因衰老导致的运动功能衰退和因疾病造成的运动障碍是导致运动失能的重大风险因素。便捷易用、高性价比的运动功能实时评估与长期监测方案对维护老年人运动健康、辅助运动障碍患者的个性化诊疗，具有重要意义。目前，针对老年人和以帕金森病为代表的运动障碍患者的功能评估仍主要依赖临床资源，无法实现规模化、及时性的评估和居家监测。本研究提出了一种基于姿态解算、机器学习和临床专家知识的实时动作姿态识别和运动功能精细化分析方法，可用于老年人和帕金森病患者的运动功能评估和监测。该方法采用嵌入通用运动手环的小型、低成本惯性测量单元（inertial measurement unit, IMU），结合机器学习和人体运动信号分析，实现实时的动作姿态识别和运动功能评估，具有较高的准确性和稳定性，可为人体日常运动功能监测提供有效途径，并可应用到一些运动功能障碍患者的诊疗辅助与康复训练中，如监测帕金森病患者运动症状进展，并为其后续治疗方案评估提供依据。本研究参照运动障碍评估量表，设计了一组运动任务，招募65名健康老年人和8名帕金森病患者进行实验。每名患者在实验前均接受了临床医生的医学评估。通过收集健康老年受试者完成各运动任务时的手腕部IMU数据、视频数据及光学运动捕捉数据，以及帕金森病患者的IMU和视频数据，建立了针对特定运动任务的多模态数据集。基于健康老年人的IMU数据，结合特征提取与支持向量机，训练实时动作姿态和运动状态识别模型，准确率达到92%，并进行了健康老年人各任务分类特征权重分析以及患者不同任务之间特征分布的显著性差异比较。基于动作时的运动信号，提取运动功能精细化分析指标，通过视频标注和三维光学运动捕捉数据验证了IMU结果的准确性。结果表明，本研究提出的方法可以有效地评估健康老年人和帕金森病患者的肢体运动控制能力，其中，患者评估结果与临床评估指标具有高度一致性。本研究为实时动作姿态识别与精细化运动功能评估提供了一种低成本、可规模化推广的可行方案，在老年人与帕金森病患者的日常运动功能评估与监测方面具有较高的实际应用潜力与价值，可为因医疗资源不足造成的评估与监测困难提供解决途径，并能为运动功能障碍患者提供更为灵活便捷的临床诊疗辅助手段。
Declining motor function due to aging and motor impairment due to disease are significant risk factors for motor disability. Convenient and cost-effective real-time assessment and long-term monitoring tools for motor functions are of great importance for maintaining motor health of the elderly and assisting precise diagnosis and treatment for patients with movement disorders. Currently, functional assessments for the elderly and patients with movement disorders such as Parkinson‘s disease, still largely rely on clinical resources, making it almost impossible to achieve large-scale, real-time assessments and home monitoring.This study proposes a real-time activity recognition and fine-grained analysis method of motor functions based on pose estimation, machine learning and clinical knowledge, which can be used for the evaluation and monitoring of motor functions of the elderly and patients with Parkinson‘s disease. This method uses a small and low-cost inertial measurement unit (IMU) embedded in a commercial smartwatch. Real-time gesture recognition and motor function evaluation with relatively high accuracy and stability are achieved with machine learning and human motion signal analysis, providing an effective way for monitoring daily motor function of the human body. It can be applied to the diagnosis and treatment assistance and rehabilitation training for patients with motor dysfunctions, such as monitoring the progression of motor symptoms in patients with Parkinson‘s disease, and providing support for the evaluation of follow-up treatment plans. In this study, a set of motor tasks was designed with reference to the Movement Disorder Assessment Scale, and 65 healthy elderly people and eight patients with Parkinson‘s disease were recruited. Each patient received a medical evaluation by a clinician prior to the experiment. By collecting wrist IMU data, video data and optical motion capture data of healthy elder participants, and IMU and video data of patients with Parkinson‘s disease, a multimodal dataset for specific movement tasks was established. Based on the IMU data of healthy elderly people, combined with feature extraction and support vector machine, the real-time action posture and motion state recognition model was trained, yielding an accuracy of 92%. The weight analysis of the classification features of healthy elderly people and the relationship between different tasks of patients were carried out. Based on the IMU motion signal during movements, the refined analytic indicators of motor function were extracted, and the accuracy was verified by video annotation and 3D optical motion capture data. The results show that the method proposed in this study can effectively assess the limb motor control ability of healthy elderly and patients with Parkinson‘s disease. It can provide a solution to the evaluation and monitoring difficulties caused by insufficient medical resources, and can provide more flexible and convenient clinical diagnosis and treatment assistances for patients with movement disorders.
