Hidden hearing loss and efficient information representation in the mammalian auditory midbrain
Hidden hearing-loss (HHL) is a term that encompasses hearing difficulties in noisy environments, that do not imply changes in hearing sensitivity (thresholds), hence, currently unobservable in clinical diagnosis such as pure-tone audiometry. It has been shown in small mammals that a single dose of loud noise-exposure for two hours damages permanently the first stage of afferent synapses in the cochlea, whilst leaving thresholds unchanged (a lesion also known as cochlear synaptopathy), affecting how sound is transduced, and potentially, impairing neural adaptation processes throughout the auditory pathway.
Measuring the impact of this lesion on neural adaptation and perception remains a central problem in hearing research, and even though recent biophysical modelling can predict anatomical and electrophysiological consequences, further development is needed to understand and quantify its repercussions. This thesis addresses the gap by proposing the use of normative brain theories, such as efficient coding, to assess the level of optimality achieved by neural populations in the midbrain of small mammals, unifying observed phenomena triggered by this lesion in a single computational framework. To this end, this work assumes that the auditory midbrain fulfils a specific signal-processing function of adaptively quantising the sound intensity dimension, just as modern analogue-to-digital converters work to efficiently transduce signals in general, and particularly, sound.
The framework is tested using three single- and multi-unit spike-train datasets from the auditory midbrain of gerbils and mice, consisting of responses to a stimulus used in previous studies to test neural adaptation to changing sound intensity contexts. Two datasets are composed by noise-exposed animals (NE; HHL group) and sham-exposed controls, whilst the third dataset by healthy animals that were ear-plugged for two weeks (conductive hearing loss; CHL), recorded before and after the removal of the earplug. To compare these groups, each of them is associated to an optimal solution space via maximum a-posteriori estimation, which allows to make observations among groups in three dimensions relevant to hearing loss phenomena: firing-rate dynamics (neural gain), coding utility (optimality) and entropy.
Observations made through this framework reproduce and unify recently reported adaptation changes of hearing loss: 1) neural gain is up-regulated by the loss (HHL) or attenuation (CHL) of afferent inputs; 2) gain up-regulation is detrimental in general to neural coding utility, which systematically increases the associated entropy as well; 3) reafferentiation processes due to synaptic damage make undamaged auditory fibres achieve higher coding utility for the intensity contexts they encode (quiet contexts), down-regulating gain. The consistency of these results suggest that the employed framework might be a solid starting point to construct a measure of hearing-loss stemming from computational neuroscience principles, and simultaneously, support a novel perspective of the auditory midbrain as a well-characterised device, an adaptive and efficient analogue-to-digital converter.