Comparison of Machine Learning and Deep Learning Approaches for Decoding Brain Computer Interface: An fNIRS Study

Abstract

Recently, deep learning has gained great attention in decoding the neuro-physiological signal. However, which one (classical machine learning or deep learning) has better performance for decoding the functional near-infrared spectroscopy (fNIRS) signal is still lack of full verification. Thus, in this paper, we systematically compared the performance of many classical machine learning methods and deep learning methods in fNIRS data processing for decoding the mental arithmetic task. The classical machine learning methods such as decision tree, linear discriminant analysis (LDA), support vector machine (SVM), K-Nearest Neighbor (KNN) and ensemble methods with strict feature extraction and screening, were used for performance comparison, while the long short-term memory-fully convolutional network (LSTM-FCN) method as a representative of deep leaning methods was applied. Results showed that the classification performance of SVM was the best among the classical machine learning methods, achieving that the average accuracy of the subject-related/unrelated were 91.0% and 83.0%, respectively. Furthermore, the classification accuracy of deep learning was significantly better than that of the involved classical machine learning methods, where the accuracy of deep learning could reach 95.3% with subject-related condition and 97.1% with subject-unrelated condition, respectively. Thus, this paper has totally showed the excellent performance of LSTM-FCN as a representative of deep learning in decoding brain signal from fNIRS dataset, which has outperformed many classical machine learning methods.

Classification of Game Demand and the Presence of Experimental Pain Using Functional Near-Infrared Spectroscopy

“…The results illustrated that: (1) SVM yielded the highest level of accuracy (66.4%) from fNIRS features, (2) maximum accuracy for HR features was 53.9% using a NB classifier, and (3) only the SVM classifier using fNIRS features achieved a level of classification performance that was significantly above chance levels. This absolute level of accuracy is somewhat lower than similar laboratory-based, subject-independent, binary classifications of mental workload reported in earlier studies, e.g., 0.83 (Lu et al, 2020), 0.84 (Naseer et al, 2016), and applied tasks, e.g., 0.71 (Benerradi et al, 2019), 0.80 (Gateau et al, 2015); however, direct comparisons between the current study and related research are problematic, as earlier experiments manipulated workload using task simulation (e.g., aviation) or standardised laboratory tasks, as opposed to a computer game. In general, the classification of game demand using heart rate features compared poorly with features derived from fNIRS for all classifiers except NB, indicating the superiority of fNIRS features for classification of game demand, with the caveat that the number of features selected for the fNIRS-based model was significantly higher.…”

“…However, other researchers working with fNIRS data have reported superior classification with artificial neural networks (Naseer et al, 2016) and deep learning techniques, such as fully convolutional networks (Lu et al, 2020); the latter reported an accuracy level above 97% for subjectindependent classification of cognitive demand with deep learning. A subject-independent, unsupervised approach could be explored in future work, especially if a large training dataset could be generated that produced a high level of subject-independent classification.…”

Explainable artificial intelligence model to predict brain states from fNIRS signals

“…As data scarcity is one limiting factor of Deep Learning techniques (Ghonchi et al, 2020 ; Lu et al, 2020 ; Nagabushanam et al, 2020 ) the sliding window also augments the data by increasing the data’s sample size, as explained in Section “Proposed sliding window approach”.…”

“…A recent study by Z. Shi et al (Lu et al, 2020) comparing machine learning models like SVM, KNN, and LDA with deep learning models like long short-term memory-fully convolutional network (LSTM-FCN), found that deep learning 10.3389/fnhum.2022.1029784 models could achieve a decode mental arithmetic task with a classification accuracy of 97%.…”

Deep learning in fNIRS: a review

“…Leave-one-super-trial-out is another available rigorous CV technique 94 . Many of the reviewed papers carry out a 10-fold CV, whereas a few execute LOSO CV 48 , 49 . The most common metrics used for evaluation are accuracy, specificity, and sensitivity.…”