Learning to Automatically Evaluate Pathological Gait: A Data-Driven System for Characterizing Disability and Informing Therapeutic Interventions

能够平衡是我们大多数人认为理所当然的事情。然而，大约35%的40岁及以上的美国公民受到前庭相关平衡问题的影响。前庭系统位于内耳，是为我们的中枢神经系统提供平衡和空间定向信息的几个感觉系统之一。当一个人的前庭系统因疾病或受伤而受损时，除了头晕和眩晕外，他/她还可能会出现平衡和步态缺陷。有身体，情感和金钱的成本与基于感觉的平衡残疾，如前庭残疾，和福尔斯，通常伴随着平衡不稳定的发作。大多数跌倒相关损伤发生在行走（步态）过程中，但治疗步态不平衡具有挑战性。目前用于评估前庭残疾人的步态病理（由于受伤或疾病导致的步态异常）的临床工具不能完全捕获身体运动，忽略了基于步态的活动期间与感觉相关的残疾的潜在关键特征。该项目的目标是开发和测试用于表征病理运动的数据驱动算法（问题解决指令）。这项工作将导致评估感觉相关步态障碍的新方法，并支持开发新的康复策略。作为这项研究的一部分，大型运动传感网络将与机器学习算法相结合，以识别和测量前庭残疾人的步态异常。虽然这个项目的重点是前庭残疾，开发的方法可以推广到广泛的平衡障碍源于感官残疾，受伤，神经残疾，运动残疾和衰老。这项研究还将有助于通过顶点设计项目，临床沉浸经验，以确定未满足的康复需求，以及开发和实施开放获取，在线教育模块，专注于机器学习对社会影响的应用。该项目的主要目的是开发和评估数据驱动的机器学习（ML）识别和量化前庭残疾人病理步态的算法，目的是为创建新的评估技术提供信息，并支持开发新的康复策略。 研究计划有三个目标。 第一个目标是创建一个共享的数据库的步态测量与前庭残疾的主题。 活动包括：a）招募具有前庭缺陷的参与者和年龄匹配的健康对照，B）在实验阶段期间收集运动学数据，其中受试者装备有全套被动标记和多达17个伊穆斯（惯性测量单元），眼震电图测试组，以及d）基于由一小群物理治疗师（PT）在1-5视觉模拟量表上观看和评级的录像步态康复练习来收集物理治疗师（PT）标签，以及d）通过将数据组织成可以以本地数据库格式下载的表格来共享数据。 第二个目标是开发强大的数据驱动ML算法，用于自动评估和表征前庭残疾人的病理步态模式。 根据设计用于学习数据驱动模型的子目标组织活动，以a）自动区分前庭残疾受试者与健康对照，B）通过为每个临床亚组开发原型步态病理学概念来表征亚组，以及c）量化残疾程度并生成关于根本感觉运动或生物力学问题的假设。 第三个目标是开发和前瞻性地评估一个便携式系统进行实时评估。 活动包括：a）开发便携式智能手机步态评估工具，该工具将使用从不超过7个伊穆斯获得的数据在基于步态的康复训练期间生成实时评级，以及B）在涉及10名成年人的概念验证研究中前瞻性地测试该系统。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Being able to balance is something most of us take for granted. However, approximately 35% of U.S. citizens 40 years and older are affected by vestibular-related balance issues. The vestibular system, located in the inner ear, is one of several sensory systems that provides our central nervous system with balance and spatial orientation information. When a person's vestibular system is impaired by disease or injury, he/she can experience balance and gait deficits in addition to dizziness and vertigo. There are physical, emotional, and monetary costs associated with sensory-based balance disabilities, such as vestibular disabilities, and the falls that typically follow bouts of balance instability. Most fall-related injuries occur during walking (gait), but treating imbalance during gait is challenging. Current clinical tools for assessing gait pathologies (gait abnormalities due to injury or disease) in people with vestibular disabilities do not fully capture body motion, neglecting potentially critical features of sensory-related disabilities during gait-based activities. The goal of this project is to develop and test data-driven algorithms (problem solving instructions) for characterizing pathological motion. This work will lead to new methods for assessing sensory-related gait disorders and support the development of novel rehabilitation strategies. As part of this research, large motion sensing networks will be combined with machine learning algorithms to identify and measure gait abnormalities in people with vestibular disabilities. Though the focus of this project is on vestibular disabilities, the methods developed can be generalized to a wide range of balance impairments stemming from sensory disabilities, injuries, neural disabilities, motor disabilities, and aging. This research will also contribute to the training of both undergraduate and graduate students through capstone design projects, clinical immersion experiences to identify unmet rehabilitation needs, and the development and implementation of an open access, online educational module focused on applications of machine learning for societal impact.This project's primary purpose is to develop and assess data-driven machine learning (ML) algorithms that identify and quantify pathological gait in people with vestibular disabilities for the purposes of informing the creation of new assessment techniques and supporting the development of novel rehabilitation strategies. The Research Plan is organized under three objectives. The first objective is to create a shareable database of gait measurements from subjects with vestibular disabilities. Activities include: a) recruiting participants with vestibular deficits and age-matched healthy controls, b) collecting kinematic data during an experimental session in which subjects are instrumented with a full set of passive markers and up to 17 IMUs (Inertial Measurement Units), c) collecting clinical vestibular testing diagnostic data, e.g., electronystagmography test battery, and d) collecting Physical Therapist (PT) labels based on videotaped gait rehabilitation exercises that are viewed and rated on a 1-5 visual analog scale by a small cohort of PTs and d) sharing data by organizing data into tables that can be downloaded in a local database format. The second objective is to develop robust data-driven ML algorithms for automatically evaluating and characterizing pathological gait patterns in people with vestibular disabilities. Activities are organized under sub-objectives designed to learn data-driven models to a) automatically differentiate subjects with vestibular disabilities from healthy controls, b) characterize subpopulations by developing a notion of prototypical gait pathologies for each clinical subgroup and c) quantify the extent of the disability and generate hypotheses regarding the root sensorimotor or biomechanical problem. The third objective is to develop and prospectively evaluate a portable system for real-time assessment. Activities include: a) developing a portable smartphone gait assessment tool that will generate real-time ratings during gait-based rehabilitation exercises using data obtained from no more than 7 IMUs and b) prospectively testing the system in a proof-of-concept study involving 10 adults.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.