Exploring Machine Learning Approaches for Classifying Mental Workload using fNIRS Data from HCI Tasks

Functional Near-Infrared Spectroscopy (fNIRS) has shown promise for being potentially more suitable (than e.g. EEG) for brain-based Human Computer Interaction (HCI). While some machine learning approaches have been used in prior HCI work, this paper explores different approaches and configurations for classifying Mental Workload (MWL) from a continuous HCI task, to identify and understand potential limitations and data processing decisions. In particular, we investigate three overall approaches: a logistic regression method, a supervised shallow method (SVM), and a supervised deep learning method (CNN). We examine personalised and generalised models, as well as consider different features and ways of labelling the data. Our initial explorations show that generalised models can perform as well as personalised ones and that deep learning can be a suitable approach for medium size datasets. To provide additional practical advice for future brain-computer interaction systems, we conclude by discussing the limitations and data-preparation needs of different machine learning approaches. We also make recommendations for avenues of future work that are most promising for the machine learning of fNIRS data.