Data Augmentation Techniques for Accurate Action Classification in Stroke Patients with Hemiparesis

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Science

Background:

• Stroke survivors with hemiparesis need effective home-based rehabilitation.
• Deep learning models for action detection in rehabilitation face challenges due to sparse and heterogeneous data.

Purpose of the Study:

• To investigate data augmentation and model training strategies to improve deep learning-based action detection for stroke rehabilitation.
• To address data sparsity and heterogeneity issues in classifying patient data.

Main Methods:

• Tested three data transformations with varying data volumes to assess classification performance.
• Evaluated transfer learning with a pre-trained one-dimensional convolutional neural network (Conv1D) and the InceptionTime model, incorporating data augmentation.
• Compared joint training strategies using non-disabled (ND) and stroke patient data with rotational augmentation.

Main Results:

• Joint training with ND and rotationally augmented stroke data improved the F1-score to 60.9% (vs. 47.3% baseline) for Conv1D.
• Transfer learning with ND data achieved 60.3% accuracy; joint training with InceptionTime reached 67.2% accuracy.
• Rotational augmentation proved most effective for data with initially lower performance and smaller participant subsets.

Conclusions:

• Joint training on rotationally augmented ND and stroke data enhances classification performance in sparse data scenarios.
• Data augmentation techniques, particularly rotational augmentation, are crucial for improving deep learning models in stroke rehabilitation.
• The findings suggest a viable approach to overcome data limitations in developing effective AI-driven rehabilitation tools.