Classification of three categories from prefrontal cortex using LSTM networks: fNIRS study

In the conventional neuroscience studies using functional near infrared spectroscopy (fNIRS), the region of interest (ROI) needs to be investigated. In this paper, to reduce the classification and processing time without classification accuracy loss, we have investigated the use of a long short-term memory (LSTM) networks in the fNIRS data analysis. Three mental tasks including mental arithmetic, mental counting, and puzzle solving were performed to obtain fNIRS data from the prefrontal cortex. To validate our proposed method, we have compared our method with the conventional classifiers, that is, linear discriminant analysis (LDA) and support vector machine (SVM), and feature extraction methods. Using the conventional methods, the averaged classification accuracies were 37.50% with LDA and 37.96% with SVM. However, in the case of LSTM, the maximum classification accuracy was 83.3%. The results show that LSTM networks can be used as a fNIRS classifier. Also, our results represent the possibility of online classification in fNIRS studies, which does not need the
t
-test process using the expected hemodynamic response.