TL;DR
This paper introduces an intelligent agent designed to assist in stroke rehabilitation by providing automated assessment and guidance for exercises. By combining expert knowledge with data-driven models, the agent aims to improve the effectiveness of in-home rehabilitation and provide therapists with detailed performance reports.
What Problem Does This Paper Address?
Stroke survivors often require extensive rehabilitation to regain physical abilities. While in-home rehabilitation regimens are common due to therapist availability constraints, patients can struggle with correct exercise execution and tracking progress. Existing computer-assisted systems are often passive, lacking real-time feedback and personalized guidance, and generate overwhelming amounts of data that are difficult to interpret. This research tackles the challenge of creating an active, intelligent system that can bridge the gap between therapist expertise and patient needs during unsupervised rehabilitation.
Methods
The core of this research is an intelligent agent that integrates two key components:
- Interactive Machine Learning for Assessment: This approach moves beyond simple detection of incorrect movements to quantify performance scores. It uses Reinforcement Learning with Double Q-learning to train an agent that interactively selects salient features for assessment. This allows the agent to learn which features are most important for evaluating an exercise, making the assessment process more efficient and interpretable. The agent then uses these features to generate an interpretable performance report and personalized suggestions for improvement.
- Robotic Rehabilitation Guidance: A guidance model is learned using an Encoder-Recurrent-Decoder approach. This model predicts desirable sequential joint positions for exercises, leveraging demonstrations of an unaffected side's movements. The guidance is provided at a high-level, directional instruction, mimicking how therapists often guide patients. The agent, equipped with both assessment and guidance models, is intended to be deployed on a Socially Assistive Robot.
Datasets
The study utilized data collected from:
- 15 stroke patients with varying functional abilities.
- 11 healthy subjects.
Exercises included three task-oriented, upper-limb stroke rehabilitation movements: 'Bring a cup to the mouth', 'Switch a light on', and 'Move forward a cane'. Data was collected using a Kinect v2 sensor, capturing joint trajectories and video frames at 30 Hz. Patients performed 10 repetitions per exercise (affected and unaffected sides), and healthy subjects performed 15 repetitions (dominant side). Ground truth scores were obtained from two therapists who evaluated recorded videos based on three performance components: 'Range of Motion', 'Smoothness', and 'Compensation'.
Key Insights and Ideas
- Bridging the Gap: The research proposes a novel way to combine expert domain knowledge with data-driven models, creating a more comprehensive and interpretable rehabilitation tool.
- Interpretable AI for Rehabilitation: Moving beyond "black box" models, the interactive machine learning approach aims to make the assessment process transparent, aiding therapists in understanding patient performance.
- High-Level Guidance: The guidance model's focus on directional joint instructions, rather than precise coordinates, aligns with common therapeutic practices and simplifies the learning process.
- Personalized Feedback: The agent's ability to generate personalized suggestions for improvement holds the potential to significantly enhance patient engagement and outcomes.
Key Results
The paper reports promising results from preliminary experiments:
- The interactive assessment approach performed significantly better than basic interpretable models and non-interactive neural network approaches (p < 0.05).
- Previous work by Lee et al. (2019) demonstrated decent agreement levels (0.7571 - 0.8254 F1-scores) for automatically assessing exercises and relevance with clinically validated Fugl Meyer scores (R² = 0.95).
Limitations and Future Work
- Validation with Therapists: The authors acknowledge the need for future studies with therapists to explore whether automatically transcribed reports can offer new insights for decision-making.
- User Studies: A planned future step involves conducting user studies with therapists and post-stroke survivors to analyze the usability of the integrated system.
- Learning Entire Motion Space: While learning the entire range of possible motions is challenging, the approach leverages prior motion history to reduce the complexity of predicting next possible motions.
- Discrete Action Space: The use of an eight-direction action space for joint guidance is assumed to be sufficient, but its effectiveness across a wider range of exercises and patient conditions will require further validation.
Summary updated