Informational-effects quantification of ‘hidden hearing loss’ using the information bottleneck method

Cochlear synaptopathy (CS) is a lesion characterised by abnormal synaptic activity between inner hair cells and auditory nerve fibres. Its quantification and specific relationship to hearing loss is unclear, but likely relates to problems listening in noise, with intact hearing thresholds, following exposure to loud sounds (so called ‘hidden hearing loss’, HHL). Here, we address this gap using the information bottleneck method (IBM). The IBM generates reduced representations of random variables, that are maximally informative about another (relevance) variable. Assuming the predictive coding hypothesis, the IBM has been used to yield a set of reduced representations for the time series of auditory oddball stimuli, enabling the construction of a theoretical prediction error, that is correlated to cortical neural responses. As such, neurons that exhibited significant correlations with the error signal were used to extract parameters of the internal representation (IR) such as, its complexity and predictive power.

We extend the IBM framework for HHL, to quantify and compare several IR’s parameters of noise induced HHL animals, when exposed to a stimulus composed of broadband noise with time-varyingrandom intensity. First, we express the stimuli capturing its time-varying component, and then derive a Bayesian estimator for the variance (energy) as a function of past information. Next, we obtain reduced representations of the estimator, and calculate a theoretical error signal that we correlate with neural recordings in the auditory midbrain, with the aim of testing the hypothesis that animals exposed to loud sounds exhibit IR with less predictive power or less complexity