Transformer Model for Functional Near-Infrared Spectroscopy Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Functional near-infrared spectroscopy (fNIRS) is a key neuroimaging technique for brain-computer interfaces (BCI).
• fNIRS signal classification remains a central challenge in BCI development.
• Transformer models have shown great success in natural language processing and computer vision.

Purpose of the Study:

• To propose novel Transformer-based networks (fNIRS-T and fNIRS-PreT) for improved fNIRS classification.
• To enhance data utilization and network representation capacity for fNIRS signals.
• To develop an end-to-end network for fNIRS classification without traditional preprocessing steps.

Main Methods:

• Developed fNIRS-T, a Transformer-based network leveraging spatial-level and channel-level representations.
• Introduced fNIRS-PreT, an end-to-end network incorporating a novel preprocessing module (1D average pooling and layer normalization).
• Evaluated models on three open-access fNIRS datasets, comparing against traditional classifiers, CNN, and LSTM.

Main Results:

• fNIRS-T achieved the highest accuracy in ternary classification tasks (75.49%) compared to CNN (72.89%) and LSTM (61.94%).
• fNIRS-T significantly outperformed traditional classifiers, showing at least 27.41% higher accuracy than statistical features.
• fNIRS-PreT demonstrated competitive performance using raw data, highlighting the effectiveness of the end-to-end approach.

Conclusions:

• The proposed Transformer-based models (fNIRS-T and fNIRS-PreT) significantly advance fNIRS-based BCI performance.
• Exploring spatial and channel-level representations is crucial for maximizing fNIRS data utility.
• The end-to-end fNIRS-PreT offers a simplified yet effective approach to fNIRS signal processing and classification.