TLDR: Researchers have developed an AI-based system, RAST-G@, for home-based stroke rehabilitation. It uses an RGB-D camera and wearable sensors to capture movements, a mobile app for guidance, and an AI server for assessment. The RAST-G@ model, utilizing spatio-temporal graph convolutional networks and temporal attention, evaluates exercise quality based on physiotherapist-annotated data. Tested on KIMORE and a new NRC dataset, the system shows improved accuracy over baselines. It provides user-friendly feedback, including progress trends and detailed movement heatmaps, offering a scalable solution for quantitative and consistent domiciliary rehabilitation.
Stroke recovery is a journey that often requires continuous and consistent rehabilitation, ideally integrated into a patient’s daily life at home. However, without the constant supervision and expert feedback of a physiotherapist, performing exercises correctly and effectively can be challenging, potentially leading to improper movements or even injury. Addressing this critical need, researchers Suhyeon Lim, Ye-eun Kim, and Andrew J. Choi have developed an innovative AI-based system designed to bring expert-level stroke rehabilitation assessment directly into the home environment.
The proposed system, detailed in their research paper AI-Based Stroke Rehabilitation Domiciliary Assessment System with ST-GCN Attention, offers a comprehensive solution for home-based rehabilitation. It comprises three main components: a hardware setup featuring an RGB-D camera and wearable sensors to accurately capture movement, a mobile application that guides users through exercises, and an AI server responsible for assessing performance and providing feedback.
When a stroke patient performs exercises guided by the mobile app, the system records detailed skeleton sequences of their movements. These sequences are then analyzed by a sophisticated deep learning model named RAST-G@. This model is built upon a spatio-temporal graph convolutional network (ST-GCN), which is particularly adept at extracting intricate features from skeletal data, understanding how different body joints move and interact. To further enhance its assessment capabilities, RAST-G@ integrates a transformer-based temporal attention mechanism. This allows the model to focus on crucial moments and patterns within the continuous flow of movement, effectively determining the quality of the action performed.
A significant aspect of this research involved creating a specialized dataset, known as the NRC dataset. This dataset includes 10 upper-limb activities of daily living (ADL) and 5 range-of-motion (ROM) exercises. Data was collected from both stroke patients and non-disabled participants, with licensed physiotherapists providing expert score annotations for each movement. This patient-centered approach to assessment is crucial, as it recognizes that the goal of neurological rehabilitation is functional improvement and independence, rather than simply matching the movements of a non-disabled individual.
The effectiveness of RAST-G@ was rigorously tested on both the KIMORE and NRC datasets. The results demonstrated that the model significantly improved performance compared to existing baseline models across various metrics, including Mean Absolute Deviation (MAD), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE. This indicates that the system can provide quantitative and consistent assessments, even for complex and varied rehabilitation movements.
Beyond just assessment, the system is designed to provide user-friendly feedback. The mobile application offers two types of feedback: period feedback and discrete feedback. Period feedback summarizes performance trends over time, showing monthly progress and allowing users to track improvements across different exercise categories. Discrete feedback, on the other hand, provides immediate results for each exercise, including visual aids like skeleton heatmap visualizations that highlight the most influential joints during the movement, a task-specific score, and optional comments from therapists. This intuitive feedback helps patients understand their performance and allows clinicians to monitor progress and adjust treatment plans effectively.
Also Read:
- Advancing Gesture Recognition: A Robust Pipeline for Dynamic Arm Signals
- Brain-Powered Mobility: A New Era for Wheelchair Control
In conclusion, this AI-based system represents a scalable and robust approach to domiciliary stroke rehabilitation assessment. By providing quantitative, consistent, and patient-centered evaluations, it has the potential to significantly enhance the quality and accessibility of rehabilitation, empowering patients to continue their recovery journey from the comfort of their homes.
