Machine Learning Recognizes Frequency-Following Responses in American Adults: Effects of Reference Spectrogram and...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Electrophysiology studies brain responses to sound.
• Frequency-following response (FFR) is a key measure of auditory encoding.
• Limitations exist in current FFR recording and analysis.

Purpose of the Study:

• To enhance frequency-following responses (FFRs) using an improved source-separation machine learning algorithm.
• Investigate the efficacy of a specific machine learning algorithm (SSNMF) for FFR enhancement.
• Assess the impact of different training references on FFR quality.

Main Methods:

• Recruited 28 normal-hearing native English speakers.
• Recorded electroencephalographic (EEG) signals during auditory stimulation (/i/ and /da/ tokens).
• Applied a source-separation non-negative matrix factorization (SSNMF) algorithm, trained with individual, grand-averaged, or stimulus token spectrograms.

Main Results:

• FFRs were significantly enhanced (p < .001) when SSNMF was trained using individual and grand-averaged spectrograms.
• No significant enhancement was observed when training with stimulus token spectrograms.
• Similar enhancement patterns were found for both /i/ and /da/ stimulus tokens.

Conclusions:

• The SSNMF machine learning algorithm effectively enhances FFRs when trained appropriately.
• Training with individual and grand-averaged spectrograms is crucial for FFR improvement.
• This advancement holds promise for improving FFR obtainment, analysis, and clinical utility.