User-driven Retrospectively Supervised Classification Updating (RESCU) system for robust upper limb prosthesis control

摘要 在美国，大约有41，000人患有上肢损失（至少失去一只手）。幸运的是， 假体装置在过去几十年中随着灵巧的发展而有了相当大的进步， 拟人化的手然而，可能最有前途的控制策略，肌电控制， 缺乏相应的高水平的性能，因此使用灵巧的手仍然很高， 有限公司对上肢假肢控制进行全面改革的必要性在以下方面得到了很好的强调： 肌电装置的废弃率，在经肱骨的情况下可高达40% 截肢者在过去的十年里，最受关注的研究领域是“模式 识别”，这是一种基于信号处理的控制方法，使用多通道表面 肌电图作为控制输入。虽然模式识别提供了多个应用程序的直观操作， 假肢的自由度，它缺乏鲁棒性，需要频繁，往往每天校准。因此，它已经 尚未达到预期的临床接受度。 我们的团队提出了一种新型的高度适应性上肢假肢控制系统的临床翻译， 结合了两个主要的进步：1）机器学习（通过实现非边界分类的鲁棒分类） 基于算法），以及2）通过回顾性地合并来自日常生活活动（ADL）的用户数据进行训练。 拟议的系统将使机器智能与用户输入假肢控制。我们的工作是 组织如下： 第一阶段：（a）首先，我们将实施一种全新的机器智能技术，即极限学习 具有自适应稀疏表示分类（EASRC）的机器，对未经训练的 用户在现实世界中可能遇到的噪声条件，并且需要比传统肌电更少的数据 信号处理(b)同时，我们将实现一个自适应学习算法Nessa，它允许 用户重新标记在使用过程中记录的错误分类数据，然后更新EASRC分类器，以适应任何 信号中的主要外在或内在变化。EASRC和Nessa共同组成了 追溯监督分类更新（RESCU）系统。 一旦RESCU实现完成，我们将通过与Johns的共同努力来优化系统 霍普金斯大学，并完成了一个迭代台式RESCU评估与焦点小组3截肢者 受试者和他们的假肢专家 第二阶段：将完成RESCU的验证和确认，最终进行第三方确认测试 和认证。最后，我们将完成临床评估，包括自我报告的主观测量， 和长期临床研究中的真实世界使用度量。

ABSTRACT Approximately 41,000 individuals live with upper-limb loss (loss of at least one hand) in the US. Fortunately, prosthetic devices have advanced considerably in the past decades with the development of dexterous, anthropomorphic hands. However, potentially the most promising used control strategy, myoelectric control, lacks a correspondingly high-level of performance and hence the use of dexterous hands remains highly limited. The need for a complete overhaul in upper limb prosthesis control is well highlighted by the abandonment rates of myoelectric devices, which can reach up to 40% in the case of trans-humeral amputees. The area of research that has received the most focus over the past decade has been “pattern recognition,” which is a signal processing based control method that uses multi-channel surface electromyography as the control input. While pattern recognition provides intuitive operation of multiple prosthetic degrees of freedom, it lacks robustness and requires frequent, often daily calibration. Thus, it has not yet achieved the desired clinical acceptance. Our team proposes clinical translation of a novel highly adaptive upper limb prosthesis control system that incorporates two major advances: 1) machine learning (robust classification by implementing a non-boundary based algorithm), and 2) training by retrospectively incorporating user data from activities of daily living (ADL). The proposed system will enable machine intelligence with user input for prosthesis control. Our work is organized as follows: Phase I: (a) First, we will implement a fundamentally new machine intelligence technique, Extreme Learning Machine with Adaptive Sparse Representation Classification (EASRC), that is more resilient to untrained noisy conditions that users may encounter in the real-world and requires less data than traditional myoelectric signal processing. (b) In parallel, we will implement an adaptive learning algorithm, Nessa, which allows users to relabel misclassified data recorded during use and then update the EASRC classifier to adapt to any major extrinsic or intrinsic changes in the signals. Taken together, EASRC and Nessa comprise the Retrospectively Supervised Classification Updating (RESCU) system. Once, the RESCU implementation is complete, we will optimize the system through a joint effort with Johns Hopkins University, and complete an iterative benchtop RESCU evaluation with a focus group of 3 amputee subjects and their prosthetists. Phase II: Verification and validation of RESCU will be completed, culminating in third-party validation testing and certification. Finally, we will complete a clinical assessment including self-reporting subjective measures, and real-world usage metrics in a long-term clinical study.