Estimation of force direction from functional near-infrared spectroscopy signals using sparse logistic regression

The brain-machine interface (BMI) has been used as a communication tool for a person who has lost body function. Extracting functional information from brain signals is important for controlling a BMI in a realistic and natural way. For a BMI, a pattern classification algorithm, such as linear discriminant analysis (LDA) and support vector machine (SVM), has commonly been used. However, the classifier using brain signals tends to suffer from overfitting because there are too many obtained features compared with the number of samples. On the other hand, sparse logistic regression (SLR), which has been proposed as a new pattern classification method for brain signals, can select small number of features to classify and interpret brain functions. Thus, overfitting can be prevented using SLR. In this study, we measured functional near-infrared spectroscopy (fNIRS) signals during isometric arm movements in four directions and performed direction classification. The features to classify force direction were selected from obtained data sets using SLR and were used in a SVM. We compared the types of fNIRS signals (OxyHb and DeoxyHb) and feature selection methods. As a result, the classification accuracy was highest when both OxyHb and DeoxyHb were used as the features and both time and channel were selected. The peak time of the signal, when the task ends, and a few seconds after the task ends, were particularly well selected.