Classifying Common Knee Rehabilitation Exercise Mistakes Using IMU Data

Physical therapy following major surgeries is a branch of medicine that has seen its fair share of technologically inspired advances. One important facet of physical therapy, the “at-home exercises” patients are prescribed to do, is still somewhat of a “black box” to many physical therapists (PTs). PTs have no way of knowing (1) whether the patient is doing the home exercises, or (2) whether the patient is doing the exercises in the correct and healthy manner. This lack of awareness makes it difficult for the PT to guide the patient, which can often lead to prolonged rehabilitation periods or (sometimes) can create life-long health problems for patients. In this thesis, we provide a means for a PT to remotely monitor patient’s performance of at-home exercises. We combined the capabilities of wearable motion sensors with computational algorithms to provide patients feedback on the quality of their performed exercises. We evaluated this approach by asking 20 healthy volunteers to perform popular knee-rehabilitation exercises with various mistakes while wearing motion sensors. After preprocessing and extracting features from the sensor data, we trained machine-learning models on the extracted features. The models showed a high rate of accuracy during testing, which brings us a step closer to giving physical therapists and doctors a tool to automatically and objectively classify certain exercises and mistakes made during those exercises