A computational model to simulate spectral modulation and speech perception experiments of cochlear implant users

Area of Science:

• Auditory Neuroscience
• Biomedical Engineering
• Computational Audiology

Background:

• Cochlear implant (CI) users exhibit significant variability in speech understanding.
• This variability is linked to peripheral auditory system factors like electrode-nerve interface and neural health.
• Computational models offer a controlled environment to study CI performance differences.

Purpose of the Study:

• To evaluate performance differences between three HiRes Fidelity 120 (F120) sound coding strategy variants using a computational model.
• To investigate the impact of varying auditory nerve fiber (ANF) degeneration on speech understanding metrics.
• To compare sequential versus simultaneous stimulation in simulated CI scenarios.

Main Methods:

• A computational model incorporating a sound coding strategy, 3D electrode-nerve interface, ANF models, and feature extraction was developed.
• The Simulation Framework for Auditory Discrimination Experiments (FADE) was used as the simulation backend.
• Two speech-relevant experiments were conducted: spectral modulation threshold (SMT) and speech reception threshold (SRT), under healthy and degenerated ANF conditions.

Main Results:

• Worse neural health conditions generally led to worse predicted performance, though the effect was smaller than observed clinically.
• Speech reception threshold (SRT) performance with simultaneous stimulation (especially F120-T) was more negatively impacted by neural degeneration than sequential stimulation (F120-S).
• No significant performance differences were found in spectral modulation threshold (SMT) experiments across conditions.

Conclusions:

• The computational model can simulate SMT and SRT experiments but requires further refinement for reliable prediction of real CI user performance.
• Neural degeneration disproportionately affects speech reception thresholds under simultaneous stimulation strategies.
• Future improvements should focus on the ANF model, feature extraction, and predictor algorithms.