Background & Summary

Stroke, also known as cerebrovascular accident (CVA), is an acute disease caused by the interruption of blood flow to the brain, resulting in damage to brain tissue1. Motor dysfunction is a common sequela among stroke patients, manifesting as limb paralysis, muscle weakness, or impaired coordination2. This motor dysfunction is closely associated with injury to specific regions of the brain following a stroke, especially those responsible for regulating motor function, such as the motor cortex, basal ganglia, and spinal cord3,4. To regain motor function, patients typically require long-term rehabilitation that includes professional assessments and treatments during hospitalization as well as home-based rehabilitation training post-discharge, aimed at maximizing neural plasticity and enhancing quality of life5,6,7,8.

Rehabilitation assessment and training are critical for restoring motor function. Assessment serves as the foundation of rehabilitation, enabling healthcare teams to gain deep insights into patients’ functional status and develop personalized rehabilitation plans9,10,11. Standardized scales are commonly used, such as the Brunnstrom stages12 for assessing the recovery status of hemiplegic limbs and the Fugl-Meyer Assessment (FMA)13 for assessing fine motor skills. Despite their value, clinical assessments still face challenges of being time-consuming and subjective, highlighting the need for intelligent enhancements14,15,16. Meanwhile, home-based rehabilitation training is also crucial, especially the activity-based task-oriented training17. Its effectiveness relies heavily on patients’ adherence and gradual progression18. While home programs offer the advantages of long-term exercise and increased convenience, they also present limitations. Key challenges include the lack of professional guidance, low compliance with exercise regimens, limited access to equipment and technological support, insufficient assessment mechanisms, and inadequate feedback systems19,20. To address the above challenges, a widely recognized intelligent solution consists of three key steps:

Step 1: Establish a data acquisition system. Common kinematic sensors include inertial measurement units (IMUs) and cameras. IMUs can be used to measure acceleration and angular velocity to assess movement patterns and coordination21. Cameras can provide intuitive visual-level motion information. However, considering the convenience and privacy requirements of home-based rehabilitation, camera-based sensors are often restricted22,23. In home rehabilitation, low-cost and portable wearable devices are more practical and promote broader adoption24,25. Therefore, IMUs represent a practical and cost-effective solution for rehabilitation-oriented motion monitoring, providing portable, privacy-friendly, and comprehensive kinematic sensing capabilities in home-based rehabilitation scenarios26,27.

Step 2: Establish data analysis models. To analyze the data collected by the acquisition system, mathematical models will be established, such as machine learning or deep learning models. By leveraging motion data, these individual models can achieve motor function assessment28, rehabilitation behavior recognition29, movement accuracy assessment30, fatigue calculation31, and so on. These models can reduce workload, eliminate subjectivity, and enhance rehabilitation effectiveness.

Step 3: Build a comprehensive rehabilitation system. A comprehensive application-oriented software platform is then developed, linking the hardware and algorithms, and featuring an intuitive and visualized user interface. The platform integrates the functions of individual models, visualizes the model outputs, and provides data recording, storage, and comprehensive analysis32. Through this platform, rehabilitation progress can be continuously tracked, and detailed feedback reports can be generated. By offering real-time visual and auditory feedback, the system assists patients in correcting movements, improving adherence, and enhancing rehabilitation effectiveness.

Both the second and third steps require large volumes of pre-existing kinematic datasets for pre-training individual models, model validation, and system verification. However, existing datasets are insufficient for training models specifically designed for patients with impaired motor function, as they suffer from the following limitations:

  1. (1)

    Lack of rehabilitation-specific tasks, limiting dataset relevance: Existing kinematic datasets (UCI33, Pamap234, HANDY35, and36) mainly contain activities of daily living (ADL), such as walking and sitting29. Although they can be used to build pre-trained kinematic-based models, they lack rehabilitation-specific tasks, such as functional training exercises, fine motor movements, and resistance-based training, which are necessary to adjust the model and form a specialized one37. This limitation results in insufficient diversity of rehabilitation-related movements during model training, reducing accuracy when handling complex or specialized rehabilitation tasks. Consequently, the models fail to comprehensively reflect the real needs of patients with motor impairments.

