A Novel Model to Generate Heterogeneous and Realistic Time-Series Data for Post-Stroke Rehabilitation Assessment

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Rehabilitation Science

Background:

• Machine learning in tele-rehabilitation is limited by insufficient data.
• Generative Adversarial Networks (GANs) are used for data augmentation but suffer from mode collapse, failing to capture complete data distributions.
• Mode collapse in GANs hinders the generation of diverse and representative synthetic data for training rehabilitation models.

Purpose of the Study:

• To address the mode collapse issue in GAN-based data augmentation for post-stroke tele-rehabilitation.
• To propose and evaluate a novel Time Series Siamese GAN (TS-SGAN) for improved synthetic data generation.
• To enhance the accuracy of post-stroke assessment models through better data augmentation.

Main Methods:

• Applied a standard GAN to post-stroke rehabilitation datasets, confirming mode collapse.
• Developed and implemented a Time Series Siamese GAN (TS-SGAN) incorporating a Siamese network and an additional discriminator.
• Utilized Longest Common Sub-sequence (LCSS) for data uniformity analysis and Gramian Angular Field (GAF) with ResNet-18 for classification accuracy evaluation.

Main Results:

• The standard GAN exhibited mode collapse, generating limited data variations.
• TS-SGAN demonstrated uniform data generation across testing datasets, overcoming the mode collapse issue.
• Classification accuracy using ResNet-18 increased by 35.2%-42.07% with TS-SGAN generated data compared to the original GAN.

Conclusions:

• TS-SGAN effectively resolves mode collapse in GAN-based data augmentation for tele-rehabilitation.
• The proposed TS-SGAN significantly improves the performance of post-stroke assessment models.
• This advancement offers a promising solution for data scarcity challenges in machine learning-driven rehabilitation.