CBSA: A Deep Transfer Learning Framework for Assessing Post-Stroke Exercises

A stroke occurs when the brain's blood flow is disrupted, leading to cell death. Post-stroke patient rehabilitation helps patients to recover from the impairments produced by stroke by using physical therapy. In an inpatient rehabilitation center, the patient performs these therapies under the close watch of medical professionals. However, due to the scarcity of resources and expensive therapeutic services, people are getting demotivated to attend inpatient rehabilitation centers. With recent technological advancements, a person's skeletal joint movement data can be recorded using sensors (e.g., Vicon, Kinect), and future applications powered by artificial intelligence can evaluate the effectiveness of the exercises. However, applications evaluating exercises face challenges due to limited amounts of annotated data, diverse use of sensors, and various movements. In this work, we proposed a novel transfer learning-enabled deep learning model named CBSA
^1
1
Abbreviations: CBSA - CNN-BiLSTM-SelfAttention, which effectively addresses the challenges faced by applications in real-world exercise evaluation. We evaluated our model's efficiency using the publicly available UI-PRMD dataset and achieved an average MAE of 0.0071, which outperformed the state-of-the-art model. Moreover, our model provided 37% performance improvement compared to the state-of-the-art model. Utilizing our model applications would be able to determine the quality of exercises by the patients more proficiently, which would help them achieve quick recovery.