Enhanced cochlear implant coding using stochastic beam-forming

人工耳蜗（CIs）是通过电刺激耳蜗神经（听觉神经）来恢复重度耳聋患者的听力。理论上，人工耳蜗使用电刺激来激发神经活动（即激发耳蜗神经纤维），模仿正常听力中可能发生的神经活动。然而，常规CI刺激无法模仿的正常神经活动的一个方面是耳蜗神经纤维的随机“放电”——这种活动即使在没有输入刺激的情况下也会发生。在正常的耳朵中，这种粗大的随机活动是由毛细胞内的噪声源引起的，这些噪声源将声音信号转化为电信号。这种噪音有几个来源，比如：毛细胞的立体纤毛（毛发）的布朗运动，以及将信号从毛细胞传递到耳蜗神经纤维的化学物质的随机释放。这些噪声源在耳聋中是不存在的，因为深度耳聋与毛细胞的完全丧失有关。然而，现在有相当多的证据表明，这些噪声源是正常神经编码的重要组成部分，因此我们之前提出，应该通过将噪声源合并到ci中，将它们重新引入失聪的耳朵。但传统上，噪音被视为令人讨厌的东西，而且有充分的理由；如果以不受控制的方式增加噪音，几乎肯定会导致植入人工耳蜗的人的语言理解能力下降。为了发挥作用，激发特定耳蜗神经纤维的噪声波形必须与刺激邻近纤维的噪声波形不同——这将确保相邻纤维的放电是独立的。简单地在每个手术植入的电极上施加噪声电流是不可能产生期望的独立性的；这是因为耳蜗中充满了导电盐溶液，使得电极产生的电流在整个耳蜗中扩散；因此，噪声电流相互作用并产生在宽空间范围内强相关的有效刺激。为了避免这个问题，我们开发了一种技术，可以减少当前价差的影响。每个电极的噪声电流来自独立噪声源的总和，每个噪声源通过加权项进行缩放；可以选择这些权重来产生具有指定去相关长度（刺激变得不相关的距离）的空间随机场。通过这种方式，准独立的放电可以在一群耳蜗神经纤维中实现。我们将这种技术称为随机波束形成，因为它依赖于噪声源的非相干求和来产生零相关的“波束”——这个概念类似于天线阵列中的波束形成。初步的计算研究表明，这种方法是可行的，并且具有很强的鲁棒性。我们建议扩展我们的初步研究，使用更完整的电刺激耳模型。关键的是，我们将在Clarion人工耳蜗（Advanced Bionics Ltd）的用户身上测试这种方法。我们将测量我们的策略能在多大程度上实现独立的噪声刺激，我们将测量植入者对语言理解的改善。这些测试将在圣托马斯医院（伦敦）进行，并与莫妮塔·查特吉博士（马里兰大学）和先进仿生学公司合作。这项研究的重要性怎么强调也不过分。在我们之前和目前由epsrc资助的建模工作中，我们已经清楚地展示了使用噪声来提高语音理解的潜力。然而，这种方法只有在我们能够在耳蜗神经纤维群的神经冲动之间获得更大的独立性的情况下才能在实践中起作用。这项工作是至关重要的一步，它将使我们能够实现随机编码策略所带来的好处。

Cochlear implants (CIs) are used to restore the hearing of profoundly deaf people by electrical stimulation of the cochlear nerve (the nerve of hearing). Notionally, a CI uses electrical stimulation to evoke neural activity (i.e. firing of the cochlear nerve fibres) that mimics the neural activity that would have occurred in normal hearing. However, one aspect of normal neural activity that is not mimicked by conventional CI stimulation is the gross random 'firing' of cochlear nerve fibres - this activity occurs even in the absence of an input stimulus. In the normal ear, this gross random activity is caused by sources of internal noise in the hair cells that transduce sound signals into electrical signals. The noise has several sources such as: Brownian motion of the stereocilia (hairs) of a hair cell, and the random release of the chemicals that transmit the signal from a hair cell to cochlear nerve fibres. These noise sources are absent in the deafened ear because profound deafness is associated with complete loss of the hair cells. There is now considerable evidence, however, that these noise sources are an essential part of normal neural coding and we have therefore previously proposed that they should be re-introduced back into the deafened ear by incorporating noise sources into CIs.But traditionally noise is regarded as a nuisance, and for good reason; if the noise is added in an uncontrolled manner it will almost certainly lead to worse speech comprehension for cochlear implantees. To be useful the noise waveform that excites a particular cochlear nerve fibre must be dissimilar to those that stimulate neighbouring fibres - this will ensure that the firing of adjacent fibres will be independent. Simply applying a noisy current to each surgically implanted electrode is unlikely to produce the desired independence; this is because the cochlea is filled with conductive salt solutions that causes the currents from the electrodes to spread throughout the cochlea; the noise currents therefore interact and result in an effective stimulus that is strongly correlated over a wide spatial range. To circumvent this problem we have developed a technique that reduces the effect of the current spread. The noise currents for each electrode are derived from a sum of independent noise sources, each scaled by a weighting term; these weights can be chosen to produce a spatial random field with a specified de-correlation length (the distance over which the stimulus becomes uncorrelated). In this manner quasi-independent firing can be achieved across a population of cochlear nerve fibres. We refer to this technique as stochastic beamforming because it relies on the incoherent summation of the noise sources to produces 'beams' of zero correlation - this concept is similar to beamforming in antenna arrays. A preliminary computational study has shown that this approach appears feasible and extremely robust.We propose to extend our preliminary study and use more complete models of the electrically stimulated ear. Critically, we will test the approach with users of the Clarion cochlear implant (Advanced Bionics Ltd). We will measure the extent to which our strategy enables independent noise stimulation and we will measure the improvement to the speech comprehension of implantees. These tests will be done at St Thomas' Hospital (London) and in collaboration with Dr Monita Chatterjee (University of Maryland) and Advanced Bionics. The importance of this study cannot be overstated. In our previous and current EPSRC-funded modelling work, we have clearly demonstrated the potential for using noise to improve speech comprehension. The method, however, will only work in practice if we can get greater independence between the nerve impulses for the population of cochlear nerve fibres. This work is the essential step that will enable us to realise the benefits that stochastic coding strategies promise.