acceptodds
Under review as a conference paper at ICLR 2027

Disentangling Device and Neural Contributions to Spatial Hearing Deficits in Cochlear Implants

Abstract

Spatial hearing is essential for spatial orientation, auditory scene analysis, and speech perception. Cochlear implant (CI) users localize sounds poorly, and these deficits can arise from both device-related factors that distort spatial cues and from neural degeneration in the peripheral and central auditory pathways. Because these factors are typically intertwined, their individual contributions to localization deficits are difficult to isolate. To disentangle these contributions, we introduce Hybrid-CINN, a model that combines a biophysical CI simulation with a convolutional neural network trained to localize natural sounds and rendering spatial signals with behind-the-ear (BTE) head-related transfer functions (HRTFs) to reflect realistic CI listening. Our model, trained under ideal conditions reproduces key characteristics of CI sound localization effects, including lateral response compression and a strong reliance on interaural level-difference (ILD) cues. We then introduce device-related factors and neural degeneration individually, retraining the network under each condition to simulate chronic adaptation. We find that electrode insertion mismatch degrades localization performance, while linking bilateral automatic gain control improves it; and modeled central degeneration has a greater impact on localization than peripheral degeneration. Together, these results establish Hybrid-CINN as a testbed for disentangling the factors that limit spatial hearing in CI users and for evaluating CI processing and fitting strategies before they reach patients.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuro-AI, Cochlear Implant, Behavior Modeling, Sound Localization, ML for Health, Bio-medical, Hearing Research

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1xuLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how device settings and simulated neural loss affect sound localization with bilateral cochlear implants. Hybrid-CINN combines natural sounds spatialized using behind-the-ear head-related transfer functions (HRTFs), a stochastic CI and auditory-nerve simulation, and a convolutional localization decoder. Baseline evaluation uses held-out sounds, simulated rooms, and HRTFs; factor experiments use reduced training data and one held-out HRTF. Without fitting human localization responses, the model exhibits lateral response compression, stronger reliance on interaural level differences than interaural time differences, and poorer low-frequency localization, while achieving lower error than the compared CI cohort. Separate decoders are trained from scratch for each device or neural condition. Within the tested configurations, unilateral insertion shifts increase localization error, linked automatic gain control provides a small improvement, and asymmetric or basal fiber loss is more detrimental than symmetric or apical loss, respectively. At matched nonzero nominal deletion fractions, silencing units across multiple decoder stages produces larger errors than deleting peripheral fibers.

Critical strengths
  1. Controlled comparisons of peripheral loss patterns
    The interaural-asymmetry experiment holds total fiber loss constant, and the basal-versus-apical experiment removes equal numbers of fibers (Fig. 4). These controls show that, within Hybrid-CINN, the distribution of surviving inputs across ears and cochlear regions matters beyond their total number. This is a useful contribution because it yields specific predictions about the loss patterns that impair localization, extending the analysis beyond overall neural survival.

  2. Qualitative behavioral agreement without fitting human responses
    The model reproduces lateral response compression, ILD-dominant cue use, and poorer low-frequency localization without fitting human localization responses (Fig. 2). Baseline testing uses held-out sounds, rooms, and HRTFs. Although its overall localization error is lower than that of the compared CI cohort, agreement across these qualitative signatures provides a useful behavioral check before the framework is used to study controlled perturbations.

Critical weaknesses
  1. The central versus peripheral interpretation depends on the lesion design
    Figure 4B compares deletion of auditory-nerve fibers at the input with silencing channels and hidden units across multiple decoder stages, with inverted scaling of surviving units (Appendix A.8). Equal deletion fractions do not establish comparable losses of information or computational capacity. The authors acknowledge the information mismatch in Section 4.3, but the explanation that the decoder has less spare capacity remains dependent on this architecture and lesion topology. Appendix A.6 describes alternative architectures' intact performance without reporting corresponding lesion responses. The observed ranking is therefore supported for the tested interventions, while a broader capacity interpretation requires additional evidence. This distinction matters because the central versus peripheral comparison is a principal result; the conclusions should consistently reflect its operational scope.

