CAREER: Generalizable, Robust, and Closed-Loop Brain-Machine Interface Control Architectures

脑机接口(BMI)旨在恢复数百万残疾人的运动。尽管实验室演示成功，但缺乏可推广性和健壮性，以及低性能仍然是阻碍临床生存的关键挑战。BMI应该能够控制各种假肢设备，并利用任何神经信号形式作为它们的控制信号。这项研究开发了用于神经假体控制的通用性、健壮性和闭环式BMI架构，并将这些架构应用于构建熟练的神经假体和研究这种控制的大脑机制。这项研究与外展和教育活动完全结合在一起，包括与残疾退伍军人的互动，以及对妇女和代表性不足的少数民族的指导。现有BMI缺乏通用性和稳健性的一个主要原因是，它们没有对所有BMI环境中单一共同组件的行为进行建模，即控制运动的大脑。此外，当前BMI的性能已经被牺牲，因为它们没有适应所记录的神经信号模式的统计特性，并且对任何模式使用标准信号处理算法。这项研究开发了一种BMI，它可以使用不同的神经记录模式，以不同的动力学方式控制假肢。它建立了一种新的闭环控制的大脑模型，针对不同的记录模式构建了原则性随机模型，并将这两种模型结合起来设计了一种自适应监督学习和解码算法。它还使用该架构来研究神经假体控制的大脑机制。这项研究实现了通用和原则性的神经假体架构，取代了特别的方法；它显著改善了神经假体的性能；最后，它允许更深入地了解神经假体和运动控制的基本大脑机制。

Brain-machine interfaces (BMI) aim to restore movement in millions of disabled people. Despite successful laboratory demonstrations, the lack of generalizability and robustness, and the low performance remain key challenges hindering clinical viability. BMIs should be able to control a variety of prosthetic devices, and to exploit any neural signal modality as their control signal. This research develops generalizable, robust, and closed-loop BMI architectures for neuroprosthetic control, and applies these architectures both to build proficient neuroprosthetics and to investigate the brain mechanisms underlying such control. This research is fully integrated with outreach and education activities including interactions with disabled Veterans, and mentoring of women and underrepresented minorities.One main reason for the lack of generalizability and robustness in existing BMIs is that they do not model the behavior of the single common component in all BMI settings, i.e., the brain, which controls the movement. Moreover, performance of current BMIs has been sacrificed because they have not been adapted to the statistical properties of the recorded neural signal modality and have used standard signal processing algorithms for any modality. This research develops a BMI that can control prosthetics with various dynamics and using different neural recording modalities. It builds a novel model of the brain in closed-loop control, constructs principled stochastic models for different recording modalities, and combines these two models to devise an adaptive supervised learning and decoding algorithm. It also uses the architecture to investigate the brain mechanisms underlying neuroprosthetic control. This research enables a universal and principled neuroprosthetic architecture, replacing ad-hoc approaches; it significantly improves neuroprosthetic performance; finally, it allows for a deeper understanding of the fundamental brain mechanisms underlying neurprosthetic and motor control.