An adaptive gain model for predicting auditory brain activity in mice outperforms standard methods for predicting cortical speech tracking in human EEG

Abstract

Human brain activity tracks slow fluctuations in continuous sound, including speech, and modelling this relationship provides insight into how the brain encodes naturalistic auditory input. Traditional approaches to modelling brain tracking of speech with linear regression often use the amplitude envelope of the stimulus as a regressor, implicitly assuming that neural responses scale linearly with sound intensity. However, previous studies of central auditory processing in both animals and humans have shown that adaptive mechanisms adjust auditory responses based on recent sound history. Here, we find that a simple nonlinear transformation of the stimulus envelope, derived from studies of mouse auditory thalamic responses to temporally varying sounds, significantly improves modelling of human cortical speech tracking. This Adaptive Gain transformation essentially normalizes sound level by its recent context. Improvements in modelling of cortical speech tracking using the Adaptive Gain stimulus representation are robust across experiments, participant groups, and languages, indicating that short-timescale adaptation to recent sound level is a general feature of auditory processing. We further show that the optimal adaptation time constant for human cortical responses to continuous speech is approximately 50-100 ms, longer than previously observed for auditory thalamic responses to temporally varying sounds in mice. In summary, the Adaptive Gain transformation is a mathematically simple alternative to the standard envelope representation that captures dynamic adaptation in auditory processing and reliably improves prediction of human cortical EEG signals during listening to continuous speech. Significance StatementStandard models of human brain responses to continuous speech use the raw stimulus envelope as a predictor of EEG responses, implicitly assuming stationary sensitivity to sound intensity. Here, we use instead a simple nonlinear transformation of the stimulus envelope, Adaptive Gain, which normalizes sound level based on recent sound history over tens to hundreds of milliseconds. Originally derived from studies of auditory processing in mice, the Adaptive Gain transformation robustly improves prediction of EEG responses to continuous speech in humans across experiments, participant groups, and languages. Results demonstrate that dynamic adaptation to sound level is a core feature of auditory processing, and can be approximated in a mathematically simple form to improve prediction of cortical speech tracking.