Machine learning performance in EEG-based mental workload classification across task types: a systematic review.

The literature features a variety of tasks and methodologies to induce mental workload (MWL) and to assess the performance of MWL estimation models. Because no standardized benchmark task or set of tasks exists, the comparison of different machine learning (ML) solutions in this field is difficult, as their performance is significantly dependent on these factors. In this paper, we present the first comprehensive examination of ML models' performance in EEG-based MWL classification across task types. To achieve this, we categorized ML studies based on the task type used in their experiments and compared models' performances across these categories. Notably, a significant drop in MWL classification accuracy was observed among the best-performing models in multitasking studies where MWL was rated based on quantitative task load, compared to those in single-tasking studies and studies where MWL was subjectively rated. This points to the inherent challenges associated with estimating MWL in more complex tasks such as multitasking. This is particularly relevant for practical applications, as real-world tasks typically involve some degree of multitasking. By comparing ML models' performances across task types, this review provides valuable insights into the state-of-the-art of EEG-based MWL estimation, highlights existing gaps in the field, and points to open questions for further research.