characterizes multiple aspects of the electrode-neuron interface. The PECAP method is more complex, but may be better at isolating neural from non-neural factors of the electrode-neuron interface.

Author Declarations

Conflict of Interest

F.S. is nominally funded by EIROL Antwerp, but Cochlear funded the EMD analyses specifically leveraged for this article. The other authors have no conflicts of interest to disclose.

Ethics Approval

Ethical approval to conduct the study was granted by the National Research Ethics Committee for the East of England [Integrated Research Application System (IRAS) No. 118230]. Informed consent was obtained from all research participants prior to taking part in the study.

Data Availability

The data corresponding to Experiment 1 can be found here in association with this article https://osf.io/pae78/.

The authors are not at liberty to share the data corresponding to experiment #2 as it was made available from collaborators at different institutions. The data may be made available upon reasonable request from the corresponding author after obtaining sufficient permissions from the relevant parties.

The authors retain a CC BY-NC-SA license to all shared data associated with this article.