VR-Aided Ankle Rehabilitation Decision-Making Based on Convolutional Gated Recurrent Neural Network.

Area of Science:

• Neurology
• Rehabilitation Engineering
• Artificial Intelligence

Background:

• Stroke rehabilitation for ankle impairments traditionally relies on physician expertise.
• Current methods face challenges in online decision-making and progress assessment.
• Existing fuzzy neural network approaches require enhancement for complex rehabilitation scenarios.

Purpose of the Study:

• To develop a novel Virtual Reality (VR)-aided ankle rehabilitation decision-making model.
• To leverage a convolutional gated recurrent neural network (C-GRNN) for enhanced assessment.
• To provide tailored rehabilitation parameters based on patient recovery stages.

Main Methods:

• Utilized wearable motion inertial sensors to collect data (range of motion, velocity, jerk, performance scores) during VR rehabilitation.
• Employed data augmentation techniques to address limited dataset challenges.
• Simulated five rehabilitation stages based on the Brunnstrom staging scale.
• Compared C-GRNN performance against Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks.

Main Results:

• The optimized C-GRNN achieved a high classification accuracy of 99.16% and a Macro-F1 score of 0.9786.
• Demonstrated a strong correlation (r > 0.9) with clinical rehabilitation expert assessments.
• Outperformed both CNN and LSTM models in classification performance.

Conclusions:

• The VR-aided C-GRNN model offers an effective and accurate approach for ankle rehabilitation decision-making in stroke patients.
• The model shows significant potential for real-world application in personalized rehabilitation.
• This AI-driven method enhances objective assessment and supports clinical decision-making in stroke recovery.