Toward Decoding Selective Attention From Single-Trial EEG Data in Cochlear Implant Users.

Area of Science:

• Neuroscience
• Auditory Neuroscience
• Biomedical Engineering

Background:

• Decoding attended speech from electroencephalography (EEG) is possible in normal hearing (NH) listeners by reconstructing the speech envelope.
• The impact of reduced spectral resolution and cochlear implant (CI) electrical artifacts on speech decoding performance remains unclear.

Purpose of the Study:

• To investigate the effect of decreased spectral resolution and CI-induced artifacts on the accuracy of decoding attended speech from EEG.
• To compare speech decoding performance between NH listeners, NH listeners with simulated CI conditions, and bilateral CI users.

Main Methods:

• EEG data were recorded from NH listeners and bilateral CI users presented with dichotic speech stimuli.
• Speech envelope reconstruction was performed using decoders trained with a regularized least square estimation method.
• NH listeners underwent a vocoder condition to simulate CI spectral resolution limitations.

Main Results:

• Decoding accuracy was 80.9% in NH listeners with original sound files.
• Performance decreased to 73.2% in NH listeners using a vocoder and to 71.5% in CI users.
• A notable drop in decoding accuracy was observed with reduced spectral resolution.

Conclusions:

• Speech envelope reconstruction from EEG is feasible for decoding attended speech, even with simulated CI spectral degradation.
• While accuracy decreases, decoding attended speech remains possible in cochlear implant users, though more training data may be beneficial.
• This study highlights the challenges and possibilities of auditory attention decoding in individuals with impaired hearing.