Construction of an Artificial Intelligence Upper Limb Multi-Joint Motion State Recognition System Based on IMU Signals—A Preliminary Study for the Development of an Artificial Intelligence Motor Function Assessment and Detection System after Stroke

doi:10.3969/j.issn.1673-5765.2025.04.003

Abstract

Abstract: Objective This study aims to develop and construct an upper limb multi-joint motion state recognition system based on low-cost inertial measurement unit (IMU) signals, for the rapid and reliable decoding of multi-joint (forearm, elbow joint, and shoulder joint) motion states in human activities, providing support for motion pattern recognition and daily motion monitoring in post-stroke upper limb rehabilitation assessment.
Methods This study enrolled four healthy subjects, collecting their six-dimensional (triaxial acceleration+triaxial angular velocity) motion signals through IMUs deployed on the wrist and upper arm, each subject repeating 10 times. Based on the flexor synergy movement in the Fugl-Meyer motor assessment-upper extremity, eight subtasks were designed, each corresponding to a ternary state label of the forearm (supinated or not), elbow joint (flexed or not), and shoulder joint (elevated or not). A multi-label classification framework based on single-task migration (i.e., independently training single-joint classifiers and then merging the outputs) was constructed. At the algorithm level, traditional machine learning methods (time-frequency domain features+random forest) were compared with deep learning algorithms (long short-term memory-based end-to-end learning). Five-fold cross-validation was used to evaluate the accuracy of the upper limb multi-joint motion state recognition system, and ablation experiments were designed to analyze the impact of sensor configuration (e.g., wrist-only vs. wrist+arm) on decoding performance, exploring hardware optimization potential.
Results A total of 320 motion data samples were collected from four healthy subjects in this study. The results demonstrated that the motion state recognition system designed in this study performed well in multi-joint state decoding of the upper limb. The average accuracy of elbow joint state classification by the traditional machine learning methods was 79.37%, while the deep learning model IBNet reached 87.5%, indicating a stronger pattern-learning capability. The ablation experiment showed that the accuracy of elbow joint state classification exceeded that of dual IMU configuration (92.5% vs. 87.5%) when wrist IMU was used only, and the difference was not significant in other tasks. This suggested that optimizing sensor deployment (e.g., reducing upper arm IMUs) can reduce system complexity while maintaining high performance.
Conclusions This study successfully constructed a low-cost IMU-based upper limb motion state recognition system. The results showed that deep learning algorithms were superior to traditional machine learning in decoding complex motion patterns, and a single-wrist IMU could replace the dual-sensor configuration in specific tasks, providing a basis for hardware optimization.

Key words: Motor function assessment; Stroke; Multi-joint decoding; Inertial measurement unit

摘要： 目的本研究旨在构建并研发一种基于低成本惯性测量单元（inertial measurement unit，IMU）信号的上肢多关节运动状态识别系统，用于快速、可靠地解码人类活动中的上肢多关节（前臂、肘关节、肩关节）运动状态，为卒中后上肢康复评估中的运动模式识别和日常运动监测提供支持。
方法本研究纳入4名健康受试者，通过部署于手腕和上臂的IMU采集受试者的6维（3轴加速度+3轴角速度）运动信号，每名受试者重复10次。基于Fugl-Meyer运动功能评定量表上肢部分的屈肌协同运动，设计8个子任务，每个任务对应前臂（是否旋后）、肘关节（是否屈曲）、肩关节（是否上提）的三元状态标签。构建基于单任务迁移的多标签分类框架（即独立训练单关节分类器后融合输出）。在算法层面，对比了传统机器学习方法（时频域特征+随机森林）与深度学习算法（长短期记忆网络的网络端到端学习）。通过5折交叉验证评估上肢多关节运动状态识别系统的准确性，并设计消融实验分析传感器配置（如单腕 vs. 腕+臂）对解码性能的影响，以探索硬件优化空间。
结果研究共采集4名健康受试者的320个运动数据样本，结果提示，本研究设计的运动状态识别系统在上肢多关节状态解码中表现良好：传统机器学习方法在肘关节状态分类中的平均准确率为79.37%，而深度学习模型IBNet（Inception-BinaryNet）的准确率达到87.5%，显示出更强的模式学习能力。消融实验发现，仅使用手腕IMU时，肘关节状态分类准确率超过了双IMU配置（92.5% vs.87.5%），且在其他任务中差异不显著，这表明优化传感器部署（如减少上臂IMU）可降低系统复杂度，同时保持较高性能。
结论本研究成功构建了一种基于低成本IMU的上肢运动状态识别系统。研究发现，深度学习算法在解码复杂运动模式时优于传统机器学习，且单手腕IMU在特定任务中可替代双传感器配置，为硬件优化提供了依据。

关键词: 运动功能评估; 卒中; 多关节解码; 惯性测量单元

CLC Number:

R74
R318

CHENG Xiangxin, ZHANG Shuo, DU Songjun, LIU Ziyang, ZHOU Hongyu, JIA Weili, LI Zixiao, LIU Tao. Construction of an Artificial Intelligence Upper Limb Multi-Joint Motion State Recognition System Based on IMU Signals—A Preliminary Study for the Development of an Artificial Intelligence Motor Function Assessment and Detection System after Stroke[J]. Chinese Journal of Stroke, 2025, 20(4): 401-409.

