Transforming hearing aids through large-scale electrophysiology and deep learning

听力损失影响全球约5亿人（英国1100万人），使其成为残疾生活年数的第四大原因（英国第三大原因）。由此产生的负担给个人和社会带来了巨大的后果。听力损失阻碍沟通，导致社会孤立，并导致生活质量和福祉下降。它也被确定为偶发性痴呆的主要可改变风险因素，并造成了巨大的经济负担，据估计，英国每年的成本超过300亿英镑。随着听力损失的影响不断增加，对改进治疗方法的需求变得越来越迫切。在大多数情况下，唯一可用的治疗方法是助听器。不幸的是，许多佩戴助听器的人实际上并没有使用它们，部分原因是目前的助听器只不过是简单的放大器，在高音量和背景噪音的社交环境中往往没有什么好处。因此，英国约有300万人患有未经治疗的致残性听力损失，这是一个巨大的未满足的临床需求。助听器使用者和非助听器使用者都有一个常见的抱怨：“我能听到你说的话，但我听不懂你说的话。”由于助听器的目的是促进交流和减少社交孤立，不能在典型的社交环境中感知语言的设备从根本上来说是不充分的。单纯通过扩音就能矫正听力损失的想法过于简单化了；虽然听力损失确实会降低灵敏度，但它也会导致许多其他问题，这些问题会严重扭曲耳朵发送给大脑的信息。为了提高性能，下一代助听器必须采用更复杂的声音转换来纠正这些失真。不幸的是，这说起来容易做起来难。事实上，几十年来，工程师们一直在尝试用手设计助听器来实现这一目标，但收效甚微。幸运的是，最近实验和计算技术的进步为一种完全不同的方法创造了机会。改进助听器的关键困难在于，有无数种可能转换声音的方法，而我们对听力损失的基本原理还不够了解，无法推断出哪种转换是最有效的。然而，现代机器学习技术将允许我们绕过我们理解中的这一差距；如果有一个足够大的声音数据库，以及它们在正常听力和听力受损的情况下引发的神经活动，那么深度学习就可以用来识别声音转换，从而最好地纠正扭曲的活动，并尽可能地恢复正常的感知。所需的神经活动数据库还不存在，但我们在过去几年里一直在开发收集这些数据所需的记录技术。这种能力是独一无二的；世界上没有其他研究小组可以做这些录音。我们已经证明了在计算机上解决机器学习问题的可行性。我们现在提议收集大规模的神经活动数据库，以充分开发基于深度神经网络的新型助听器算法的工作原型，并证明其对听力损失人群的功效。

Hearing loss affects approximately 500 million people worldwide (11 million in the UK), making it the fourth leading cause of years lived with disability (third in the UK). The resulting burden imposes enormous personal and societal consequences. By impeding communication, hearing loss leads to social isolation and associated decreases in quality of life and wellbeing. It has also been identified as the leading modifiable risk factor for incident dementia and imposes a substantial economic burden, with estimated costs of more than £30 billion per year in the UK.As the impact of hearing loss continues to grow, the need for improved treatments is becoming increasingly urgent. In most cases, the only treatment available is a hearing aid. Unfortunately, many people with hearing aids don't actually use them, partly because current devices, which are little more than simple amplifiers, often provide little benefit in social settings with high sound levels and background noise. Thus, there is a huge unmet clinical need with around three million people in the UK living with an untreated, disabling hearing loss. The common complaint of those with hearing loss, "I can hear you, but I can't understand you", is echoed by hearing aid users and non-users alike. Inasmuch as the purpose of a hearing aid is to facilitate communication and reduce social isolation, devices that do not enable the perception of speech in typical social settings are fundamentally inadequate. The idea that hearing loss can be corrected by amplification alone is overly simplistic; while hearing loss does decrease sensitivity, it also causes a number of other problems that dramatically distort the information that the ear sends to the brain. To improve performance, the next generation of hearing aids must incorporate more complex sound transformations that correct these distortions. This is, unfortunately, much easier said than done. In fact, engineers have been attempting to hand-design hearing aids with this goal in mind for decades with little success. Fortunately, recent advances in experimental and computational technologies have created an opportunity for a fundamentally different approach. The key difficulty in improving hearing aids lies in the fact that there are an infinite number of ways to potentially transform sounds and we do not understand the fundamentals of hearing loss well enough to infer which transformations will be most effective. However, modern machine learning techniques will allow us to bypass this gap in our understanding; given a large enough database of sounds and the neural activity that they elicit with normal hearing and hearing impairment, deep learning can be used to identify the sound transformations that best correct distorted activity and restore perception as close to normal as possible.The required database of neural activity does not yet exist, but we have spent the past few years developing the recording technology required to collect it. This capability is unique; there are no other research groups in the world that can make these recordings. We have already demonstrated the feasibility of solving the machine learning problem in silico. We are now proposing to collect the large-scale database of neural activity required to fully develop a working prototype of a new hearing aid algorithm based on deep neural networks and to demonstrate its efficacy for people with hearing loss.