Ethical Considerations for Language Modeling within Brain-Computer Interfaces

项目摘要 机器学习（ML）和自然语言处理（NLP）有可能改变通信 为神经退行性疾病患者提供个性化和实时的辅助和替代治疗， AAC（Audio Communication）设备。有严重沟通障碍的人，他们不再能控制 他们的日常对话或参与以前的生活角色需要AAC设备。他们想让他们工作- 可靠、有效、快速。ML和NLP正在成为连接当前技术和 下一代设备，为个人与最严重的语言和身体障碍，如RSVP Keyboard™，一种由父母资助开发的脑机接口（BCI）。BCI系统用于 这些通信被称为AAC-BCI。NLP努力将大型公共数据集与私有数据集结合起来， 例如个人电子邮件消息，承诺给予有通信障碍个人他们自己的 个性化的语言模型，这些模型足够强大，可以更接近实时通信。的 然而，专注于让AAC-BCI与机器学习一起工作，导致了该领域的一个关键疏忽： 对个人为什么需要下一代设备以及他们愿意做出哪些权衡的理解不足 以实现更快、更个性化的沟通。转向ML使这一疏忽得到了极大的缓解。 个人应该提供关于用于构建其个人语言模型的数据集的输入，但这 提出了一些重要的伦理问题，比如个人的价值观是什么，他们如何理解自己的身份，以及 他们愿意相对于他们的个性化通信数据做出的权衡。本补充的目的是 是填补这一理解上的空白，以便研究人员能够将ML应用到下一代AAC-BCI系统中， 以一种对未来用户的道德问题敏感的方式。这种道德有四个组成部分 补充：（1）设计一个专门针对BCI提出的道德问题的道德小插曲工具箱， （2）每月对有严重道德问题的个人进行基于小插曲的在线道德调查 由于运动神经元疾病引起的通信损伤（例如，ALS）（n=25）或运动障碍（例如， 帕金森氏病）（n=25）;（3）进行半结构化的基于插图的访谈与个人与前 运动神经元疾病（n=10）或运动障碍（n=10）导致的临床或轻度沟通障碍。 组件（2）和（3）将采用迭代、并行混合方法方法。Likert风格的在线趋势 严重沟通障碍队列中对伦理小插曲的回应将用于告知和修改 半结构化访谈提示询问临床前或轻度损伤队列。同时，主题 从访谈的直接内容分析中得出的结论将用于改进在线调查问题。成果 将使用迭代的混合方法（4）来概述核心伦理领域的框架和初步的 AAC-BCI研究人员可以用来评估伦理问题的工具（插图和讨论提示）， 开发和迭代地改进个性化语言模型的通信技术。

Project Summary Machine learning (ML) and Natural Language Processing (NLP) have the potential to transform communication for patients with neurodegenerative disease through personalized and real-time augmentative and alternative communication (AAC) devices. Individuals with severe communication impairments who can no longer control their daily conversations or participate in previous life roles want AAC devices. And they want them to work – to be reliable, effective, and fast. ML and NLP are emerging as promising tools to bridge current technology and next generation devices for individuals with the most severe speech and physical impairments, like the RSVP Keyboard™, a brain-computer interface (BCI) being developed by the parent grant. BCI systems for communication are referred to as AAC-BCIs. NLP efforts to combine large public data sets with private data sets, such as personal email messages, promise to give individuals with communication impairments their own personalized language models, models that are sufficiently robust to get closer to real-time communication. The focus on getting AAC-BCIs to work with machine learning, however, has led to a critical oversight in the field: an inadequate understanding of why individuals want next-generation devices and what trade-offs they are willing to make for faster and more personalized communication. The turn to ML brings this oversight into sharp relief. Individuals should provide input about the data sets used to construct their personal language models, but this raises important ethical questions about what individuals value, how they understand their identity, and what trade-offs they are willing to make relative to their personalized communication data. The goal of this supplement is to fill this gap in understanding so that researchers can implement ML into next generation AAC-BCI systems in a way that is sensitive to the ethical concerns of future users. There are four components to this ethics supplement: (1) to design a toolbox of ethics vignettes tailored to ethical concerns raised by both BCI communication and ML; (2) to administer monthly vignette-based online ethics surveys to individuals with severe communication impairments due to motor neuron disease (e.g., ALS) (n=25) or movement disorders (e.g., Parkinson's disease) (n=25); (3) to conduct semi-structured vignette-based interviews with individuals with pre- clinical or mild communication impairment due to motor neuron disease (n=10) or movement disorder (n=10). Components (2) and (3) will employ an iterative, parallel mixed-method approach. Trends in Likert-style online responses to ethics vignettes in the severe communication impairment cohort will be used to inform and modify the semi-structured interview prompts asked of the pre-clinical or mild impairment cohort. In parallel, themes emerging from direct content analysis of interviews will be used to refine online survey questions. Results of this iterative, mix-methods approach will be used (4) to outline a framework of core ethical domains and preliminary tools (vignettes and discussion prompts) that AAC-BCI researchers can use to assess ethical concerns while developing and iteratively refining communication technology for personalized language models.