  2. Generalization of the main factor effects across HRTFs is not established
    The baseline evaluation uses five leave-one-HRTF-out folds, but the device and neural-factor experiments use reduced training data and a single held-out test HRTF, b_ci8 (Appendix A.5). Repeated runs quantify variation across training runs and, for neural lesions, lesion masks within this setting; they do not establish that effect sizes or condition rankings are stable across HRTFs. These comparisons provide much of the paper's added scientific contribution, so their robustness across HRTFs materially affects the strength of the conclusions. The results remain informative within the tested configuration. Selected key contrasts across the available folds would strengthen the evidence, with the reduced training regime retained as an explicit qualification.

Questions for authors
  1. Clarification of the central versus peripheral interpretation. Section 4.3 suggests that the decoder has less spare capacity, while acknowledging that lesions are matched by fraction rather than information. Is this interpretation intended to be limited to the current CNN and lesion design? If a broader claim is intended, what evidence supports it? Do the alternative-architecture evaluations include lesion responses beyond the intact results reported in Appendix A.6?
  2. Additional evidence for the main factor claims. Appendix A.5 states that a five-fold evaluation is planned. Can the authors provide selected central/peripheral, basal/apical, and asymmetric/symmetric contrasts across the available held-out HRTFs, expressed relative to each fold's own intact control? Consistency in the directions and magnitudes of these effects would materially affect my assessment; a full replication of every dose curve is not necessary.
Reviewer_9tFxOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

A new framework called Hybrid-CINN was proposed, which couples a biophysical bilateral-CI front end to a CNN decoder trained to localize natural sounds, and uses a "lesion-and-retrain" protocol to attribute localization deficits to device versus neural factors. Sounds are rendered with BTE HRTFs and simulated rooms, passed through a stochastic auditory-nerve model, and decoded by a small depthwise-separable CNN. This computationally designed "ideal" model reproduces lateral error growth, lateral response compression, ILD-over-ITD reliance, and poor low-frequency localization. Each factor is then degraded one at a time and the decoder retrained from scratch: interaural insertion mismatch, AGC linking, peripheral fiber loss, central unit loss, asymmetric loss, and basal vs. apical loss. The authors conclude that the implant signal itself limits localization, that high-frequency level cues and interaural symmetry are the key determinants, and that the framework can serve as a testbed for fitting strategies.

Critical strengths
  1. A well-motivated and elegant experimental concept. Replacing the hand-designed read-out of mechanistic CI models with a task-optimized decoder separates "information available in the signal" from "ability to extract it". Retraining under each impairment, so that costs reflect what adaptation cannot compensate, is a thoughtful way to approach a question that cannot be answered in patients, where factors are intertwined and neural survival cannot be measured. Held-out rooms and HRTFs, and the fact that the model never sees human responses, make the qualitative agreement (lateral compression, ILD reliance, low-frequency penalty) a meaningful, non-trivial result rather than a fit.
  2. Breadth and care of the factor experiments. The paper sweeps device and neural factors with many seeds, reports full dose response curves with piecewise-linear fits and bootstrap CIs, tests combined factors for additivity, and documents the pipeline thoroughly (appendices, parameters, compute). The use of a biophysical CI model for in-silico experiments is a creative route to hypotheses (e.g., place- and ear-specific nerve health, interaural place alignment) that are actionable for CI fitting and testable in the clinic.
Critical weaknesses

1. The key attributions seems not yet clearly isolated from design choices, some conclusions may reflect the setup rather than the underlying factor.

  • Connection between CNN and the human brain

    • Expanding more on the modifications in CNN to clarify any analogy with the human brain's auditory signal processing would be needed.
    • For example, it is not argued why silencing decoder units should correspond to central auditory degeneration (e.g., brainstem vs. cortex), or what evidence suggests this CNN's internal units correspond to neural populations. There are lots of explanations on the 'effect of silencing' in human auditory pathway, but I believe it does not necessarily explain the properties of silencing the decoder units.
  • Peripheral vs. central loss (Sec. 4.3, Fig. 4B).

