According to the Korea Disease Control and Prevention Agency (KDCA), the number of patients suffering from Stroke in South Korea has been steadily increasing. Patients who experience hemiparesis after a stroke often experience functional decline in da...
According to the Korea Disease Control and Prevention Agency (KDCA), the number of patients suffering from Stroke in South Korea has been steadily increasing. Patients who experience hemiparesis after a stroke often experience functional decline in daily activities, which necessitates long-term rehabilitation. While the demand for rehabilitation treatment for stroke patients has increased, the supply of such services remains insufficient. To address this, there is a growing need for remote monitoring of stroke patients in order to expedite recovery, provide rehabilitation services beyond specialized medical institutions, and alleviate the burden on healthcare professionals. A key component of remote monitoring is the recognition of the patient's movements. However, existing motion recognition technologies in the field of human activity recognition are designed for able-bodied individuals, making them inadequate for classifying movement data of hemiparetic stroke patients, whose movement patterns are significantly different. Furthermore, in the research on remote rehabilitation systems, the evaluation of these systems has often been based on publicly available movement data from able-bodied individuals, which makes it difficult to confirm the practical effectiveness for stroke patients.
In response, this study selects 15 indoor rehabilitation movements specifically for patients with upper limb hemiparesis after a stroke and develops a system that simultaneously collects data using IMU sensors and RGB cameras while presenting these movements to the user. Experiments using the developed system were conducted to collect movement data from both able-bodied individuals and acute/subacute stage upper-limb hemiparetic stroke patients performing the same movements.
In the field of motion recognition models, to maximize the recognition accuracy of the multimodal ensemble model, we proposed a model structure and data preprocessing method aimed at optimizing the accuracy of sensor-based unimodal human activity recognition models. Specifically, by reshaping the multivariate time-series data collected from IMU sensors into a form similar to 3-channel 2D image data, we were able to take advantage of the lower computational load of sensor data while utilizing the ViT (Vision Transformer) model for classification, which enhanced the motion recognition accuracy.
In the data domain, we compared data from upper-limb hemiparetic stroke patients with that from able-bodied individuals and found that the stroke patients' movement data differs significantly from that of able-bodied individuals. Consequently, models that show high classification accuracy in human activity recognition for able-bodied individuals were found to be inadequate for stroke patient data. Additionally, we proposed a ViT model and time-series data preprocessing method that captured the movement characteristics of both stroke patients and able-bodied individuals, achieving stable classification accuracy.
Furthermore, we developed a multimodal ensemble deep learning model that combines skeleton data extracted from RGB video with motion data, and confirmed that it achieved high classification accuracy when trained on both stroke patient and able-bodied data. Unlike unimodal models, the multimodal model was able to compensate for the weaknesses of each modality, improving overall performance.
The findings of this research have the potential to lead to the development of a system in which upper-limb hemiparetic stroke patients can autonomously perform rehabilitation exercises without the assistance of a professional therapist. This system could not only focus on motion recognition but also enable future research to assess the quality of rehabilitation movements of stroke patients using skeleton or IMU sensor data. It is believed that this research will contribute significantly to meeting the rehabilitation needs of hemiparetic stroke patients and effectively alleviating the burden on healthcare professionals.