Enhancing rehabilitation in stroke survivors: a deep learning approach to access upper extremity movement using...

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Artificial Intelligence in Healthcare

Background:

• Upper extremity (UE) rehabilitation is vital for stroke survivors' independence.
• Current assessment methods lack objectivity and real-world applicability.
• Accurate UE performance measurement is needed for effective rehabilitation.

Purpose of the Study:

• To apply deep learning models (CNNs with Dense layers) to accelerometry data for classifying UE movements in stroke survivors.
• To evaluate both intrasubject and intersubject models for UE movement classification.
• To assess the impact of including non-paretic arm data on classification accuracy.

Main Methods:

• Utilized wrist-worn accelerometers to collect raw data on UE movements.
• Developed Convolutional Neural Networks (CNNs) combined with Dense layers for movement classification.
• Trained and tested intrasubject and intersubject deep learning models.

Main Results:

• The intrasubject model achieved 0.90 ± 0.05 accuracy for paretic UE movements.
• The intersubject model reached 0.79 ± 0.06 accuracy, improving to 0.88 ± 0.10 with non-paretic arm data.
• The deep learning approach eliminated manual feature extraction, outperforming previous methods.

Conclusions:

• Deep learning with accelerometry provides an objective, cost-effective method for monitoring UE function post-stroke.
• This approach enhances the ability to track real-world UE use.
• Findings support personalized rehabilitation strategies and improved clinical practice for stroke survivors.