CNNs improve decoding of selective attention to speech in cochlear implant users

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Speech comprehension in noise is challenging for cochlear implant (CI) users.
• Auditory attention decoding (AAD) uses electroencephalography (EEG) to identify target speech, guiding CI devices.
• Deep neural networks (DNNs) show promise for improving AAD in normal-hearing individuals.

Purpose of the Study:

• To evaluate DNNs for enhancing AAD in bilateral CI users.
• To establish DNNs as a state-of-the-art approach for neuro-steered CIs.
• To assess the impact of CI artifact removal on AAD accuracy.

Main Methods:

• Collected EEG data on selective auditory attention from 25 bilateral CI users.
• Implemented and compared linear models and convolutional neural networks (CNNs) for AAD.
• Developed a novel CI artifact removal strategy and used support vector machine (SVM) for speaker classification.

Main Results:

• CNNs outperformed linear models in AAD across decision window sizes (1-60s).
• CI artifact removal led to a modest improvement in CNN decoding accuracy.
• The CNN decoder achieved a peak population-level accuracy of 74% with SVM classification for a 60s window.

Conclusions:

• CNN-based decoding shows superior potential for neuro-steered CIs.
• This technology could significantly improve speech perception for CI users in complex listening situations.
• The findings support the advancement of brain-computer interfaces for hearing assistance.