程相鑫, 张烁, 杜松骏, 刘子阳, 周宏宇, 贾伟丽, 李子孝, 刘涛. 基于IMU信号的人工智能上肢多关节运动状态识别系统构建——卒中后人工智能运动功能评估与检测系统建设前导研究[J]. 中国卒中杂志, 2025, 20(4): 401-409.

References

[1] 王拥军，李子孝，谷鸿秋，等. 中国卒中报告2020（中文版）（1）[J]. 中国卒中杂志，2022，17（5）：433-447.
WANG Y J，LI Z X，GU H Q，et al. China stroke statistics 2020（1）[J]. Chin J Stroke，2022，17（5）：433-447.
[2] WANG X，ZHANG J，XIE S Q，et al. Quantitative upper limb impairment assessment for stroke rehabilitation：a review[J]. IEEE Sensors Journal，2024，24（6）：7432-7447.
[3] PROIETTI T，BANDINI A. Wearable technologies for monitoring upper extremity functions during daily life in neurologically impaired individuals[J/OL]. IEEE Trans Neural Syst Rehabil Eng，2024，32：2737-2748[2025-02-10]. https://doi.org/10.1109/TNSRE.2024.3435042.
[4] WANG S，LIU J W，CHEN S，et al. Recognizing wearable upper-limb rehabilitation gestures by a hybrid multi-feature neural network[J/OL]. Eng Appl Artif Intell，2024，127：107424[2025-02-10]. https://doi.org/10.1016/j.engappai.2023.107424.
[5] SONG X Y，CHEN S G，JIA J，et al. Cellphone-based automated Fugl-Meyer assessment to evaluate upper extremity motor function after stroke[J]. IEEE Trans Neural Syst Rehabil Eng，2019，27（10）：2186-2195.
[6] UEYAMA Y，TAKEBAYASHI T，TAKEUCHI K，et al. Attempt to make the upper-limb item of objective Fugl-Meyer assessment using 9-axis motion sensors[J/OL]. Sensors（Basel），2023，23（11）：5213[2025-02-10]. https://doi.org/10.3390/s23115213.
[7] CONTRERAS R C，PARNANDI A，COELHO B G，et al. NE-motion：visual analysis of stroke patients using motion sensor networks[J/OL]. Sensors（Basel），2021，21（13）：4482[2025-02-10]. https://doi.org/10.3390/s21134482.
[8] OUBRE B，DANEAULT J F，JUNG H T，et al. Estimating upper-limb impairment level in stroke survivors using wearable inertial sensors and a minimally-burdensome motor task[J]. IEEE Trans Neural Syst Rehabil Eng，2020，28（3）：601-611.
[9] OUBRE B，DANEAULT J F，WHRITENOUR K，et al. Decomposition of reaching movements enables detection and measurement of ataxia[J]. Cerebellum，2021，20（6）：811-822.
[10] OUBRE B，LEE S I. Detection and assessment of point-to-point movements during functional activities using deep learning and kinematic analyses of the stroke-affected wrist[J]. IEEE J Biomed Health Inform，2024，28（2）：1022-1030.
[11] SIAHROUDI S K，KUDENKO D. An effective single-model learning for multi-label data[J/OL]. Expert Syst Appl，2023，232：120887[2025-02-10]. https://doi.org/10.1016/j.eswa.2023.120887.
[12] CHRIST M，BRAUN N，NEUFFER J，et al. Time series feature extraction on basis of scalable hypothesis tests（tsfresh—a python package）[J/OL]. Neurocomputing，2018，307：72-77[2025-02-10]. https://doi.org/10.1016/j.neucom.2018.03.067.
[13] LUBBA C H，SETHI S S，KNAUTE P，et al. Catch22：canonical time-series characteristics[J/OL]. Data Min Knowl Discov，2019，33：1821-1852[2025-02-10]. https://doi.org/10.1007/s10618-019-00647-x.
[14] BREIMAN L. Random forests[J]. Machine Learning，2001，45（1）：5-32.
[15] ISMAIL FAWAZ H，LUCAS B，FORESTIER G，et al. InceptionTime：finding AlexNet for time series classification[J]. Data Min Knowl Discov，2020，34（6）：1936-1962.
[16] SINGH S P，SHARMA M K，LAY-EKUAKILLE A，et al. Deep ConvLSTM with self-attention for human activity decoding using wearables[J]. IEEE Sensors Journal，2021，21（6）：8575-8582.
[17] DING K J，WANG J P，WANG X J，et al. System for detection and quantitative evaluation of compensatory movement in post-stroke patients based on wearable sensor and machine learning algorithm[J]. IEEE Sensors Journal，2024，24（14）：22830-22842.