fNIRS-GANs: data augmentation using generative adversarial networks for classifying motor tasks from functional...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Functional near-infrared spectroscopy (fNIRS) is a key technology for brain-computer interfaces (BCIs).
• Acquiring sufficient fNIRS data for training classification models is challenging due to participant discomfort during lengthy measurements, leading to reduced BCI accuracy.

Purpose of the Study:

• To enhance fNIRS-BCI accuracy by investigating a novel data augmentation technique.
• To evaluate the efficacy of Wasserstein generative adversarial networks (WGANs) in generating artificial fNIRS data for improved classification performance.

Main Methods:

• Employed WGANs to generate artificial fNIRS data from measurements acquired during hand-grasping tasks.
• Utilized support vector machines and simple neural networks to assess classification performance with augmented datasets.
• Compared the characteristics of WGAN-generated fNIRS data with measured data.

Main Results:

• WGAN-generated fNIRS data exhibited temporal profiles similar to measured data, with an additional noise component.
• Augmenting training datasets with WGAN-generated data significantly improved classification accuracy for four distinct task types.
• Accuracy improvements were observed irrespective of the classification algorithms employed.

Conclusions:

• The WGAN-based data augmentation method effectively generates useful artificial fNIRS data.
• This approach holds promise for improving the performance and practical applicability of fNIRS-BCI systems.