Deep Learning-Based Image Automatic Assessment and Nursing of Upper Limb Motor Function in Stroke Patients

Area of Science:

• Rehabilitation Medicine
• Biomedical Engineering
• Artificial Intelligence

Background:

• Assessing upper limb mobility after stroke is crucial for patient recovery.
• Traditional clinical assessments can be subjective and time-consuming.
• Objective and automated methods are needed for accurate and efficient evaluation.

Purpose of the Study:

• To develop and validate an automated system for upper limb mobility assessment post-stroke.
• To utilize Kinect sensor technology for precise tracking of upper limb bone points.
• To construct and evaluate a Gated Recurrent Neural Network (GCRNN) model for predicting Fugl-Meyer Assessment (FMA) scores.

Main Methods:

• Clinical assessment knowledge of upper limb mobility was reviewed.
• A Kinect sensor was employed for spatial location tracking of upper limb bone points.
• A Gated Recurrent Neural Network (GCRNN) model was constructed and trained.
• Experimental data acquisition involved setting up a unique environment and evaluation tasks based on FMA items.
• Comparative experiments were conducted against LSTM and CNN deep learning algorithms.

Main Results:

• The GCRNN model successfully predicted FMA scores using upper limb bone point data.
• Optimal prediction performance (coefficient of determination = 0.89) was achieved using a combination of specific tasks (task 1, 2, and 5).
• The proposed GCRNN method demonstrated superior performance compared to LSTM and CNN algorithms.

Conclusions:

• The GCRNN model effectively extracts spatial and temporal motion features of the upper limb.
• Automated assessment using GCRNN and Kinect sensor technology offers a promising approach for post-stroke upper limb mobility evaluation.
• The developed system achieved high prediction accuracy, indicating its potential clinical utility.