Measurement Criteria for Neurophysiological Monitoring and Neuroadaptive Interfaces

Human Factors and Psychology have long been interested in the neurophysiological basis of mental states such as mental workload or fatigue due to their relevance to safety and well-being at work. With the recent emergence of Neuroergonomics, research on the human mind in applied settings is becoming more commonplace in an attempt to improve ecological validity and to address the limitations of artificial laboratory environments. This thesis dealt with the applied use of neurophysiological metrics for continuous mental state monitoring, a rapidly growing area thanks to affordable consumer sensing technologies such as smartwatches and eye-trackers. However, neuroimaging techniques like Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) remain largely confined to research contexts. A key question this thesis aimed to answer was the viability of wearable neuroimaging sensors in comparison to lab-grade devices for mental state monitoring. Secondary aims concerned the exploration of novel task-irrelevant probing techniques for mental workload monitoring and ensemble machine learning strategies to overcome the challenge of cross-subject generalisability in mental state monitoring. To achieve this, a two-day experimental paradigm was designed, collecting multimodal data from 80 participants across four datasets using various sensor configurations across three levels of task load in two different tasks. The first was an artificial working memory task called the n-back, and the second was a more ecologically valid multi-tasking paradigm called the Multi-Attribute-Test-Battery. The results revealed task-dependent variations in traditional neurophysiological workload metrics and highlighted the impact of aperiodic contributions to canonical EEG band power metrics, demonstrating varying effects on different frequency ranges, which could clarify some conflicting findings in the existing literature. Additionally, the thesis identified and estimated biases in passive Brain-Computer Interface evaluation methods, which complicate reproducibility in the field. Most importantly, it provided evidence that 64-channel lab-grade EEG and sparse seven-channel sponge-based EEG can achieve comparable mental workload classification performance, along with strategies to address accuracy differences when they occur. However, across the four datasets collected, subtle differences in mental workload - such as between a 0-back and a 1-back task, or the low-effort operation of the MATB compared to medium effort scenarios - proved difficult to distinguish with state-of-the-art methods, with accuracy levels around 60% rather than the 80% - 90% for more extreme workload differences, regardless of sensing hardware. Overall, the results offered evidence for the potential of lightweight, user-friendly EEG devices for neuroadaptive technologies and provided several future research directions towards robust cross-subject mental workload monitoring.