Authors: Julie Van Heghe¹; Sarah Verhulst¹; Guy Torfs¹
¹Ghent University
Background: Cochlear implants (CIs) restore hearing by electrically stimulating the auditory nerve, yet current sound-processing strategies produce neural representations that differ substantially from those in normal hearing. This mismatch limits perceptual performance, particularly in complex listening environments. Approaches that directly target neural response similarity may help bridge this gap.
Method: We propose a closed-loop optimization framework that combines computational auditory models with deep learning to design CI stimulation strategies. The system simulates both normal-hearing and cochlear-implant auditory pathways, generating spiral ganglion neuron responses to the same acoustic input. Two neural networks replace the conventional sound-coding stage and jointly learn electrode activation patterns and current amplitudes through end-to-end training. Optimization minimizes the difference between normal-hearing and CI-evoked neural responses using differentiable approximations of auditory periphery models.
Results: In a proof-of-concept evaluation using speech stimuli, the learned stimulation strategy substantially reduced the mismatch between normal-hearing and CI neural response envelopes compared to the conventional Advanced Combination Encoder (ACE), achieving a 6.6-fold reduction in envelope error. While the current implementation does not yet fully account for all practical constraints of clinical CI devices, these results indicate improved alignment of simulated neural activity.
Conclusion: These findings demonstrate the feasibility of closed-loop, model-based optimization for cochlear implant processing. By integrating auditory neuroscience models with machine learning, the proposed framework offers a promising direction for developing next-generation CI strategies that more closely reproduce neural responses associated with normal hearing.