  2. (2)

    Insufficient patient data limits model generalization and clinical applicability: Currently, systematic kinematic datasets based on real patients suffer from insufficient data volume. For instance, IntelliRehabDS38 (IRDS) contains kinematic data tailored to physical rehabilitation but is constrained by a limited number of enrolled patients. By contrast, most large-scale kinematic sensor datasets are constructed using data from healthy individuals, such as39. This scarcity of patient-specific data has compelled many studies40,41 to rely on self-collected datasets, which typically include only healthy participants or a small cohort of patients. Consequently, models trained on such inadequate datasets fail to accurately capture the unique movement characteristics and recovery trajectories of patients. This not only impairs the models’ generalization ability across diverse pathological conditions but also restricts their practical applicability in clinical settings.

  3. (3)

    Dominance of vision-based data, limiting privacy and convenience in home rehabilitation: Many kinematic datasets (NTU RGB + D42, Kinetics-Skeleton43,44) primarily rely on video data, which, although effective for action recognition, raises privacy concerns in home rehabilitation settings45,46,47. Recent advances in computer vision and single-camera body tracking frameworks have reduced the hardware and deployment requirements of vision-based systems. However, continuous video monitoring in personal living environments may still limit user acceptance and long-term applicability in home rehabilitation scenarios.

  4. (4)

    Inconsistent labeling and low annotation quality: To establish an intelligent model that meets clinical requirements, accurate annotation and professional labeling of the data are required. Lack of rehabilitation-specific clinical labels, such as Brunnstrom stages, Fugl-Meyer scores, or fatigue levels, will hinder the model’s ability to assess rehabilitation outcomes and predict recovery stages.

To address the aforementioned limitations, we present REHAB, a wearable sensor-based kinematic dataset designed for stroke rehabilitation research. Compared with existing datasets, REHAB is characterized by a streamlined data acquisition setup, clinically grounded data collection, and comprehensive coverage of rehabilitation-related movements. In contrast to many previous IMU-based datasets that were primarily collected from healthy participants performing activities of daily living (ADLs), REHAB specifically focuses on post-stroke patients with hemiplegic motor impairments and rehabilitation-oriented movement tasks. All data in the dataset were collected from stroke patients undergoing real clinical rehabilitation procedures, thereby providing movement characteristics that more accurately reflect pathological motor patterns and rehabilitation needs in clinical practice. Furthermore, unlike conventional ADL-oriented datasets, the movements included in REHAB are entirely composed of clinically relevant rehabilitation assessment and rehabilitation training tasks, making the dataset more suitable for rehabilitation assessment, motor function analysis, recovery monitoring, and intelligent rehabilitation applications. The dataset was acquired using a deployable and easy-to-operate sensing system composed of five wearable sensor nodes, integrating inertial measurement units (IMUs) and flex sensors (FSs). This configuration enables the collection of sufficient kinematic information while maintaining a low hardware burden, facilitating potential deployment in both clinical and home-based rehabilitation settings. REHAB comprises real-world data collected from a randomized controlled clinical trial involving 120 post-stroke patients with motor impairments over a two-week rehabilitation program. All recordings were obtained in clinical environments and accompanied by standardized clinical labels, ensuring data reliability and clinical relevance. The dataset includes kinematic recordings from 27 standardized rehabilitation assessment movements and 16 rehabilitation training movements, covering a broad range of clinically relevant motor functions. While a subset of the assessment-related kinematic data has been utilized in our previous study48,49 for intelligent rehabilitation assessment, these data have not been publicly released. In contrast, the kinematic data corresponding to the rehabilitation training movements are reported here for the first time. By integrating data from both rehabilitation assessment and training scenarios, REHAB provides a unified resource for studying motor assessment, recovery progression, and rehabilitation outcome analysis.

The remainder of this Data Descriptor is organized as follows. The Methods section details the clinical trial design, sensor system configuration, experimental protocol, and data preprocessing procedures. The Data Records section describes the dataset organization, data formats, and variable definitions. The Technical Validation section evaluates data quality and reliability through systematic validation analyses.

More at link.