High-resolution auditory restoration by a brain implant operated with a bio-mimetic deep encoder

Descripción del proyecto

Nuevo implante cerebral para mejorar el sonido

La pérdida de audición es un importante problema de salud. Los implantes cocleares son eficaces en casos graves, pero tienen limitaciones. Para mejorar la interacción del sonido en el cerebro, se están desarrollando nuevos implantes de corteza auditiva que requieren algoritmos de codificación innovadores. Un proyecto anterior demuestra que estas transformaciones pueden reproducirse utilizando redes neuronales profundas. El proyecto BRAINCODER, financiado por el Consejo Europeo de Investigación, pretende crear una red que comprima los datos sonoros para el conjunto de estimuladores del implante siguiendo las reglas de codificación del sonido. Asimismo, el proyecto permitirá estimar con precisión la pérdida de información y comparar los implantes corticales con los cocleares. También conseguirá una patente, desarrollará un sistema de estimulación cortical de alta densidad, construirá un demostrador de «software» y establecerá un plan de negocio para poner en marcha una empresa emergente que aproveche estas innovaciones.

Objetivo

Hearing loss is a leading cause of disease burden. While the cochlear implant has emerged in the last decades as a major improvement for severe forms of deafness, this medical device still suffers from the lack of resolution which degrades the quality of restored hearing and from contraindications in a significant number of patients. To circumvent these issues, pioneering projects envision the use of auditory cortex implants placed on the last stage of the auditory system in the central brain, as they offer more interaction surface and address a highly plastic network. However, compared to cochlear implants, these new medical devices require novel encoding algorithms adapted to the neuronal code for sounds in the auditory cortex at a high level in the brain. To achieve this, the Braincoder project proposes to leverage recent results made in the context of the ERC CoG Deepen project which characterised essential transformations of the auditory information from the cochlea to the cortex and established that they could be reproduced with deep neural networks. Following these results we have implemented a network which solves the difficult problem of finding the mathematical transformations that compress information about arbitrary sounds onto the stimulator array of the implant, while respecting the key rules of sound encoding in the target brain region. The network architecture chosen allows estimating precisely the information loss in the compression process and to benchmark cortical implant performance against cochlear implants. This technology, along with auditory rehabilitation using a high density cortical stimulation system, has been the subject of a recent patent filing. The goal of the Braincoder project is to consolidate this patent filing and implement these benchmarks to transform this research breakthrough into a robust software demonstrator as well as to establish the business plan to fully exploit our innovations via the creation of a startup company.