Enhancing Activity Recognition After Stroke: Generative Adversarial Networks for Kinematic Data Augmentation.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Rehabilitation Robotics

Background:

• Machine learning (ML) models for wearable monitoring in stroke rehabilitation face generalizability issues due to limited and heterogeneous data.
• Existing data augmentation techniques often produce unrealistic synthetic data, hindering classifier performance.
• Accurate patient monitoring is crucial for effective stroke rehabilitation and intervention tailoring.

Purpose of the Study:

• To employ Conditional Generative Adversarial Networks (cGANs) for generating realistic synthetic kinematic data for stroke rehabilitation.
• To enhance the variability and scale of training datasets for ML models used in wearable monitoring.
• To improve the task classification accuracy of ML models by incorporating synthetic data.

Main Methods:

• Utilized a publicly available dataset of stroke survivor reaching movements.
• Employed Conditional Generative Adversarial Networks (cGANs) to generate synthetic kinematic data mimicking real movements.
• Trained deep learning models on both real and synthetic datasets to evaluate classification accuracy.

Main Results:

• Synthetic data generated by cGANs closely mimicked the temporal dynamics and movement patterns observed in stroke survivors.
• ML models trained with both real and synthetic data achieved an overall accuracy of 80.0%.
• Models trained solely on real data achieved a significantly lower accuracy of 66.1%.

Conclusions:

• cGANs effectively generate realistic synthetic kinematic data, significantly enhancing ML model performance in stroke rehabilitation.
• The improved task classification accuracy enables more precise patient progress monitoring.
• This approach offers potential for clinicians to tailor rehabilitation interventions more effectively.