| dc.description.abstract | Working performance is closely related to human mental workload (MWL). Today's researchers are paying more and more attention to implicit brain-computer interface (implicit BCI) that automatically adapts to user mental states to avoid critical risks and improve work outcomes. Some of those investigations have demonstrated functional Near-infrared spectroscopy (fNIRS) as one potential non-invasive system for human MWL analysis in both off-line and on-line manners. In this study, we developed two studies that focused into the applications of supervised and unsupervised machine learning techniques of fNIRS parameters to support human workload classification. In the experimental setup, five subjects engaged in ten-loop memorizing tasks that were devised into two MWL levels while fNIRS signals were being monitored over their frontal lobes. Statistical analysis of the signals demonstrated the regional activations over prefrontal cortex (PFC) while subject was performing two graded levels of MWL. The averages of fNIRS parameters also indicated the statistical differences according to two MWL levels.
In the first study, supervised classification framework based on principle component analysis (PCA) and support vector machine (SVM) are applied for data dimensional reduction and classification. Classification features included the PCA projections of oxyHemoglobin (HbO), deoxyHemoglobin (HbR) and their derived derivatives differences (Diff). Inputs to classification were either univariate features (the projected signals) or multivariate features (the combinations of different univariate features). Classification with SVM suggested that fNIRS parameters monitored over PFC could be applied to quantify two MWL levels effectively and achieved classification results higher than conventional methods with mean features and 3NN classifier. In conclusion, PCA-SVM framework combined with voting method on multivariate features obtained the highest classification performance.
In the second study, unsupervised techniques were applied to extract the basis functions of fNIRS signals and used them to extract features for classification. Two fundamental unsupervised leaning including principle component analysis (PCA) and independent component analysis (ICA) were applied on a set of unlabeled random fNIRS data to extract the basis functions. Then two-dimensional convolutional matrices, which are sets of convolutional coefficients of the input signal with learned basis functions, are implemented as the inputs for MWL classification using convolutional neural network (CNN). Study of generalized linear model demonstrated that basis functions extracted using ICA and PCA are more effective when illustrating the activation regions over PFC than previous using available constructed hemodynamic response functions (HRF). Besides, ICA basis functions demonstrated the sparseness and showed its potential for further study of fNIRS based on its contained basis functions. | en_US |
