System Derived Spatial-Temporal CNN for High-Density fNIRS BCI.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• High-density (HD) functional Near-Infrared Spectroscopy (fNIRS) offers rich spatial information for brain-computer interfaces (BCIs).
• Extracting both spatial and temporal features is crucial for accurately interpreting brain activity in fNIRS data.
• Existing methods may not fully leverage the spatial resolution of HD fNIRS for complex BCI tasks.

Purpose of the Study:

• To propose and demonstrate an intuitive and generalizable approach for spatial-temporal feature extraction using HD fNIRS.
• To enhance motor-task classification in brain-computer interfaces (BCIs).
• To leverage Frequency-Domain (FD) fNIRS signals for improved BCI performance.

Main Methods:

• Utilized a HD fNIRS probe design to generate layered topographical maps of Oxy/deOxy Haemoglobin changes.
• Developed and trained a 3D convolutional neural network (CNN) to simultaneously extract spatial and temporal features.
• Employed a mixed-subjects training scheme for evaluating classification performance.

Main Results:

• The proposed spatial-temporal CNN effectively exploited spatial relationships in HD fNIRS data.
• Achieved an average F1 score of 0.69 across seven subjects in motor-task classification.
• Demonstrated improved subject-independent classification compared to a standard temporal CNN.

Conclusions:

• The developed spatial-temporal CNN approach is effective for feature extraction in HD fNIRS-based BCIs.
• This method enhances the classification of functional haemodynamic responses, particularly in subject-independent scenarios.
• The approach shows promise for advancing the capabilities of fNIRS-based brain-computer interfaces.