Subject Dependent and Subject Independent Classification of Mental Workload with Electroencephalographic Brain Computer Interfaces

Mental workload assessment is important to optimize human performance in critical settings such as medical interventions. Approaches to real time assessment include electroencephalographic (EEG) brain-computer interfaces (BCI). However, a current problem is the ability of models of data to generalize well to new users, tasks, or data in different time points. Here we present a solution to generalization of task transfer to new subjects – subject independent task transfer. We follow up previous results and findings for classification of mental workload in subject dependent task transfer with EEG brain signals (Gallegos Ayala et al., 2023), and successfully extend, apply and implement our methods and approaches to preprocessing –filtering, trimming and scaling–, feature extraction based on averages of negative voltage amplitudes –HAN (“Highest upper Alpha Negative”) for upper alpha, and HTN (“Highest Theta Negative”) for theta band filtered subsets– with a leave one out analysis to determine feature vectors with the same dataset, and obtain 99.54 and 93.14% classification performances between low and no workload, and 99.41 and 93.26% between high and no workload for HTN and HAN, respectively, all significantly different from chance with five-dimensional feature vectors, and similar results, significantly different from chance, for HTP and HAP based on averages of positive amplitudes. With a suitable selection of the number of features we overcome the challenge of data treated with independent component analysis (ICA) –to remove eye activity– not previously solved for the subject dependent case, this time generalizing to subject independent task transfer with 93.04 and 97.24% performances between low and no workload, and 91.93 and 97.43% between high and no workload, for HTN and HAN respectively, significantly different from chance, with four-dimensional feature vectors. We achieve very high and stable performances for wide ranges of universal scaling factors, also for data treated with ICA. We look for inner, essential, and permanent characteristics of brain activity reflecting mental workload, and benefit from the maximum permissible amount of data available in our procedures to discover trends that appear after averages over large amounts of data, as in other brain phenomena as the “Bereitschaftspotential”. After ICA and HAN and HTN feature extraction, we verify a statistically significant increase in voltage amplitudes of EEG signals in the parietal-occipital area after cognitive task execution, consistently identifying feature channels that successfully generalize task transfer to new subjects. Our methods could be used in further studies of mental workload –or similar analyses in other EEG data para-digms– opening a way to avoid subject dependent adjustments in each session, closer to universal classification of EEG brain signals, reducing the gap between the limitations of lab experiments and the requirements of real world BCI applications.