Learning to assess the quality of stroke rehabilitation exercises
summary
An automated system uses a Kinect sensor to record 16 joints and extracts kinematic features for range of motion, smoothness, and compensations, summarizes them with frame statistics, trains binary classifiers (NN, decision tree, SVM, LSTM) and applies a threshold model to turn classifier outputs into ordinal scores. Data: 11 healthy + 15 post-stroke subjects, 485 trials; agreement with the primary therapist achieves F1 = 0.8436, 0.8264, 0.7976 on three tasks, the threshold model matches or outperforms multi-class/regression, and summed scores correlate R2 = 0.95 with Fugl-Meyer.
AI-written summary · about
chapters (11)
- 0:00Introduction to stroke rehab assessment
- 1:31Computer support in rehabilitation
- 2:37Research goals and approach
- 3:20Study design and exercises
- 4:43Performance components explained
- 5:26Automated assessment method
- 6:33Feature extraction process
- 7:17Classifier and threshold model
- 8:20Validation study details
- 9:23Binary classification results
- 10:31Model performance comparison
related talks
28mEnhancing Upper Limb Rehabilitation with Unsupervised Machine Learning Models
33mAI Healthcare – A Cybernetic Perspective
22mUsing data to improve physical therapy for people with neurological diseases
47mMachine Learning for medical rehabilitation
31mData as an instrument for more efficient collaboration with robots
22mSeeing, Sensing and Recognizing Laban Movement Qualities
44mFrom Athlete to Algorithm: Transforming Canoe Technique Analysis with AI
28mDeep Learning-Based Human Activity Recognition for Physical Skills...