    • "Central degeneration was more damaging" is compared at matched fractions of units, but units in a 43k-parameter decoder (first layer: 12 channels) and fibers in a 980-fiber input are not comparable in redundancy. The authors note this, but the abstract states the conclusion without that caveat.
  • Basal vs. apical loss (Fig. 4D). By my reading of A.4/A.8, the 12 electrodes span 5–26 mm and the apical half of the fibers spans 17.5–35 mm, so roughly half of the "apical half" lies apical to the last electrode and is barely stimulated, while the basal half contains 7 of 12 electrodes (vs. 5). The 2.7° gap may therefore partly reflect how many electrode-driven fibers (or spikes) were removed rather than frequency region. Matching by spike count or electrode coverage would clarify this.

  • The baseline is not decomposed.

    • The ideal model is ~3x worse than normal hearing, and the paper attributes this to "the implant signal", but no reference run (e.g., the same decoder on a normal-hearing periphery with BTE HRTFs, and with ear-canal/KEMAR HRTFs) quantifies how much of the 16.6° comes from BTE microphone placement, the front-end settings (pre-emphasis, 298–6679 Hz filterbank, CIS), or the nerve model.
    • Since BTE rendering is presented as a contribution relative to Banerjee et al. (2025), an ablation showing it matters would strengthen the paper.
    • Likewise, the excitation-inhibition reference decoder is described (Sec. 3.2, A.6) but its results are never reported, and the alternative architectures are summarized without numbers. This makes the motivating claim, that a fixed read-out confounds the analysis, not to be sufficiently demonstrated.
Questions for authors

Clarification requests (answers could change my assessment):

  1. What are the RMSEs of the excitation–inhibition decoder and of the alternative architectures (S4, RNN, larger CNN)? Why are the ear-swap check (A.7) and the 50%/90% partial-lesion results not reported?

Suggestions (would strengthen the paper but are not required to evaluate it):

  1. Please report the factor experiments on all five HRTF folds with the full corpus, and add listener-level uncertainty (and a multiple-comparison treatment) to the factor results.
  2. Adding a baseline decomposition would be nicer: NH-periphery decoder with BTE vs. ear-canal/KEMAR HRTFs, to quantify how much of the 3× deficit stems from BTE rendering vs. the CI front end.
1 comment

I read the rebuttal made by the authors, including General Response and the other responses to all reviewers.

According to the authors' promises, most of my concerns (will) have been resolved. I am adjusting my rating from the original score 'Weak acceptance' to 'Clear acceptance'.

Reviewer_pNgJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hybrid-CINN, which couples a biophysical bilateral cochlear implant (CI) and auditory-nerve simulation with a CNN trained to localize natural sounds rendered with HRTFs, in order to disentangle device-related and neural contributions to CI localization deficits. The ILD dominance appears to be largely determined by the front-end design rather than emerging from the model.

Critical strengths
  • The paper addresses a clinically relevant question that is difficult to study in patients.
  • The modular framework allows each device and neural factor to be varied in a controlled way.
  • The experiments are carefully designed, with held-out rooms and HRTFs, and the limitations are discussed transparently.
Critical weaknesses
  • I am not fully convinced about the novelty for ICLR. What is the differences between this work and the framework in Banerjee et al. (2025)? It seems that the methodological contribution to machine learning is limited.
  • Using unit dropout in a small CNN to simulate the central degeneration seems like a loose analogy to me. It mainly reduces network capacity. The conclusion that central loss is more harmful than peripheral loss may reflect the small decoder rather than biology, and the two sweeps remove different kinds of units, so the comparison is not well matched.
  • The comparison with human data is qualitative, which the authors have also acknowledged. With only five HRTFs, the model is evaluated with different stimuli from the human studies, it's hard to say how closely the model matches real CI listeners.
  • The factor experiments use a single fold and 10% of the data. It would be good to know if the conclusions hold up with more folds and the full dataset.
Questions for authors

Please see weakness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.