Subspace-based dual quadratic discriminant analysis for motor signal classification using functional near-infrared spectroscopy

Abstract

Functional near-infrared spectroscopy (fNIRS) has been used as a non-invasive modality for brain-computer interfaces (BCIs). Feature extraction is crucial for computationally efficient and performance-effective brain signal classification in fNIRS-BCI. For example, mean, peak, slope, and variance in the oxygenated hemoglobin (HbO) signal were typically used. However, there are limits to these features to exploit all the information in the measurements, and optimizing feature extraction in multi-channel signals is challenging. This paper proposes a subspace-based dual quadratic discriminant analysis (SD-QDA) to address these issues. It comprises dual quadratic and linear terms, and we optimized them for maximum discriminability within the training dataset. The difference between orthogonal projections was utilized in the quadratic terms to mitigate the small sample size problem, enabling the effective use of all multi-channel signals. We validated the method using simulation and experimental datasets (right-hand squeezing and imagining) and showed that it improved the classification accuracy compared to conventional methods. We also demonstrated that using the deoxygenated hemoglobin (HbR) signal in conjunction with HbO significantly improved the performance of the proposed method in the experimental datasets, thereby fully exploiting the fNIRS measurements in potential BCI applications. The proposed method exhibited the (maximum) accuracy for the motor signal (actual movement) at 84.5% on average and 74.7% for the motor signal imagination.